EP3732485A1

EP3732485A1 - Predicting the metabolic condition of a cell culture

Info

Publication number: EP3732485A1
Application number: EP19700004.5A
Authority: EP
Inventors: Wolfgang Paul; Arthur MOHR; Sayuri HORTSCH
Original assignee: F Hoffmann La Roche AG
Current assignee: F Hoffmann La Roche AG
Priority date: 2017-12-29
Filing date: 2019-01-02
Publication date: 2020-11-04
Also published as: JP7092879B2; CN112119306B; JP2021508872A; WO2019129891A1; CN112119306A; US20230313113A1; US20200377844A1

Abstract

The invention relates to a method for predicting the metabolic condition of a cell culture of cells of a specific cell type. The method comprises providing (102) a metabolic model (402) of a cell of the specific cell type and carrying out the following steps at each of a plurality of points in time during the cultivation of the cell culture: receiving (106) measured concentrations of several extracellular metabolites and a measured cell density in the culture medium; inputting (108) the received measured values as input parameter values into a trained machine learning program logic – MLP (218); predicting (110) extracellular flows (408) of the extracellular metabolites at a future point in time by means of the MLP; carrying out (112) a metabolic flow analysis for calculating the intracellular flows at the future point in time on the basis of the predicted extracellular flows and the stoichiometric equations of the metabolic model.

Description

Prediction of the metabolic state of a cell culture

description

Technical part

The invention relates to the prediction of the metabolic state of cells, in particular cells kept in cell culture. State of the art

In recent years, in the pharmaceutical industry, the endeavor has significantly increased the processes of development, production and production

To make quality assurance more efficient by placing a stronger focus on process analysis, control and This tendency also and particularly relates to the operation of cell culture reactors used to produce pharmaceutically interesting molecules, especially macromolecules such as proteins.

The modeling of cell metabolism and fermentation processes in the

However, pharmaceutical research and development is a major technical challenge: The metabolism of cells, especially eukaryotic cells, is characterized by very complex non-linear chemical reaction cascades that are difficult to simulate or model analytically.

For example, the analysis of metabolic fluxes using metabolic flux analysis (MFA) is known. The MFA is used in particular for processes in which the intracellular flux distribution is time-invariant. This is approximately given for the exponential growth phase in the sentence procedure ("Batch

Bioreactor ") or for cultivation in the chemostat. However, in today's predominantly used fed-batch processes, the cells are

Exposed to environmental conditions that are constantly changing. Therefore, the intracellular fluxes also vary in the course of the processes. So it is not enough to do a single metabolic flux analysis to map the state of the cells throughout the process. In addition, it is unfavorable in terms of process control, if only descriptively determined that the cells of a bioreactor have already gone into a metabolic unfavorable state. It would be better to predict this and counteract early if necessary. The use of MFA for analyte metabolite fluxes is described, for example, in Ahn WS, Antoniewics MR (2012): "Towards dynamic metabolic flux analysis in CHO cell cultures", Biotechnology Journal 7, 61-74. However, attempting to predict the state of the cells of a cell culture at a future date is also very difficult for several reasons. For the prediction of time courses kinetic models are usually used. These consist of a system of

Differential equations describing the temporal changes of substance concentrations or substance quantities. The equations are mostly based on

mechanistic knowledge, which can be modeled for example by means of a Michaelis-Menten kinetics. However, this kinetic knowledge is difficult to obtain. For the description of a whole intracellular

Metabolic network using kinetic expressions are usually not sufficient for the existing mechanistic information and the number of parameters to be estimated would be enormous.

Kinetic hybrid models that use metabolic flux analyzes with kinetic

However, in practice they are usually unable to generate reliable predictions of the future behavior of a cell culture. Creating these models has proven to be labor intensive

exposed. In addition, the models thus created are not very flexible and can not be adapted to the metabolism of other cell types without considerable manual effort. The use of hybrid models is described, for example, in Covert MW, Xiao N, Chen TJ, Karr JR (2008), "Integrating metabolic, transcriptional regulatory and signaling transduction models in Escherichia coli", Bioinformatics Vol.

24 no. 18, 2044, as well as in Nolan RP, Lee K (2011) "Dynamic model of CHO cell metabolism", Metabolic Engineering 13, 108.

Summary

Against this background, there is a need for improved methods for

Prediction of the metabolic status of cells and at corresponding

improved systems in that the aforementioned disadvantages are at least partially avoided.

The object of the invention is specified in the independent claims. Embodiments of the invention are described in the dependent claims. The embodiments and examples of the invention described herein are freely combinable with each other unless they are mutually exclusive.

In one aspect, the invention relates to a method for predicting the

metabolic state of a cell culture of cells of a particular cell type. The method comprises providing a metabolic model of a cell of the particular cell type, wherein the metabolic model comprises a plurality of

intracellular and extracellular metabolites as well as a variety of

intracellular and extracellular flows, and wherein the metabolic model comprises stoichiometric equations comprising at least one

specify stoichiometric relationship between one of the intracellular and one of the extracellular metabolites.

The method further comprises, at each of a plurality of times during the culture of the cell culture, performing the following operations:

- receiving a plurality of measurements measured at that time, the measurements comprising concentrations of several extracellular metabolites of the metabolic model in the culture medium of the cell culture and a measured cell density of the cells in the cell culture;

- Input of the received measured values as input parameter values in a trained Machinearning program logic (MLP);

- Predicting extracellular fluxes of extracellular metabolites at a future time point by the MLP using the received readings, the time of the future being a time subsequent to the reception of the readings, the extracellular flows taking up extracellular metabolite uptake rates into a cell and / or delivery rates extracellular metabolites from a cell into the medium;

Perform a metabolic flow analysis to calculate the

intracellular flows at the time of the future using the predicted extracellular fluxes of the extracellular metabolites and the stoichiometric equations of the metabolic model.

This can be advantageous since in this way the knowledge which is already present in the literature for many important metabolic processes of cells, for example in the form of stoichiometric reaction equations, and which in Representable form of metabolic models can be linked in a very advantageous way with the use of an MLP, which allows a high degree of flexibility. Thus, some experimental preliminary studies have shown that the use of MLP-based approaches, for example, the use of neural networks to predict future metabolic states of cells purely based on, for example, the concentration of their

Excretion products is unreliable. The complexity and dynamics of cellular metabolism also seem to make it impossible for sophisticated, modern MLP methods to make a reliable prediction of the metabolic state of a cell culture in the future purely based on actual cell concentrations, especially in fed-batch bioreactors with eukaryotic cells. On the one hand, this may be due to the fact that the cell-internal state is characterized only very indirectly and insufficiently by the measurement of extracellular metabolite concentrations, on the other hand there is the danger of "over-fitting" to the training data in the course of generating the MLP in the training phase. However, it has been shown that MLP-based methods, as described here for embodiments, are at least for a not too distant future

At this point in time, very reliable predictions can be made regarding the concentration and fluxes of extracellular metabolites. By combining this knowledge with the knowledge of stoichiometric and transport-related dependencies between intracellular and extracellular metabolites

between each other and embodiments of the invention it is possible to accurately predict the metabolic state of the cells up to the strength of individual intracellular material flows.

Although a rough characterization of the metabolic state of cells based on excreted substances (such as lactate) has been possible to date, the cell's internal processes ultimately remained a "black box". A fine granular prediction of cell metabolism down to the level of individual cell-internal fluxes was not possible with the currently known hybrid models.

In a further advantageous aspect of the invention, there is also provided a highly flexible method of predicting cell metabolism. It has been shown that already by cultivating a few cell cultures with the cells of a certain cell type a sufficiently large training data set can be obtained to train an MLP which is able to make reliable predictions about the metabolic state at a future time for cell cultures of this cell type. The extraction of

Training data and the subsequent training of the MLP can be done largely automatically. In contrast, the metabolic model used for metabolic flow analysis, which may require extensive literature study and manual adaptation steps, may often be reused for cells of different cell types. Thus, unlike using hybrid models known in the art, it is often possible using embodiments of the inventive method without the elaborate creation of new models solely through the cultivation of some cell cultures to generate a training data set and

then training an MLP on that training data set to provide a prediction method which can be specific metabolic

Characteristics of a particular cell type are considered.

In another advantageous aspect, the MFA now becomes a prediction

intracellular flows successfully used. So far, MFA has been used essentially only for processes in which the intracellular flux distribution is time-invariant. This is approximately given for the exponential phase in batch procedures or for cultivations in the chemostat. However, in today's predominantly used fed-batch processes and perfusion-like processes involving cell retention, the cells are exposed to environmental conditions that are constantly changing. Therefore, the intracellular fluxes also vary during the course of the process. Approaches that attempt to map the state of cells throughout the process have failed. In accordance with the invention, however, an approach has been found to accurately predict the metabolic state of fed-batch reactors, at least for the coming hours and days, by defining a dynamic metabolic model describing the intracellular flows as a function of the (possibly discretized) process time.

In a further advantageous aspect, embodiments of the method according to the invention make it possible to control the quality of the prediction or the intermediate results obtained during this prediction during the cultivation of the cell culture. The method described above involves the prediction of extracellular fluxes. These are easily determinable from the change in concentration of extracellular metabolites during fermentation by repeated sampling and analysis of culture medium samples. Thus, during the fermentation a

Verification of the prediction of MLP by repeatedly comparing the predicted and measured extracellular flows.

If significant differences occur here, the calculation can be stopped immediately and the reasons for the deviation can be researched. Other metabolic flux prediction methods known in the art, on the other hand, use 13C measurement data in the steady state around intracellular and

predict extracellular fluxes. Thus, these methods require 13C-labeled analyzes and experiments under steady state conditions to directly predict intra- and extracellular flows. 13C analyzes are time-consuming and during normal operation of a fermenter, stationary conditions can generally not be assumed. In addition, the

predicted intracellular flows are not empirically verifiable, so that in such methods, no intermediate step takes place, which can be compared with real measured data.

According to embodiments of the invention, the MLP receives input data in the form of extracellular concentrations of extracellular metabolites at a current time to assess the corresponding extracellular flows at a future time (next sampling time) under dynamic conditions

To predict fermentation conditions. The intercellular MFA is calculated on the basis of the stoichiometric metabolic model (for the future and optionally the current sampling time). This makes it possible to check the prediction on the basis of measured extracellular metabolite concentrations during the fermentation.

Embodiments of the invention thus provide a method of predicting the metabolic state of cell cultures that is both accurate (as it provides detailed information about intracellular metabolic processes and experimentally demonstrated that the extracellular flux predictions compare well with the real measured extracellular fluxes agree) and which is very flexible, because it can be easily adapted to cells of another cell type.

According to embodiments, at least some of the extracellular and intracellular flows of the metabolic model are not based on a decomposition of a known metabolic network into elemental flow modes. In the literature, simple intracellular network fluxes have been estimated using MLPs. Initially, the elementary flow modes of the network were generated. Elementary flow modes represent a set of permissible flow distributions, from which different possible weights make all other possible

Flow distributions can be combined. The MLP gives the

corresponding weights. However, the number of these elementary flow modes is so large for most biochemical networks that an effective and robust estimation of the weights is not possible.

According to the invention, however, only the extracellular flows are estimated via neural networks or other MLPs, and the coupling to the metabolic flux analysis then establishes the connection to the intracellular flux distribution.

Thus, according to embodiments of the invention, the machine learning program logic is adapted to selectively predict only extracellular flows but not intracellular flows.

This may be advantageous since the prediction is thus limited to values which can easily be measured during the operation of a bioreactor so that the extracellular flows represent intermediate values for the prediction of intracellular flows, whereby these intermediate values can easily be compared with real measured data to quickly identify errors in the MLP.

In a further aspect, the method can be advantageous for the following reasons: The metabolism of the cells depends essentially on the conditions in the bioreactor. If factors such as pH value, pO 2 value, pressure and temperature are kept constant, it is above all the metabolite concentrations in the reaction medium that are decisive for the behavior of the cells, which is why they were selected as inputs for the MLP. The prediction of future metabolite fluxes has increased compared to Prediction of future metabolite concentrations to the next

Sampling time has two main advantages:

On the one hand, the MLP trained in this way is more flexible in terms of the choice of

Time intervals. Namely, when networks with metabolite concentrations are trained as outputs, the prediction that can be made with such a trained MLP always relates to the same time interval as in

Training data set. When generating the training data, it should therefore be noted that the intervals are chosen as uniformly as possible, otherwise

Contradictions can arise. On the other hand, if flows are predicted, this results in some independence of the time interval for the current prediction from the time intervals used during training of the MLP. The method is thus more robust and flexible with regard to the choice of time intervals for the prediction.

On the other hand, the metabolite concentration depends on continuous and / or pulse-like nutrient additions, some of which can be handled variably. However, this variability is not present in the prediction if the trained MLP does not treat the doses that were carried out in the training data record separately, but learns them blindly. By using extracellular metabolite fluxes, future concentrations can be extrapolated using equation 4.4 of the appendix. With regard to the time interval, there is some flexibility, provided the flows are proportional to the

Sampling intervals are not subject to strong fluctuations. To

In preferred embodiments, the time interval between the times used for the extracellular flow prediction is less than or equal to or at most 20% greater than the time intervals used in the creation of the training data set.

According to embodiments, the calculated extracellular flows are calculated to be adjusted for drug doses (e.g., glucose additions). In the context of the prediction of the one or more extracellular flows, the information that in the next time interval a feeding of a specific extracellular metabolite takes place can be used to determine the

Prediction of the concentrations of this extracellular metabolite adjust accordingly. It is also possible to adjust the supplements based on the prediction if the MLP predicts a deficiency in this extracellular metabolite. It should be noted, however, that dosing may generally affect the mean flow as it will result in altered concentrations in the reactor. If the dosing profile differs greatly from that of the training data, therefore, inaccuracies could result

According to embodiments of the invention, the temporal and quantitative profile of the metabolite feeds to a cell culture whose metabolic state is to be predicted is selected to be identical or similar to the feed profile used to generate the training data with one or more training bioreactors, the MLP was trained on these training dates.

According to embodiments of the invention, the measurements for collecting the measurement data and the prediction of extracellular flows are made in real time, that is, during the operation of the bioreactor in which the cell culture is contained. The time period between the collection of the measurement data and the prediction of the metabolic state of the cells is typically small and in the range of a few seconds or minutes, typically less than 15 minutes, whereas the time intervals for individual predictions typically in the range of 1 -48 hours and esp. 6-24 hours are.

In a further advantageous aspect, embodiments of the

Invention thus predicts the metabolic state of a cell culture in real time, since the information basis on which the prediction is based exists in the already defined metabolic model, the already trained trained MLP as well as measurement data, whereby the measurement data can easily be collected in real time. Thus, for example, by regularly taking a sample from the culture medium and determining the cell number and metabolite concentration in this sample, the measurement data can be obtained, which is used to carry out the

Prediction are required.

According to embodiments of the invention, the method for predicting the metabolic state of a cell culture at a future time in real time is continuously performed during the operation of a bioreactor containing the cell culture. According to embodiments, the method further comprises generating the MLP by machine learning.

The generation of the MLP comprises generating a training data set, wherein generating the training data set comprises performing the following operations at each of a plurality of training times during the cultivation of at least one training cell culture of cells of the particular cell type:

Receiving a plurality of measured values measured at this training time, the measurements comprising concentrations of a plurality of extracellular metabolites of the metabolic model in the culture medium of the at least one training cell culture and a measured cell density of the cells in the at least one training cell culture;

- Receiving the time of the current training time; and

Calculation of extracellular fluxes of the extracellular metabolites as a function of the measurements received at that time and at the respective preceding time, the extracellular fluxes being uptake rates of the extracellular metabolites into the cell and / or delivery rates of the extracellular metabolites into the medium;

Train the MLP, where the training includes:

o Enter the readings taken at each of the training times

are input as input parameter values into the MLP, and input of the extracellular flows calculated at the time following that training time for that subsequent time as output parameter values associated with these input parameter values into the MLP; and

o performing a learning process by the MLP such that the MLP learns to predict the respective associated output parameter values based on the input parameter values;

Storage of the trained MLP in a volatile or non-volatile storage medium.

The use and generation of an MLP according to embodiments of the invention may be advantageous because, on the one hand, the choice is more appropriate

Machine learning algorithms are achieved a high accuracy of the prediction on the other hand an adaptation to the metabolic conditions in other cell types very easily and without much manual effort or

Literature study can be done. The operation of some training cell cultures with continuous collection and recording of the above-mentioned measurement data and the extracellular flows calculated from these is sufficient to one

To provide a training data set on the basis of which an MLP can be trained and generated specifically for the cell type of cells used in the training cell cultures.

According to embodiments of the invention, the training data set is generated so that at each of a plurality of training times during the

Cultivation of multiple training cell cultures of cells of the particular cell type receiving the measurements and time and calculating the extracellular fluxes of the extracellular metabolites. Here, the cell cultures are preferably cultured in bioreactors of different types. Preferably, these bioreactor types include at least two different bioreactor types from the following: a fed-batch bioreactor, a batch bioreactor, a perfusion reactor (including cell restraint variants), a chemostat, and a split-batch bioreactor.

The use of bioreactors of different types in the generation of the training data set may be advantageous because it creates a broader database and can avoid or reduce overfitting of the MLP during training. In addition, it allows the use of the same MLP to successfully predict the future metabolic state of a cell culture in many different bioreactor types. Preferably, the training data set is collected based on training cell cultures that are in different

Bioreactor types are cultivated, wherein the bioreactor types include at least one fed-batch bioreactor and / or at least one perfusion reactor. This may be advantageous since these types of reactors are increasingly used in practice and the presentation of the metabolic state of cell cultures in these types of reactors has heretofore been a particular difficulty due to their high dynamics. Preferably, the cell culture becomes its future metabolic state

is to be predicted, cultured in a bioreactor type which was also used to generate the training records.

According to embodiments, the training data set is generated such that at each of a plurality of training times, during the culture of multiple training cell cultures of cells of the particular cell type, the measurements and times are received and the extracellular fluxes of the extracellular metabolites are calculated Bioreactors of the same type or of different types are cultivated, with all

Bioreactors do not belong to the batch bioreactor type. For example, all bioreactor types could be of the fed-batch reactor type.

Due to the continuous or pulsatile addition of additional nutrient solution during operation, the prediction of future metabolic states of cells in fed-batch bioreactors has proven to be a particular technical challenge. Embodiments of the invention are currently in the context of

Use of fed-batch bioreactors is advantageous because it has been found that the predictions according to embodiments of the invention are accurate despite the metabolic complexity of cell cultures cultured in this type of reactor.

In some embodiments, the MLP is a support vector machine or multiple support vector machine system.

In other embodiments, the MLP is a neural network or a multiple neural network (NNs) system.

Some initial tests suggest that support vector machines and neural networks can be used in addition to other MLP methods. However, the use of neural networks has yielded particularly good predictive results, and a wide range of software solutions for various network architectures are already available that allow easy handling of the neural network during the training phase as well as during the application phase. According to embodiments of the invention, the MLP is a system of several sub-MLPs (in particular individual NNs), the individual ones being in the

System-contained sub-MLPs were each trained for the prediction of the extracellular flow of a single extracellular metabolite and selective for the prediction of extracellular flow of this single extracellular

Metabolites are used at the time of the future.

This may be advantageous as it has been found that the predictive power of measured extracellular metabolite concentrations is different with respect to single extracellular flows of other extracellular metabolites. The quality of the prediction can be increased by providing individual sub-MLPs, for example, individual neural networks, each based on a specific set of input parameter values with respect to the extracellular flow of the

Extracellular metabolites are trained as output parameter value. The

Results of individual sub-MLPs can be linked by a parent MLP or other program logic such that, as a result, an extracellular flow of one or more extracellular metabolites

is returned. A "sub-MLP" is an MLP that is operatively linked to one or more other "sub-MLPs" such that the output of this and the other "sub-MLPs" are from a parent program logic, particularly another, higher level MLP , combined to a total result, eg is aggregated.

In accordance with embodiments of the invention, MLP uses measured concentrations of multiple extracellular metabolites as input parameter values to predict the extracellular flow of a single extracellular metabolite. In this case, the several extracellular metabolites, which differ as

Input parameter values for at least two of the extracellular metabolites whose extracellular flux is to be determined.

This can be advantageous since it allows a higher accuracy of the prediction to be achieved. According to embodiments of the invention, the method further comprises a

Measurement of the concentrations of all input candidate metabolites over several time points. In particular, the measurement of the creation of

Serve metabolite concentration profiles over time. The amount of

Input candidate metabolite includes all extracellular metabolites that are in a reference bioreactor with a cell culture of the specific type

are available by measurement or all extracellular metabolites of the metabolic model. For example, the reference reactor may be one or more training bioreactors, that is, reactors used to collect training data to generate the MLP. Alternatively, the reference reactor may also be another reactor having a cell culture of the same type as the currently monitored bioreactor. The determination of those extracellular metabolites, which are to serve as input parameter values for the MLP with regard to the prediction of the extracellular flows of individual extracellular metabolites of the model, thus preferably takes place during or before the training of the MLP.

For each of the extracellular metabolites, whose extracellular flux is to be predicted, a selection procedure is carried out for the identification of the several extracellular metabolites, which are known as

Input parameter values should be used to predict the extracellular flow of this single metabolite. The selection procedure, with respect to this single metabolite, respectively comprises: a) defining a first amount of extracellular metabolites, the first amount comprising all input candidate metabolites; b) Calculation of a first relevance score of each extracellular

Metabolites in the first amount as a function of the measured concentrations of that metabolite, the first relevance score indicating the predictive power of the concentration of the particular extracellular metabolite with respect to the extracellular flux of that single extracellular metabolite; c) transfer only of the extracellular metabolite having the highest first relevance score from the first amount to a still empty second Amount of extracellular metabolites removing this metabolite from the first amount; d) calculate another relevance score of each extracellular

Metabolites in the first amount as a function of the measured concentrations of this metabolite and the measured concentrations of all contained in the second amount of extracellular metabolites, wherein the further relevance score the predictive power of the concentration of the respective input candidate metabolite with respect to the extracellular flow of this single extracellular metabolite indicating the metabolites already present in the second quantity; e) transferring only that extracellular metabolite of the first quantity having the highest further relevance score into the second quantity by removing that metabolite from the first quantity, the transfer taking place only if the second quantity is absorbed by the uptake of this metabolite does not exceed a maximum limit for informative redundancy of the metabolites it contains with respect to the prediction of extracellular flow of that single extracellular metabolite; f) repeating steps d) and e) until no more metabolites can be transferred from the first to the second amount without the second amount exceeding the maximum limit for informative redundancy; and g) selectively using only the metabolites that were converted to the second amount as input parameter values for predicting the extracellular

Flow of this single extracellular metabolite.

According to embodiments, the first relevance score is calculated as the partial mutual information score (PMI score) between a metabolite of the first amount and the single metabolite whose extracellular flow is to be predicted. The second relevance score is calculated as the PMI score - between a metabolite of the first quantity and the single metabolite whose extracellular flux is to be predicted, taking into account all the metabolites already contained in the second quantity. This can be advantageous, as it allows input parameter values to be determined and used which have particularly high predictive power for the particular extracellular metabolite or its flux. Overfitting in the course of training and thus poor predictive quality can be avoided.

By identifying those metabolites responsible for a given

Extracellular flow of extracellular metabolites that have the highest expressiveness ("relevance" or "predictive relevance") ensures that the selected input parameter values will produce a good MLP

Enable predictive power. The predictive relevance is preferably determined by determining the measure of the correlation of a measured

Metabolite concentration profile of a particular extracellular metabolite with the metabolite concentration profile of the metabolite whose extracellular flow serves as the output parameter value of the MLP (ie: its extracellular flux is to be predicted). The predictive relevance can be determined by various methods, e.g. Principal component analysis or PMI ("partial mutual information") as will be described below for embodiments of the invention. The fact that a metabolite is included in the second quantity only if it does not already contain a metabolite whose concentration profile correlates strongly with the concentration profile of that metabolite (indicating a high degree of informative redundancy of this metabolite with it already in the second quantity implies that the second set also contains groups of two or more metabolites whose concentration profiles would strongly correlate and thus introduce redundant information into the second set. A high proportion of informative-redundant metabolites in the second amount would lead to overfitting effects.

For example, for the absorption of further metabolites from the first to the second amount, the achievement of a maximum value for the informative redundancy or another termination criterion can be defined so that, as a rule, only a portion of the metabolites is transferred from the first to the second amount.

Thus, according to embodiments of the invention, the selection of

Input parameters (extracellular metabolites whose concentration is measured and entered into the MLP) are completely independent of the MLP. However, it is also possible that the selection process is performed in the form of a "wrapper" as described in Chapter 4.7.3 of the Appendix, eg as a functionality provided by the neural network.

According to embodiments, the first relevance score of each of the metabolites in the first amount with respect to that single extracellular metabolite is calculated as the PMI score between that metabolite and the individual

Metabolites whose extracellular flux is to be predicted in each case.

A PMI score (as opposed to the simpler Mutual Information (Ml)) may also be used in the other iterative steps to calculate the further relevance score of any metabolite remaining in the first set. The further PMI score indicates the predictive relevance of this metabolite with regard to the extracellular metabolite, with the informative

Redundancy of the metabolite to be tested from the first quantity with respect to all metabolites already taken in the second amount received. Before a

Thus, if the metabolite is ultimately included in the second quantity, it will be examined whether, in view of the metabolites already present in the second quantity, it will at all yield a "predictive added value" beyond the metabolites already contained in the second quantity. If the measured concentration profile of this metabolite correlates strongly with a metabolite already in the second amount, this should be answered in the negative. In this case, none finds

Admission to the second set instead.

Using a test data set, the quality of the selection of the input parameters can be estimated and the architecture of the MLP (e.g., number of layers of a neural network) optimized.

The determination of the predictive relevance of a metabolite with respect to an output parameter metabolite by means of the PMI therefore allows a measurement or an estimate of the possible input variables and the dependence between each of the possible input variables with regard to the output variable. The stronger the dependency, the better the output can be determined by the

Predicted input variables, the higher the relevance score of the input value metabolite. This allows the calculation of relevance scores and score-based sorting of metabolites of the first set. In the PMI-based decision to include a metabolite in the second set, the PMI determines the dependence of this metabolite on each metabolite already in the second set. Only metabolites that contain sufficiently relevant new information compared to the metabolites already in the second quantity are included in the second quantity so that no redundancies occur.

When selecting the next relevant metabolite in the second quantity, the metabolites already selected in the second quantity are thus also included

considered.

The calculation of the dependency between the extracellular metabolites in the form of PMI (partial mutual information) may be e.g. done in the course of the

Decide if one metabolite of the first quantity in the second quantity

the calculated PMI with a PMI criterion, e.g. a PMI limit for a still acceptable degree of dependency; however, the comparison with a PMI criterion does not usually consist of a simple comparison with a threshold, but consists of a statistical test or an alternative selection procedure such as e.g. a "wrapper" as explained, for example, in the appendix in equations 2.22-2.27 and in chapter 4.7.3 of the appendix. According to the embodiment described in the appendix, first of all the "most relevant", ie the one with the highest PMI value, is always taken over into the first quantity. This creates an order of all metabolites. Finally, however, ALL metabolites are in the first lot. In a subsequent step (wrapper) it is then decided how many of the most relevant metabolites are used in the "second list" of those for prediction.

According to embodiments, MLP uses to measure the extracellular flow of each extracellular metabolite

Concentrations of multiple extracellular metabolites as input parameter values, wherein the plurality of extracellular metabolites comprise at least one, preferably at least two, amino acids. The metabolites whose concentrations are used as input parameter values often, but not necessarily, include the metabolite whose flow is used as the output parameter value. The use of amino acid concentrations in the medium as

Input parameter values for the MLP or for training the MLP may be advantageous because established methods for determining their concentration already exist and the sufficient presence of amino acids is often necessary for efficient synthesis of many target proteins in a bioreactor.

It is possible that during training, only the concentrations of this subset of metabolites are selectively used to train the MLP.

If you classify the input parameter values of the first set according to their relevance (for example according to PMI), all available inputs are included in this "set" at the end, but in the order that indicates the relevance. An additional criterion is then used to derive from this list the number of

Input parameters (concentrations) to determine, which will then be used as input in the future. According to embodiments of the invention, the selection was made during the training itself (there were different numbers

Input parameters compared with respect to their predictive power evaluated on a test data set).

Preferably, only the concentrations of this subset selectively become

Metabolites measured in cell culture whose metabolic status is to be predicted are used as input parameter values for the already trained MLP. The selective use of predictive relevant and independent input variables (rather than all metrologically available extracellular metabolite levels) may be beneficial as it reduces the problem of overfit, simplifies data collection (it may not be necessary)

Concentrations of all the extracellular metabolites of the model are measured) and the prediction of less computational resources is consumed as fewer input parameter values need to be evaluated.

According to embodiments, the metabolic model for the intracellular metabolites of the model represents a steady state assumption that the mole fraction of intracellular metabolites remains constant such that the sum of the incoming flows for each intracellular metabolite corresponds to the sum of the outgoing flows of that metabolite. According to embodiments, the plurality of times (used to predict the extracellular flows of the currently cultured cell culture) are separated by time intervals of 10 minutes to 48 hours, preferably 1 to 24 hours. According to preferred embodiments, the time interval between the times used for the prediction of the extracellular flows is less than or equal to or at most 20% greater than that in the preparation of the

Training data set used time intervals.

According to embodiments, the multiplicity of times are separated by time intervals which are the same length over the duration of the cell culture or whose length becomes smaller at the end of the duration of the culture of the cell culture. According to preferred embodiments, the profile of the change of time intervals between the times used for the extracellular flow prediction is identical or very similar to the profiles of the changes of time intervals used in the generation of the training data set

Time intervals.

The particular cell type may be a prokaryotic or a eukaryotic cell type.

In particular, the particular cell type may be a eukaryotic cell type.

It has been found that, despite the high complexity eukaryotic

Metabolic processes Embodiments of the invention are capable of accurately predicting the metabolic state of the cell, particularly for future periods ranging from hours to 1-2 days in the future.

For example, the particular cell type may be a mammalian cell type, e.g. HELA cells, etc. In one embodiment, the particular cell type is um

Chinese Hamster Ovary (CHO) cells.

According to embodiments, the particular cell type is a genetically modified cell type that is maintained and / or propagated in a bioreactor for the purpose of obtaining a biomolecule. For example, it may be a genetic modified cell line expressing a particular protein, such as an enzyme or a specific antibody, and / or expressed in particularly high amounts.

According to embodiments, the calculated intracellular flows are checked for plausibility and consistency and / or with regard to further quality criteria and if necessary by supplementing, removing or changing

modified stoichiometric equations. The measured and / or

Predicted extracellular metabolite fluxes are now transferred to the modified metabolic model to recalculate intracellular fluxes for plausibility and / or consistency. Thus, the quality of the metabolic model can be improved and possibly also to certain

Cell types or specific cell clones can be adjusted. Based on this

Plausibility criteria can according to embodiments of the invention on the basis of several experimental experiments also plausible reference values for the

intracellular fluxes are recovered.

According to embodiments, the calculation of the intracellular flow of one or more intracellular metabolites at each of the future times involves a calculation of multiple or preferably all intracellular flows of the metabolic model. Preferably, all intracellular flows of the model are calculated. The more intracellular flows are taken into account, e.g. in a plausibility assessment, the reliability of the prediction increases.

According to embodiments, the method includes an identification of all

intracellular fluxes, by a respective reference value or

Reference range deviate more than the limit. The reference values or reference value ranges may e.g. be obtained empirically and / or derived from the literature. The method further includes automatic identification of that intracellular flow that acts as a limiting factor for cell growth or production of a desired biomolecule.

For many metabolites, their approximate intracellular flux is known in the context of a particular metabolic condition, e.g. using kinetic models, 13C labeled substrates, and NMR quantification of

Isotopomers of the metabolites or by the amino acid composition of the cell proteins, etc. Strong deviations from these reference ranges thus suggest suggest that the metabolism of the cells of the cell culture in one

unfavorable or at least unexpected state. The fact that embodiments of the invention compare intracellular fluxes with reference values and not the concentration of extracellular metabolites with reference values in order then to conclude on the metabolic state of a cell, is advantageous because a more granular and better detection of (mostly unwanted) Deviations of the physiological state of a cell from physiologically usual or favorable reference values is possible.

In another aspect, the invention relates to a method for monitoring and / or controlling a bioreactor that includes the cell culture of cells of a particular cell type. The method includes a calculation

intracellular flows at a future time according to the embodiments described herein and examples of the method for predicting the

metabolic state of cells. The method further comprises a comparison of the predicted intracellular fluxes with reference values or

Reference ranges for acceptable intracellular fluxes of the respective one or more intracellular metabolites.

The method may be for monitoring the bioreactor and include outputting a warning, wherein the warning is issued when a deviation of the calculated intracellular flow from its respective reference or reference value range exceeds a threshold. The warning may e.g. be sent via a graphical user interface to a human and / or via another interface to machines or software programs that logs the deviations.

Additionally or alternatively, the method may serve to control the bioreactor and include sending a control command to the bioreactor. The control command is sent to automatically initiate steps that change the state of the bioreactor or the media it contains, so that the

Deviation is reduced. For example, the control command can go to a valve, a pump or other actuator of the reactor and there effect the addition of nutrient solutions, trace elements, oxygen, C02, pFI-regulating acids or bases or a corresponding throttling of the addition. In particular, the automatic steps may involve a change in the amount or composition of a nutrient solution. The control command can thus go, for example, to a mixer of a nutrient solution or a throttling unit on a supply line of the nutrient solution, for example, depending on the predicted flows, the amount of certain sugars, salts and / or amino acids in the

Change nutrient solution or reduce or increase the feed rate of nutrient solution in the bioreactor.

For example, according to embodiments of the method, it can be ascertained or predicted that a specific intracellular metabolic pathway is markedly less pronounced (recognizable by low intracellular flow) than is expected or desired. From this metabolic pathway it can be known that it is often limited by the amount of a particular vitamin or trace element in the medium, e.g. Iron. If, therefore, it is recognized that this particular intracellular flow is lower than expected, targeted addition of iron into the bioreactor can be much more targeted than is possible if only care is taken to ensure that physical or extracellular parameters such as temperature, pH, glucose concentration, etc. are kept constant.

According to embodiments, the method for monitoring and / or controlling a bioreactor includes identifying that reaction within the metabolic model of the cells which acts as a limiting factor for cell growth or production of a desired biomolecule according to the embodiments described herein and examples of the method of predicting the bioreactor metabolic state and multiple intracellular flows of cells. The method further comprises automatically adding selectively to those

Substances (in particular enzymes, co-enzymes, trace elements or nutrients) which (exclusively or in particular) change the limiting factor acting intracellular flow so that the cell growth or the production of the biomolecule or the quality of the biomolecule is promoted. Additionally or alternatively, the method includes outputting a request for such an addition via a user interface.

This may be advantageous because more detailed monitoring or the taking of very specific control measures for controlling a bioreactor is possible than is possible with methods known in the prior art which rely on trying to keep constant only the operating parameters of the bioreactor, including some parameters measured in the medium. For example, a strong uptake and metabolism of amino acids does not necessarily mean that the cells also use the amino acids to synthesize the desired target protein. Depending on the state of the intracellular flows of the cell, it may also be that the ingested amino acids are metabolized for completely different purposes. However, according to the invention, this can be achieved by a metabolic flux analysis based on a specific metabolic model of the cell using the predictions of MLP.

According to embodiments of the method, the predicted

extracellular and intracellular flows also used to test the quality of the model. For example, if a model formulated for a particular CHO cell clone deviates strongly from another clone or from another cell line (other tissue or other animal species), the predicted flows (intracellularly by MFA, extracellularly using the MLP) may measured concentration change rates or plausibility criteria (no unrealistically or unphysiologically high rivers, etc.) are compared. At high deviations of the predicted from the measured or plausible flows, the model is adapted or regarded as an indication of the presence of an error in the metabolic model on the basis of which the model is corrected.

According to embodiments, by comparing the predicted flows with the measured or plausible flows, measurement errors are identified in the determination of extracellular metabolite levels or biomass, e.g. by means of statistical tests.

In a further aspect, the invention relates to a method of identifying a metabolically beneficial clone of cells of a particular cell type. The

Method includes:

Cultivating different cell cultures in several bioreactors, the different cell cultures being different clones of cells of the cell

include certain cell types;

- Calculation of the intracellular flow of one or more

intracellular metabolites separated at multiple time points for each of the Cell clones according to the embodiments described herein and examples of the method for predicting the metabolic state of cells;

Identification of that of the cell clones whose calculated intracellular flow of the one or more intracellular metabolites is metabolically most advantageous.

This may be advantageous since, in the context of pharmaceutical synthesis methods, the identification and selective propagation of metabolically advantageous cell clones is often necessary. Many known methods of genetic modification of cells do not provide complete control over whether and where in the genome of a cell a particular gene encoding a target protein to be synthesized is incorporated. Depending on the location in the genome, the expression rate may be different. For example, transfection of cells with viruses is a process in which many different cell clones are formed, some of which do not contain the desired gene at all and others have incorporated the desired gene but at different sites in the genome.

According to the invention can now in real time in parallel operation of several

Bioreactors with the different cell clones are determined whether the intracellular fluxes indicate that the gene encoding the target protein was incorporated into the genome of the cell and the target protein in the cell in

significant extent is synthesized. For example, a comparison of the amino acid composition of the target protein with the intracellular fluxes for the synthesis or degradation of individual amino acids may provide an indication of whether the target protein has been incorporated. Additionally or alternatively, the intracellular flows can provide information as to whether a particular cell clone proliferates sufficiently rapidly and is vital, whether it has a low rate of formation of toxic or otherwise undesirable metabolites, etc.

However, embodiments of the method of the invention may be used not only to predict the future metabolic state of cells, but also to describe the current metabolic state of a cell.

According to embodiments, the method comprises:

- Calculation of the current extracellular flow of one or more of the extracellular metabolites by reference to the current time and concentrations of extracellular metabolites measured at the previous time point;

- Carrying out another metabolic flow analysis to calculate the current intracellular fluxes at the current point in time

Use of the calculated current extracellular fluxes of extracellular metabolites and the stoichiometric equations of the metabolic model; and

Use of calculated current intracellular fluxes as a description of a current metabolic state of the cells of the cell culture.

This may be advantageous because a very accurate description of the current metabolic state of the cells of a cell culture down to the level of individual intracellular flows is provided.

According to embodiments, the readings also include a lactate dehydrogenase (LDH) concentration, and at each of the times during the culture of the cell culture, an LDH concentration measured in the medium of the cell culture is received. The prediction of extracellular flows of extracellular

Metabolites at each of the future times are provided by the MLP below

Use of corrected rather than measured cell density.

Preferably, the calculation of the corrected cell density for each of the time points comprises a calculation of the density of lysed cells in the medium of the cell culture as a function of the measured LDH concentration. In particular, this function can be an empirically determined, heuristic and linear function which represents the dependence of the LDH concentration in the medium on the number of lysed cells of this particular cell type. The corrected cell density is then calculated as the sum of the measured cell density in the medium and the calculated density of the lysed cells.

This can be advantageous since cells which are completely or partially lysed are often not or only poorly detectable by optical methods for cell density determination, but the lysed cells possibly had an influence on the concentration of the extracellular metabolites until shortly before their lysis. The standard methods for determining cell density only detect cells whose structure is still intact. This results in a distortion of the total cell density determination, if The decay of cells occurs, which tends to be the case in a fermentation at later times. It has been observed that in some cases a systematic error in the prediction of the rivers with the here after

Embodiments described procedures was observed, so that the predicted flows were not consistent with the measurable flows but rather had a systematic error. It was observed that this error correlated with the LDFI concentration in the medium, which is an indicator of the presence of lysed cells in the medium. This enzyme is not released by an intact cell into the medium. The detection of LDFI in the reaction medium thus indicates destroyed cells.

With the exception of batch fermentation, a similar, approximately linear relationship between the LDFI concentration and the number of lysed cells was detected in many fermentation approaches. To

Embodiments, therefore, the LDFI concentration in the medium is used as a further measured value to calculate a corrected cell density.

According to embodiments of the invention, the corrected cell densities are used as input parameter values during training of the trained MLP.

According to embodiments, in the MLP-based prediction of

extracellular fluxes used in addition to the extracellular metabolite concentrations the LDFI concentration as an input variable for predicting at least some of the extracellular flows of the model. To

Embodiments, in addition or alternatively, the LDFI-corrected cell density value is used as the output variable. Accordingly, according to embodiments of the invention, cell densities measured by measured LDFI are used in training the MLP for extracellular flux calculation

Concentration has been corrected and / or an LDFI concentration is used as another input parameter value or concentration of an extracellular metabolite.

In a further aspect, the invention relates to a system for predicting the metabolic state of a cell culture of cells of a particular cell type. The system comprises one or more processors, a first interface to the Receiving readings from a bioreactor containing the cell culture and a volatile or non-volatile storage medium.

The storage medium includes a metabolic model of a cell of the particular cell type, the metabolic model comprising a plurality of

intracellular and extracellular metabolites as well as a variety of

intracellular and extracellular (408) flows, the metabolic model comprising stoichiometric equations comprising at least one

specify stoichiometric relationship between one of the intracellular and one of the extracellular metabolites. Additionally, the storage medium includes trained machine learning program logic (MLP) as well as program logic adapted to provide a metabolic prediction method at each of a plurality of times during culture of the cell culture

Condition of the cells. This method includes:

- receiving a plurality of measurements measured at that time via the first interface, the measurements comprising concentrations of several extracellular metabolites of the metabolic model in the culture medium of the cell culture and a measured cell density of the cells in the cell culture;

- input of the received measured values as input parameter values into the MLP;

Perform a metabolic flow analysis to calculate the

intracellular flows at the time of the future using the predicted extracellular fluxes and the stoichiometric

Equations of the metabolic model.

For example, the metabolic model may be in the form of an amount of stoichiometric reaction equations defined in MatLab in the storage medium be saved. The storage medium can be a main memory of a computer or an electromagnetic or optical storage medium, eg a hard disk. The storage medium can also be a distributed system of a plurality of hardware storage units, for example a cloud storage area offered by a cloud service or the IT infrastructure of a laboratory. The system can be implemented, for example, as a standard computer, or as a notebook or portable mobile device of a user. The system may be incorporated in or operatively linked to a Laboratory Information System (LIS). The system can also be a

Control computer for one or more bioreactors act or on

Control module that is reversibly or irreversibly coupled to a single bioreactor, for example, is an integral part of the bioreactor.

For example, the first interface may be a network interface that establishes a wireless or wire-based connection to one or more sensors of the bioreactor. In addition or alternatively, the first interface may also be an interface for manually entering the corresponding measured values. For example, it is possible that the

Measurements can be made by placing the bioreactor in the manual, semi-automatic or fully automated procedure

At certain times, a sample is taken, which is transported to one or more other analyzers, where then the cell density and / or the concentration of extracellular metabolites takes place. The won

Measured values can in turn be transmitted automatically from these analysis devices via the first interface to the system, or a user enters the measured values into the system by hand via a first interface designed as a graphical user interface.

According to embodiments, the system is a control unit for monitoring and / or controlling a bioreactor or a plurality of bioreactors or is operatively coupled to such a control unit. Preferably, the system further comprises a user interface for outputting the calculated intracellular flow to a user. For example, the predicted flows can be represented in tabular form, or a dynamic visualization of the flows can take place by means of a moving image, for example a moving image based on a graphic image Representation of the metabolic model underlying the prediction of intracellular fluxes.

According to execution systems, the system further comprises a second interface for sending control commands to the bioreactor. This program logic is designed to:

Comparing the predicted intracellular flux with reference or reference ranges for acceptable intracellular flow of the particular or respective intracellular metabolite;

Outputting a warning via the user interface if a deviation of the calculated intracellular flow of at least one of the intracellular metabolites from its respective reference value or reference range exceeds a threshold; and or

Sending a control command to the bioreactor via the second interface, wherein the control command is adapted to change the state of the bioreactor or the medium contained therein so that the deviation is reduced.

The user interface may be, for example, a graphical

User interface (GUI) and / or an acoustic user interface act. For example, a warning sound can be output via the acoustic user interface or it can be a warning in the form of a text,

preferably displayed with a qualified flashing on the intracellular flow for which the deviation has been detected via the GUI on a screen. The screen may be a screen coupled to the bioreactor or the screen of a computer connected to one or more bioreactors over a network, for example a desktop computer, a server or a mobile communication device , for example a smartphone of the user.

The second interface may, for example, be physically configured as a wireless or wire-based connection between the system and the bioreactor of the bioreactor. The actuators can be, for example, pumps, valves for various nutrients, buffers, pH-regulating liquids, trace elements, gases and / or stirrers or temperature regulators. A "training time" is a time during the creation of the training data set. In contrast, the "time" refers to

Claim 1 at a later time, when trained on the training data set MLP is used to predict the metabolic state of a currently monitored and / or controlled bioreactor.

A "metabolite" in the narrower sense is understood to mean an intermediate (intermediate) in a biochemical pathway. However, a metabolite according to the present invention should be broader than any substance which participates in a biochemical reaction of a cell in the form of a starting material, product or intermediate. In particular, a metabolite may be an amino acid, a sugar, fats, peptides, antibodies, proteins, citrate cycle components, glycolysis components, components of protein synthesis or degradation pathways, and the like.

By an "extracellular metabolite" is meant herein a metabolite of which one is aware or of which, according to a metabolic model of the cell type being studied, is believed to be present in the medium of a cell culture, e.g. because it is secreted by cells of the particular cell type into the medium of cell culture (e.g., lactate) or because it is added as a component of the medium or a nutrient solution to the cell culture (e.g., glucose).

By "intracellular metabolite" is meant herein a metabolite of which is known or believed to be within cells of the particular cell type, e.g. because it is taken from the medium of cell culture or is produced by the cells.

A "flow" is an amount of matter that changes per time in a given volume due to a particular transport or reaction process.

A flow is therefore also referred to as "reaction rate" or "transport rate". If several processes run concurrently in the given volume, it is possible that, due to the opposition of some processes, there is no net change in a substance concentration. , eg by substance conversion, uptake or delivery. According to embodiments, the volume to which a flow statement relates is the volume of a cell in the case of intracellular as well as extracellular metabolites.

By relating the change in the amount of substance produced by the particular process to the cell or the cell volume, a flow implicitly also gives a change in concentration of this substance per time in this volume, which is caused by the specific process

Transport process is effected on. Since, given the volume of the flow of a substance, the change in concentration of that substance in the volume implicitly results from the particular process, and vice versa, the concept of a flow here is likewise a change in a quantity of substance per unit of time in one volume and one per time concentration amount included.

Preferably, the metabolic model used according to embodiments of the invention assumes that the amount or concentration of the

intracellular metabolite remains approximately constant. However, this does not mean that the intracellular flows remain constant. Rather, a cell into whose cytosol the concentration of an intracellular metabolite rapidly increases, e.g. compensate by increasing the rate of one or more chemical reactions that metabolize this metabolite.

An "extracellular flow" is understood here as meaning the rate at which an extracellular metabolite is taken up or released by the cell via a specific transport process. More specifically, it is the amount of metabolite that is taken up into the cell by this particular transport process per cell and per time, or by the cell into which the surrounding medium is delivered. According to embodiments of the invention, the extracellular flux v of a component at time t is determined from the measured change in the concentration of the corresponding metabolite in the reactor medium, which is normalized with the live cell density.

For example, according to embodiments, the change in concentration of the extracellular metabolite in the culture medium is measured over a time interval. From this, it is possible (using the volume of the media in the reactor) to determine the absolute metabolite quantity change in the medium. The viable cell count in the medium can be measured. By normalizing the absolute change in the amount of extracellular metabolite on cell number, the extracellular flux can be given as specific (biomass-related) size. Since that

average volume of a single cell of the cultured cell type is usually known from the literature or can be measured, is after

Embodiments of the invention, a conversion of the measured

Concentration change performed on the average cell volume.

For example, the metabolic model of cell metabolism used for metabolic flux analysis (also called "flow analysis" or "flux analysis") may include extracellular stoichiometric reaction equations each associated with extracellular flow. In particular, extracellular fluxes thus describe uptake rates of extracellular metabolites into the cells of a cell culture, so that they then function as intracellular metabolites, and release rates of extracellular metabolites of the cells of the cell culture to the surrounding medium, so that they are there as extracellular metabolites

Occurrence.

By way of these extracellular flows, which represent transport flows into and out of the cell, the purely intracellular flows are linked to the concentrations and concentration changes of the extracellular metabolites in the metabolic model and allow it to be determined by means of those specified in the model

Stoichiometries, plausibility criteria and the empirically determined or

predicted changes in the concentration of extracellular metabolites in the medium to make descriptive statements about current intracellular flows of the cell and / or predictions about future intracellular flows of the cell on the basis of extracellular flows.

The intracellular and extracellular flows are thus coupled together in the metabolic model via one or more intracellular metabolites, as some substances are described as both intracellular metabolites, their rate of formation or rate of metabolism into one or more intracellular metabolites, and their import rate into the cell or delivery rate from the cell can be described by extracellular flows of the model. A change in the concentration of an extracellular metabolite and thus an extracellular flow can also be effected in some bioreactor types by an external addition or feeding of this metabolite into the cell culture medium. An extracellular flux can be approximately calculated, for example, by first determining the absolute difference of the metabolite concentration measured at two successive times and then converting this to the measured cell density. The higher the cell density, the lower the flow per cell, since the absolute measured change in substance concentration is distributed over a larger number of cells. The uptake of an extracellular

Metabolites at the time of the future can thus be calculated as the metabolite concentration difference measured in the time interval between the two times divided by the duration of the time interval, the result then being divided by the measured cell density.

By "intracellular flow" is meant the rate (a quantity related to a time interval) with which a reaction takes place within a cell, the reaction consisting of: a transport between intracellular compartments (eg transport from the cytosol into the mitochondria and vice versa) or a conversion of one or more intracellular metabolites (educts) into one or more other metabolites (products) in the cell. In an intracellular flow, all starting materials and products are intracellular metabolites.

Intracellular flows are formulated in the metabolic model for stoichiometric equations which specify a reversible or irreversible response of the above categories (intracellular transport between cell compartments, metabolism).

"Immeasurable rivers" are rivers that are not readily apparent from the

Fermentation data can be determined. As a rule, intracellular fluxes are not measurable due to the difficult observability of intracellular metabolites.

According to preferred embodiments of the invention, the MFA is based on

Determination of intracellular fluxes based on provided extracellular fluxes on the assumption that the concentration of an intracellular

Do not change metabolites. Concentrations of intracellular metabolites are

metrologically very difficult or impossible to detect. Embodiments of Invention are based on the assumption that the sum of the input and output flows on an intracellular metabolite (the concentration of which is unknown) is identical and therefore its concentration does not change. This is not correct, depending on the mode of operation of the reactor, but at least

Approximately with respect to the time intervals described here sufficiently well, since the cell metabolism quite quickly back to its chemical

Balance returns.

According to embodiments, the metabolic model used in the MFA comprises at least 10, preferably at least 20 stoichiometric equations. The model is preferably designed to be metabolic

Completely or at least approximately describing processes in the cell completely.

The use of a metabolic model which describes the cell metabolism as comprehensively as possible can be advantageous, since for the selection of particular cell clones a most complete picture of the individual metabolic activities and peculiarities of each cell clone examined is important. A pure plausibility check of the extracellular flows predicted by the MLP by means of individual stoichiometric equations would generally not provide sufficient data in order to be able to select specific cell clones on the basis of advantageous metabolic properties.

In another advantageous aspect, the use of metabolic models with a large number of stoichiometric equations in MFA can also be used to monitor cell cultures and, in the event of a deviation of one or more metabolite flows from a setpoint range, specific changes in the amount and / or or composition of

supplied nutrient solution or other control parameters (temperature, pH, partial pressure oxygen, partial pressure C02, speed stirrer etc)

automated and / or manually.

By a "metabolic model" is meant a descriptive model of the current metabolic state of cells of a particular cell type. The metabolic model is preferably reduced in complexity compared with the real reactions taking place in the cell and is limited to those parts of the cell metabolism that are of particular interest to the particular application. Preferably, the model includes several

extracellular and intracellular metabolites, reaction and transport equations with stoichiometric factors as well as extracellular and intracellular flows. The metabolic state of a cell can be characterized, at least in part, by stating the fleas of each intracellular flow at a particular time. Preferably, the temporal change of the results

Metabolite quantities in the cell from the balancing of incoming and outgoing material flows, such as those in the extracellular and intracellular

Reaction equations of the model are specified. The model is an MFA model that is suitable for performing a metabolic flux analysis and is based on the so-called steady-state assumption, according to which the amount of substance intracellular

Metabolite remains constant, which means that the sum of incoming flows for each intracellular metabolite is the sum of outgoing

equivalent.

Concentrations of extracellular metabolites measured and reported as

Input parameter values of a trained MLP are used, it is embodiments of the invention to Metabolitkonzentrationen in the strict sense. Under a metabolite concentration in the narrower sense is here a

Salary statement, which refers to a volume understood. The concentration thus indicates how much of a metabolite in a reference volume (e.g.

Cell culture medium) is present. The metabolite concentration may e.g. when

Mass concentration in the unit g / l or as molar concentration in the unit mol / l.

According to other embodiments, the concentrations of extracellular metabolites that are measured and used as input parameter values of a trained MLP are metabolite concentrations in a broader sense. A metabolite concentration in a broader sense is taken to mean measured values and values derived therefrom which are known to have a metabolite concentration in the narrower sense of the respective metabolite in a linear manner or at least approximately (at least 90% in the case in question)

Concentration range) in a linear manner. For example, an extracellular flux of an extracellular metabolite may be the metabolite concentration in the be understood in a wider sense. The measured extracellular flux refers to the change in the concentration of the metabolite in the medium in the period between a point in time in the past, eg the last measurement of the

Metabolite Concentration, and a current time at which the measured values are currently being collected and used as input to the MLP. The im

Essential linear relationship between measured flow and

Extracellular concentration results from the fact that an increase of the

extracellular flow of a metabolite by a certain amount causes a corresponding change in the concentration of the metabolite in the extracellular medium. In addition, the measured metabolite concentrations according to embodiments of the invention can still be modified in various ways by offsetting with correction and normalization factors, so that these modified values ultimately also represent metabolite concentrations in the broader sense, ie correlate with the originally measured values in a linear manner, but not identical to these are. Since the actually measured metabolite concentration and the metabolite concentration in a broader sense correlate linearly or the metabolite concentration can be derived from the measured metabolite concentration in the narrower sense, both types of metabolite concentration data can ultimately be used equally as input for an MLP.

A "descriptive model" is a model that describes the current static and dynamic metabolic state of a cell.

A "predictive model" is understood to mean a model that the

Predictions about a future static and dynamic metabolic state of a cell. In embodiments, the metabolic model used for flux analysis is a descriptive metabolic model.

A "metabolic hybrid model" is a combination of one

Mechanistic model representing reaction kinetic knowledge, and an empirical model, the measured values of a bioreactor with metabolic

States of a cell in relation sets. A metabolic flux analysis (MFA) is a computational method by which intracellular flows can be estimated from extracellular fluxes. The extracellular material flow here refers to the amount of substance that is absorbed or released by a cell per time. The intracellular flux of substances describes the reaction rate at which an intracellular metabolite is converted or formed. MFA includes

various approaches designed to determine the rate of metabolic reactions within a biological entity. Metabolism is a dynamic process and can be considered as a kind of "cellular phenotype", which is described by MFA.

A "batch bioreactor" is a bioreactor that is operated in a batch process ("batch process") or designed for operation in a batch process. The batch process is characterized in that all substrates (especially sugars and amino acids) are presented before being inoculated in the bioreactor. By contrast, gases and PH correction agents are also introduced into the system during the process. However, as an approximation, the batch method can be considered as a closed system with a constant reaction volume. Due to the gradual consumption of the substrates by the cells follows after one

initial unlimited growth phase (so-called exponential phase) a phase of stagnation (stationary phase) in which the growth and death of the cells are in equilibrium. Subsequently, the death phase follows with a decrease in cell density as a result of a strong lack of nutrients. Among the usual modes of operation of a bioreactor, the batch process is generally the least expensive and has the lowest risk of contamination. However, in the typesetting process usually no optimal product yield is achieved

Process duration is limited by the consumption of the substrates.

A "fed-batch bioreactor" is a bioreactor that is fed-batched

("Feed method") is operated or designed for operation in the fed-batch process. In the fed-batch process, additional nutrient solution is added during the process (often only after an initial batch phase). The too

Feeding may be continuous or in the form of one or more highly concentrated boluses (ie, pulse-like). As a result, an extension of the process time is possible in comparison to the typesetting process, as consumed Substrate can be postdosed. In addition, a better process control can take place. For example, inhibition phenomena and the formation of toxic by-products can be cushioned by continuously lower educt concentrations. Overall, by the fed-batch process, significantly higher cell densities and product yields than in the batch process can be achieved.

A "chemostat" ("continuous reactor") is a bioreactor to which a constant stream of nutrient solution is supplied and to the same extent cell and

product-containing reaction medium is withdrawn. Thus, that changes

Reaction volume during the fermentation is not. With a suitable choice of the volume flows, a steady state equilibrium is established in the reactor in which the cell density and the nutrient concentrations remain constant.

A "perfusion bioreactor" or "perfusion reactor" is a bioreactor in which a continuous stream of nutrient solution is added and a continuous stream of (mostly cell-free) reaction medium is withdrawn. Through targeted bleeding, a stream of cell-containing reaction medium can also be continuously removed. The selection of the volume flows is coupled to one or more process control parameters and adjusted so that the

Reaction volume in the reactor remains constant. Cell density and

Depending on the process, nutrient concentrations may be constant (based on cell density) or variable (similar to fed-batch mode).

A "split-batch bioreactor" is a batch or fed-batch type bioreactor operated to charge the reactor one or more times a substantial portion of its medium, e.g. over 10% or over 30% for the purpose of harvesting the cells contained therein.

The "PMI" (partial mutual information) is a

Data value that quantifies nonlinear direct dependencies of two parameters. In the context of some machine-learning approaches, eg neural networks, the PMI is a measure of the dependence between a random input variable X and a random output variable Y, taking into account already selected inputs. Various approaches for calculating the PMI of two variables are known, eg Sharma A (2000): "Seasonal to interannual rainfall probabilistic forecasts for improved water supply management: Part 1 - A Strategy for System predictor Identification ", Journal of Hydrology Vol. 239, Issues 1 - 4, 232-239.

Thus, a "PMI criterion" refers to a feature or characteristic that has been defined with respect to the PMI to make a decision. For example, a PMI criterion may be a limit whose overshoot or undershoot affects the course of a process.

Some embodiments of the invention will be explained in greater detail in the appendix which accompanies this application and the disclosure of which is a part of this application. In order to ensure consistency between the description and the appendix, the meaning of the variables as specified in the appendix variable listing was retained. With regard to the meaning of the variables, reference is made to the variable register of the appendix. Examples and embodiments as well as further explanations, which are described in the appendix, are freely combinable with the embodiments, examples and features described in the application text unless they are mutually exclusive.

Brief description of the figures

Embodiments of the invention will now be described in more detail by way of example, with reference to the figures, which show embodiments of the invention and individual aspects thereof, respectively

Represent embodiments.

FIG. 1 shows a flow chart of a method for the prediction of the

metabolic state of a cell;

FIG. 2 shows, by way of example, the process of obtaining information in several

Stadiums using various devices and data sources;

FIG. 3 shows a block diagram of a system for predicting the

metabolic state of a cell which may be used to monitor and / or control one or more bioreactors; Figure 4 shows a metabolic model of a cell having multiple intracellular and extracellular flows of intracellular and extracellular metabolites;

FIG. 5 shows the calculation of intracellular flows at several

successive times during the operation of a

bioreactor;

FIG. 6 shows flows of various metabolites in accordance with FIG. 4

represented metabolic model;

FIG. 7 shows a number of metabolite flows which successfully use the

Illustrate a method for generating biological knowledge;

FIG. 8 shows a number of metabolite flows which demonstrate the successful use of the

Illustrate a method for generating biological knowledge;

Figure 9 shows plots with lactate fluxes and glutamine concentrations;

FIG. 10 shows time courses of intracellular flows; Figure 11 shows several intracellular and extracellular flows at different

Time points during the cultivation of a cell culture;

FIG. 12 shows the highly correlated course of intracellular flows, which were calculated for the current time by means of descriptive MFA and were predicted by means of a combination of MLP and MFA for a future point in time;

Figure 13 shows input parameter values and output parameter values of an NN;

FIG. 14 shows a histogram of the obtained RMSE for intracellular flows in FIG

12 fed-batch fermentation runs;

FIG. 15 shows a histogram of the obtained RMSE for extracellular flows in FIG

12 fed-batch fermentation runs;

Figure 16 shows 12 plots, each with a predicted extracellular

Metabolite flux and two for identical cell clones in two different bioreactors (fed-batch and split-batch) measured extracellular Metabolitflüssen;

FIG. 17 shows eleven plots with two curves each, all obtained for a fed batch bioreactor by means of two different calculation methods; FIG. 18 shows eleven plots with two curves each of calculated extracellular flows of a cell clone ZK1;

FIG. 19 shows eleven plots with two curves each of calculated extracellular flows of a cell clone ZK2;

FIG. 20 shows eleven plots with two curves of calculated intracellular flows

a cell clone ZK1; and

FIG. 21 shows eleven plots with two curves each of calculated intracellular flows of a cell clone ZK1.

FIG. 1 shows a flow chart of a method for the prediction of the

metabolic state of a cell cultures of CFIO cells according to one embodiment, which method is equally suitable for other cell types.

Step 102: Modellerzeuauna

The processes in a bioreactor can be described mathematically. First, the mapping of changes with time of relevant substance concentrations or amounts in the reaction medium should be considered (e.g., courses of

Substrate amounts, product quantities, cell densities). The formulation is based on mass balances and consists of a term that describes the reaction of the substance and a convection term, the possibly existing

Includes material flows into and out of the reactor. It applies generally (see [51], section 4.2 in the appendix):

or in formulas: where ⁷⁷¹ denotes the amount of substance in the reaction medium, ¹ the process time, VU or ^v a the volume flow of the inlet and outlet, or. the concentrations of the respective substance in the inflow and outflow and Q the amount of substance that is converted per time and volume. In the case that amounts of substance extracellular

Metabolites are considered, the reaction term comprises primarily the uptake or release of the substance by the cells. Should the cell density in the

Fermenter, it includes the formation and death of cells.

According to embodiments, the metabolic model is based on the assumption according to the above equation that the time course of the amount of substance can be differentiated. This is justified provided that all incoming and outgoing flows are continuous. In the case of bolus supplements or sampling during fermentation, progressions are obtained which are piecewise continuous. The above equation then applies to the regions between the discontinuities.

Numerous mathematical models for process description, control, and optimization can be specified on this differential equation. These are playing an increasingly important role as a result of the growing desire to better understand bioprocesses while saving money and money

to improve time-consuming laboratory experiments in silico. Models that rely solely on such material balance and do not describe intracellular events are called black box models. You can not explain dependencies between the considered processes mechanistically. For this it would be necessary, the To model metabolism as a link between the different extracellular substances.

Ultimately, the metabolic model should enable a metabolic flux analysis to be carried out so that the extracellular fluxes on the intracellular flows can be inferred using the model. While easy-to-implement methods are usually established for determining cell density and measuring extracellular substance concentrations, the observation is

intracellular reaction rates much more complex. To appropriate

To circumvent experiments, the metabolic flux analysis (MFA) was developed - a computational method by which the intracellular flows can be estimated on the basis of the extracellular flow. The extracellular material flow here refers to the amount of substance that is absorbed or released by a cell per time. The intracellular flux is the amount of substance that is converted in an intracellular reaction per time and cell.

In order to be able to carry out the metabolic substance flow analysis, a (preferably or generally simplified) biochemical,

stoichiometric metabolic network of the considered organism, which includes the most important intra- and extracellular reactions. Extracellular reactions - analogous to rivers - are those in which metabolites are taken up or released by the cell.

It is assumed that the network consists of reactions of which ^k m is measurable and therefore known (this is usually the extracellular

Reactions), and from 1 intracellular metabolites. The reactions can then be in a stoichiometric matrix AE hold on, in which the

stoichiometric coefficients (negative for starting materials, positive for products of the individual reactions), the lines corresponding to the different metabolites and the columns corresponding to the reactions. The extracellular metabolites are omitted here. A concrete example can be found in section 2 in the appendix.

The material flow of the / -th reaction is denoted by. Continue to be the

Amount of the ί-th intracellular metabolite in a single cell. If the flows and metabolite quantities are combined to form vectors v and m, the following applies:

The equation states that the temporal change in the metabolite quantities in the cell results from the balancing of incoming and outgoing material flows.

The MFA is now based on the so-called steady-state assumption that the amount of intracellular metabolites remains constant, which means that the sum of the incoming flows for each intracellular metabolite is the sum of the intracellular metabolites

outgoing corresponds. This simplifies equation (2.2)

Divide the vector v into the subvector of the known (measurable)

and the subvector of the unknown flows and the matrix A

corresponding to the sub-matrices and so that applies becomes equation (2.3)

The determination of the unknown rivers using MFA can thus be done by solving a linear system of equations.

The next step is to classify the system of equations. The obvious formulation for the solution of the equation system

is usually not applicable, since the matrix A _{u is} usually not invertible. In such cases, the solution space can be infinitely large or

Inconsistencies can occur so that no solution exists. The following terms were used by van der Heijden et al. (1994) introduced the

Classify system of equations or substance flows according to criteria of solubility and consistency [17, 39, 40], and which according to embodiments of the invention were also used for the creation of the model:

Assertiveness: The system (2.4) is underdetermined, if applicable. In

In this case, not all unknown flows can be clearly calculated because the metabolic network contains too few restrictions. is so the system has at most one solution and is called

determined designated.

Redundancy: If so, the system is redundant. This means that

linearly dependent lines are present in A. Due to measurement errors in the

Determination of or inaccuracies in the metabolic network model leads

this usually to an inconsistent system for which no solution exists (it then applies the latter term representing the extended coefficient matrix). If the system is not

redundant and therefore always consistent.

Calculability: A flow from v _u is calculable if it is calculated using

Equation (2.4) can be calculated unambiguously, otherwise not calculable. This is based on a consistent system. If the system is underdetermined, there is at least one flow that is not calculable.

Balance: A river out means balanced, if its value has an influence on the consistency of the system, otherwise unbalanced.

Balanceable flows occur only in redundant systems.

As mentioned previously, MFA systems are often underdetermined and / or redundant (with one system at a time being underdetermined and redundant). A solution can then be formulated using the Moore-Penrose pseudoinverse A *, which is defined for all A _u :

In the case of sub-determinism, this expression yields one of the infinite solutions to the equation system; inconsistency results in a least-squares solution.

In the following, methods shall be presented which can be used to check which of the unknown flows are still calculable in the case of underdetermination and which of the measured flows can be balanced in the case of redundancy. Identification of calculable rivers

It is the von Klamt et al. published procedure for the identification of calculable rivers [17]. Be a matrix whose columns are a base of the nucleus of form. It then applies:

Every vector can be represented as a linear combination of the basis vectors. So there is an a so that applies

Equation (2.4) can be extended to:

where a is arbitrary. The left-sided application of

Pseudoinverse then yields:

By varying the vector a, one obtains the space of the solutions of the

Equation system (2.4). t has a rank greater than 0 if and only if the system is underdetermined. It can be concluded that the calculable flows in v _{u are} exactly those to which a variation of a has no influence. This is the case if and only if the corresponding row in the matrix T is a zero row.

Preferably, according to embodiments of the invention, an identification of balanceable flows takes place in the next step.

For a non-redundant system, inserting (2.5) into (2.4) returns the

formulation

with the redundancy matrix (see references [17, 39] in

Appendix).

However, if the system is redundant then equation (2.6) is only for specific ones

satisfies, which is equivalent to the solvability of equation (2.4). The column-wise notation

makes clear that a measured flow has no effect on the

Solvability of equation (2.4) if the corresponding column vector of R is the null vector. in the above formulation denotes the / the column vector of

and the / th measured flux.

The following presents ways to treat underdetermined and redundant systems or corresponding metabolic models according to embodiments of the invention.

Treatment of Underdated Systems If the system is under-determined, that is, infinitely many solutions, there are several ways to modify the probia to arrive at a definite solution. One option, if possible, would be to extend the constraints generated by extending the metabolic network model

aggravate and thus create a specific system. Furthermore, one can try to increase the number of known flows by additional experimental quantifications. The determination of intracellular flows, for example, by Labeling experiments are accomplished (see reference [43] in the appendix).

If you want to avoid this experimental effort, the common method of Flux Balance Analysis (FBA) is recommended. Here is a to be optimized

Defined objective function, which is selected according to biological plausibility and which depends on the material flows. For example, it can be assumed that the

Host organisms can tailor their material flows to maximize their growth rate, as this represents a significant evolutionary advantage. If the objective function is denoted by F, the general formulation for the FBA results:

Formulating as an optimization problem with equation constraints usually results in a clear flow distribution as a solution. Lie

Information on irreversibilities of reactions before, can be admissible Range (Allowable range: set of points for which all constraints of the optimization problem are met) by the additional ones

Ungleichungsbedingungen further restrict, wherein denotes the vector of all irreversible rivers.

The main problem with the FBA is the correct choice of the objective function on which the solution depends heavily. It is quite conceivable that cells during a fermentation change their biological target (see reference [33] im

Appendix).

Treatment of a redundant system

In the case of a redundant system of equations, there are due to

Contradictions in the balanceable rivers usually no flow distribution, which solves the equation (2.4). Even in a completely correctly defined network model, inconsistencies usually arise due to measurement errors in the determination of be caused. The actually measured flow values are to be discussed below be clearly distinguished from the true measurable values. In the literature ma finds various methods with which an approximate solution ^ a redundant MFA problem is calculated:

One possibility would be to set and select the vector such that

the mean square distance between the vectors and ^~

is minimized. The formulation as optimization problem is then:

The associated solution meets the steady-state assumption only approximately. If the system is determined, then

As a second possibility, the determination of a least-squares solution for the vector ^v , which determines the relative quadratic distance to the measured ones, is appropriate Fluxes are minimized (see reference [39] in the appendix) while satisfying the Steady-Sfate condition. The wording is:

Again, irreversibilities can be included as additional constraints.

The second method differs from the first one in that its solution satisfies the steady-state assumption (the measured flows are then called "balanced"). For this, the values of the measurable flows in the solution only approximately coincide with the actually measured values. This can certainly be considered useful if the steady-state assumption is given a higher reliability than the measured values for the flows, which are always faulty. In the solution of (2.8), all those measured flows that can be balanced are adjusted. The unbalanced, however, remain unchanged. A generalization of the just described method for the adaptation of balanceable flows results from a more statistically motivated one

Approach explained in the references cited in the appendix [39, 40]. It is based on a Weighted Least Squares approach:

Let v _{m be} again the vector of the measured, error-prone flows and let v _{m be} the vector of the corresponding true values. The measurement error vector S denotes the difference between the true and the measured values:

It is assumed that d is normally distributed with expectation vector o and

covariance matrix

Since it is unknown, a plausible, experience-based estimate of the covariance matrix must be used for the further calculations. The goal is now an estimate of which complies with the steady-state assumption and at the same time closely, taking into account the information on covariances

should be. This is achieved by using the Mahalanobis distance. The optimization problem is:

The solution v _{m of} the optimization problem has the shape

(see reference [23] according to Appendix). I is the unit matrix. R denotes

Here, the reduced form of R, which is generated by elimination of linearly dependent lines and therefore has full rank (this is not unique). It can be generated by multiplication with a non-square matrix r, which performs the corresponding row transformations:

These balanced values can now be used in Equation (2.5) to calculate the unknown flows.

Assuming that the measurements of the flows are independent (i.e.

covariance matrix be diagonal) and the standard deviation of

proportional to the amount of the measured value with uniform proportional constant b> o, the optimization problem (2.9) is equivalent to (2.8). It is true that:

The formulation (2.9) offers the advantage over (2.8) that the covariance matrix can be flexibly adapted to the quality of the measurement data. In the

Balancing rivers whose readings are considered unreliable will allow larger changes in values than likely to be measured more accurately. Before illustrating the application of the MFA in bioprocessing issues, the following section will first discuss the statistical validation of metabolic metabolic models as carried out according to embodiments of the invention, since these are based on the

based on explained considerations.

Validation of the biochemical metabolic model

To date, the quality of the postulated biochemical network model has not yet been addressed. However, it is obvious that insufficient quality in its formulation can lead to severe deficits in the results of MFA. To create a meaningful model as a compromise between high informative value and maximum simplification, are

Validation methods needed. In a publication by van der Heijden et al. from 1994, statistically motivated tests are presented and the detection of possible systematic sources of error is explained (see reference [40] of the appendix). The investigations are based on the analysis of balanceable flows and their influence on inconsistencies in the model. They are therefore only applicable if there is a redundant system.

A test for the evaluation of inconsistencies

In the previous section in Equation (2.11), the reduced form R ^{f of} the redundancy matrix has already been introduced. The residual vector is defined by

In a redundant system usually applies. The covariance matrix of can be calculated by

It is thus dependent on the covariance matrix of the measured fluxes, which takes into account the uncertainties in the measurements. For the test, the test statistics are used, whose observations are given by: It can be shown that the test statistic of a subject (cf.

[40] in the appendix). The degrees of freedom correspond to the rank of

Overall, the following hypothesis test is obtained: Tests

The inconsistency of the considered metabolic model is not significant against The inconsistency of the examined metabolic model is significant: denotes the -quantile of the distribution with

Degrees of freedom. Detection of possible, systematic error sources

If the previously defined test indicates inconsistencies, this may be due to an underestimation of the measurement noise, which is reflected in the matrix C and thus influences the outcome of the test. In reference [40] of the appendix, three further possible sources of error are discussed:

Systematic measurement errors: the measurement of the / -th flow is having a systematic mistake afflicted.

Absence of an Important Reaction in the Metabolic Network: One () important reaction is missing in the network model; the stoichiometric matrix A would have to be extended by another column. The corresponding river was

designated.

Incorrect definition of a reaction in the metabolic network: The / -te

Reaction is wrongly defined; instead of the column vector in the stoichiometric

Matrix should stand the vector.

The investigation of the error is based on the structure of the residual vector e. For each of the above-mentioned errors, a characteristic comparison vector v can be defined, the direction of which approximates the direction of e, as long as the respective error source actually exists. A statistical test assessing the similarity between the directions of the vectors is also presented. The length of £ gives an indication of the size s of the error. In the following table the corresponding comparison vectors are listed. Here r L denotes the / th column vector of R ^f . The derivation of the

Comparison vectors should be demonstrated here only for the first of the listed error sources. For further cases, see [40].

It applies

In the case of a correctly defined network model, E [s] = o. If, however, the true value for the / -th flow deviates by the systematic error p of

measured from, ie is

so applies

It is therefore to be expected that e and rL have the same directions.

Table 1: Comparison vectors and associated error quantities for three different error sources.

To evaluate the similarity between ^£ and ^v , the test statistic becomes used, the -distributed with a degree of freedom of

The following hypothesis test assesses the similarity of and

Teste The vectors and are similar

against The vectors are not similar:

The statistical derivation can be found in the appendix of reference [40] of the

Appendix.

According to embodiments, the metabolic model created according to embodiments of the invention comprises a network which should comprise the central intracellular material flows and nevertheless should have as little complexity as possible. The model exemplified here is based essentially on the

Network stoichiometries proposed in the following publications: Altamirano C, Illanes A, Becerra S, Cairo JJ, Godia F (2006): "Considerations on the lactate consumption by CHO cells in the presence of galactose", Journal of

Biotechnology 125, 547-556; Lianeras F, Pico J (2007): "A procedure for the estimation over time of metabolic fluxes in scenarios where measurements are uncertain and / or insufficient", BMC Bioinformatics 8: 421; and Nolan RP, Lee K (2011): "Dynamic model of CHO cell metabolism", Metabolic Engineering 13, 108-124.

However compartmentalizations of the cells were not considered. Due to their variety, not all reactions involving redox and energy equivalents could be included in the metabolic model. Therefore were NAD (P) H and ATP are not included in the formulation of stoichiometry. In addition, some metabolic branches were not considered in detail but integrated into biomass formation (eg the pentose phosphate pathway). In most cases, the formulated reactions are a summary of several

cascaded biochemical reactions without branching, which after steady-state assumption should have identical material fluxes

(For example, only a few intermediates of glycolysis or the citric acid cycle are explicitly listed).

The biomass balance was taken from the above publication by Nolan (2011), as well as the conversion of the living and total cell density into the unit mol / l. The formulation of the stoichiometry for product formation follows from the amino acid composition of the target protein. Also regarding the

Reversibility of the reactions were the o.g. Publications.

The resulting metabolic model is shown in detail in FIG. This shows in Figure 4a, a biochemical network model of intra- and extracellular material fluxes of CHO cells] {Biochemical network model on intra- and extracellular material fluxes of CHO cells. It models the central fluxes of metabolism, which facilitate the transport and conversion of glucose (Glc),

Lactate (Lac), alanine (Ala), glutamate (Glu), glutamine (Gin), ammonia (NH3), aspartate (Asp), asparagine (Asn), serine (Ser), glycine (Gly), total cell density (BIO) , Product (Prod), glucose-6-phosphate (G6P), pyruvate (Pyr), alpha-ketoglutarate (AKG), malate (Mal) and oxalacetate (Oxa). The reversibility is characterized by the shape of the arrows.

In the table depicted in FIG. 4b, the individual stoichiometric reactions of the metabolic model are listed. Of these, reactions 1, 3, 9, 11, 13, 14, 16, 18, 20, 21, 22 and 23 are extracellular. The model includes 13 intracellular metabolites for which the steady-state assumption is to apply. The index "e" refers to extracellular substances. The reversibility / irreversibility of

Reactions are indicated by the reaction arrow.

The network model contained in the metabolic model should encompass the central intracellular material fluxes and still minimize it

Complexity. The wording used in this work is based on Essentially on the network stoichiometries that are mentioned in the above

Publications of Altamirano et al. (2006), Lianeras et al. (2007) and Nolan et al. (2011). However compartmentalizations of the cells were not considered. Due to their variety, not all reactions involving redox and energy equivalents could be included in the metabolic model

be recorded. Therefore, NAD (P) H and ATP were not included in the formulation of stoichiometry. In addition, some metabolic branches have not been considered in detail but have been integrated into biomass formation (e.g.

Pentose phosphate pathway). In most cases, the formulated reactions are a summary of several cascaded biochemical reactions without branches, which should have identical material fluxes after steady-state assumption (for example, only a few intermediates of glycolysis or the citrate cycle are explicitly listed).

The biomass balance was taken from the above publication by Nolan (2011), as well as the conversion of the living and total cell density into the unit mol / l. The formulation of the stoichiometry for product formation follows from the amino acid composition of the target protein.

Also with regard to the reversibility of the reactions, the o.g. Publications.

To the metabolic network shown in the above-specified table, the stoichiometric matrix A was formulated. The columns which correspond to the known (in this case extracellular) material fluxes have been combined to form the sub-matrix A _m , the remainder to the sub-matrix A _u .

The characterization of the metabolic network can e.g. according to a scheme, which is shown and explained in the Appendix as Figure 4.2.

The metabolic network can be validated and modified as described in the appendix.

Thus, a metabolic model 402 of CHO cells was provided as shown in FIG. The metabolic model involves a variety of intracellular 410 and 408 extracellular fluxes, using the metabolic model specifies at least a stoichiometric relationship between an intracellular 406 and an extracellular 404 metabolite.

The following steps 106-112 are performed for a plurality of times during the culture of a cell culture in a bioreactor. In this case, a profile of actually measured extracellular material flows and for the next time (after expiration of an interval of defined length, for example 24 h) predicted extracellular flows can be created. The deviation of these two profiles from each other indicates the quality of the prediction.

Step 106: Receiving Measured Values

In one embodiment, a sample is taken at several times during the culture of a cell culture in a bioreactor 208 of this cell culture and automatically or manually transferred to one or more analyzers 250, as shown in Figure 2. The analyzer may be a system of one or more analyzers, for example, a Thomaskammer or an optical counting station for determining the cell density. For example, to determine the concentration of individual amino acids, a high performance liquid chromatograph or other suitable method known in the art may be used. The sampling can be done for example in an interval of 24 hours. The measurement data thus obtained are transmitted to a data processing system 252. For example, the data processing system 252 may be a computer that monitors and / or controls one or more bioreactors as a control unit.

According to some embodiments, at least some of the measured values, for example the cell density, are also determined by way of corresponding sensors of the bioreactor 208 themselves and transmitted to a data processing system 252.

In addition to the concentrations of selected extracellular metabolites that are known or expected to have some predictive power with respect to the concentration and flow of this or any other extracellular metabolite at a future time point, others may as well

Input parameter values are determined, in particular the current time, the current cell density, and optionally other parameters such as the LDH concentration that can be used as a correction factor for the lysed cells not included in the cell density determination. The measurement data thus obtained empirically at a particular time can now be used to predict extracellular flows at a subsequent time, for example the next day, using an MLP, as described in the subsequent step.

Step 108: Enter the Measured Values in a Trained MLP

The data processing system 252 includes an MLP, for example, a neural network (NN) or a cooperative system of multiple neural networks, which has been trained to provide one or more extracellular responses based on input parameter values measured at a particular time (in particular, extracellular metabolite and cell density) Rivers of the metabolic model 254 predicts or estimates. For example, the data processing system 252 may include program logic that automatically transfers the measurement data obtained at a time to an MLP trained on test data sets obtained from cell cultures of the same type of cell that also includes the cells of the cell culture metabolic state at the future time (for example, next day) should be predicted. Step 110: Predict future intake rates and delivery rates through the MLP

Via the neural network are then in a one-step prediction

based on currently measured concentrations of extracellular metabolites at the time extracellular flows at any chosen future time predicted or estimated. Thus, in this step, the MLP will respond in response to the input of the measured

Input parameter values are calculated and returned to the MLP from this one or more extracellular flows for the future time point.

Optional: prediction of extracellular metabolite levels by the

MLP The extracellular fluxes estimated via the neural network also become one-step predictions of embodiments of the invention

Metabolite concentrations at any chosen future time starting from the current concentrations Used at the time.

For this, equation (4.1) of the Appendix is resolved to c 2. Because the

Live cell density of the future time is unknown, it is replaced by the currently measured and possibly corrected cell density. The reformulated equation is then:

If feeds are made during operation of the bioreactor, they should preferably be in and taken into account

become.

Step 112: Perform an MFA

Based on the uptake rates and delivery rates of the extracellular metabolites (extracellular flows), as predicted by the MLP for the future time, then, according to embodiments of the invention, a metabolic flow analysis is performed, including intracellular flows and stoichiometric equations as described in US Pat formulated in the metabolic model. Since in the metabolic model extracellular and

Since intracellular flows are coupled to one another via one or more intracellular metabolites, it is mathematically possible to use the predicted ones

extracellular fluxes also predict the intracellular fluxes at the future time. Corresponding program routines for performing metabolic Flow analyzes can be implemented in Matlab and other existing software solutions on the market.

In the prediction of intracellular fluxes, both the prediction of the MLP trained on dynamic empirical data and that in the

metabolic model specified knowledge about stoichiometric

Considering correlations and reaction equations, this predictive step can also be described as a prediction of a hybrid model relationship.

The coupling of the results of the predictions of the MLP with the information of the metabolic model in the course of the substance flow analysis can for example be implemented as follows:

After prediction of extracellular fluxes in the subsequent time interval using the neural network, a metabolic flux analysis is performed

performed to estimate the flow distribution in the next time interval. First, an estimate of a covariance matrix of the flows of the metabolic model is created:

shows that both the quality of the measurements and the estimates of the measured values by the neural network are included in the covariance matrix. However, the reformulation can not facilitate the estimation since the

Differences and not independent random variables

represent and thus the expected value in the middle term of the right side is unknown. Therefore, the term used in this work is according to the Equation 4.5 in the appendix, which will be explained below, only to be understood as a rough approximation to the actual covariance matrix.

According to embodiments, the covariance matrix is chosen to be a diagonal matrix for the purpose of predicting future intracellular flows since the different metabolite flows are estimated over separate networks and therefore the errors can be considered largely independent of each other. The

By definition, diagonal entries should reflect, or at least be proportional to, the variants of the estimated flows' errors (the

Proportionality Factor plays no role in solving the MFA problem as shown in Equation 2.9 of the Appendix).

In contrast, embodiments of the invention for descriptive metabolic flux analysis of the current metabolic state of a cell use a covariance matrix in which the diagonal entries do not

depending on the amount of the rivers.

se the jth measured predictive flow in the nth time interval. The

Covariance matrix was formulated as a diagonal matrix and has the structure:

The diagonal entries were used as the medians of the quantities

selected. It is assumed in this embodiment that the "time interval in the training data set" is identical or similar to the time interval to be used for the current prediction. Then, a flow analysis is performed on the basis of this covariance matrix, as is known per se in the prior art. The covariance matrix serves to balance the flows according to equation (2.10) of the appendix. Although Equation 2.10 of the appendix refers to the descriptive MFA in which, according to the

Appendix described another covariance matrix was used. The balancing or calculation of the flows by MFA takes place in the case of the predictive MFA in the same way.

Optionally, an error analysis of the model based e.g. on a Gaussian error propagation as described in Section 4.7.8 of the Appendix.

Generation and training of a neural network

Compared to the specification of reaction kinetic models for the prediction of future material flows, the use of trained MLPs has the advantage that their generation is generally easier and faster in a semi-automatic process. An example of how an MLP in the form of an NN can be generated through training is described below. a) Cultivation of several training cell cultures

In order to obtain the broadest possible database for the training of MLP, preferably several training cultures are cultured in several bioreactors.

Preferably, these bioreactors include one or more fed-batch reactors and one or more other bioreactors from other types of reactors.

In one embodiment, there were eight fermentations of one clone

recombinant CFIO cells on a one liter scale. The

Initial conditions (volume, media composition,

Inoculum concentration) were chosen identically for each bioreactor, but different modes of operation were used:

• A bioreactor was batched until cell viability dropped below 50%.

• In a second bioreactor, the batch process was also used first. Towards the end of the exponential growth phase, a Partial harvest performed and the reactor filled with fresh medium, so that similar conditions as at the beginning of the fermentation were achieved (in terms of volume and inoculum concentration). Subsequently, the batch was continued in a batch process (so-called split-batch process).

• The fermentation in the remaining six reactors was carried out as a fed batch with initial batch phase. Both a continuous feeding and a pulsatile nutrient addition took place towards the end of the fermentation. The fed-batch fermentations differed in the

Glucose concentrations in the medium, which were adjusted by different strategies in the feeding. In the following versions, the Fed-batch approaches will often be numbered. According to this numbering, in the first two approaches a short full glucose limitation occurred in the second half of the process before the bolus additions were made. The third and fourth approaches experienced the same limitation, but the subsequent boluses adjusted higher glucose concentrations. In the fifth and sixth approach there was always a positive minimum concentration of glucose.

Temperature, pO 2 and pH were kept constant throughout the fermentation. Regular sampling took place throughout the process. The samples were examined for their live and total cell density, using a staining method that distinguishes between living and dead cells. Lysed cells were not detected. Furthermore, the content of various substances in the reaction medium by means of a COBAS INTEGRA analyzer or through

High performance liquid chromatography examined. These included glucose, lactate and ammonium, as well as the amino acids alanine, glutamine, glutamate, asparagine, aspartate, serine, glycine and the enzyme lactate dehydrogenase. The

Product concentration was also determined.

At the time of sampling, the volume of liquid in the bioreactor was measured by a fermenter scale. b) Determination of the network architecture

A neural network was created which was to estimate the mean flows of the extracellular metabolites between the current and the next sampling time from the current state in the bioreactor (so-called one-step prognosis). It was for some embodiments for each

extracellular material flow as a starting point to train its own network. When

Inputs were used to select the currently prevailing extracellular metabolite concentrations.

The neural network consisted of a two-layered perceptron with linear

Activation function in the output layer and sigmoid activation function in the hidden layer (see Figure 2.4 in the appendix). For the latter, both the sigmoid function and the hyperbolic tangent were tested. There were no significant differences visible to the results under

Using the sigmoid function were achieved. The training took place after the sequential gradient descent procedure (as exemplarily in the

Equations (2.16), (2.17) and (2.18) are illustrated in the appendix). c) Selection of input variables

According to preferred embodiments of the invention, to select the input parameter values, several or all of the extracellular metabolites named in the metabolic model are ranked according to their relevance for the prediction of the respective flow. Preferably, extracellular metabolites with redundant information content were not considered here.

A detailed description of the selection of input parameter values according to embodiments of the invention can be found in the description of FIG. 13. d) choice of training parameters, initialization of weights, and

Normalization of the data

Once the number of iterations and hidden neurons has been determined, the values for the initial learning distance in each network will now be in the output layer and chosen in the hidden layer. After every tenth of the total iteration number, the step sizes became 1/10 of the r_ 1 J_

Output value reduced. The initial values for the weights were generated on * - 10 ⁵ 10 J equally distributed random numbers. The input and output training data were standardized separately for metabolites, so that the adjusted values had the empirical mean 0 and the empirical standard deviation 0.5. The test data were compared with the means and standard deviations of the

Training data set transformed in the same way.

e) Choice of training and test records

The estimates were made by three different neural networks, which differ in the grouping of data in training and test records:

1. Network 1: The training data set of the first network consisted of the data from three of the fed-batch fermentations and the batch fermentation, the

Test data from those of the other three fed-batch fermentations and the split-batch fermentation. 2. Network 2: The second network had data from three of the fed-batch fermentations in the

Training data set and data from the remaining three fed-batch fermentations, batch fermentation and split-batch fermentation in the test data set.

3. Network 3: The third network training data set included the data from three of the fed-batch fermentations and the test data set the data from the remaining three fed-batch fermentations. f) One-step prediction of extracellular metabolite concentrations

The goal is to create a trained MLP that is as accurate as possible

Predicting the metabolism of a cell allows. If low-frequency data is generated during the generation of the training data set, it may happen that an averaging / filtering of the data would lead to an excessive loss of information. Therefore, in this case average extracellular fluxes should be approximated between two consecutive measurement points. The calculation is based on equation and will be explained below:

Since the batch and the fed-batch process should be considered and thus there is no liquid outflow, applies in any case The extracellular flow v of a component at time t, as already mentioned in section 2.2.2 of the appendix, is the amount of substance taken up by a cell per time.

is delivered. He is therefore given by

in which denotes the number of living cells in the reaction medium at time t. According to equation 2.1 in the appendix thus applies in the range between two

discontinuity

The concentration of the substance in the feed is constant over time or is assumed to be approximately constant.

Discontinuities may occur at sampling times as well as at bolus additions.

It should first be assumed that the nutrient addition has always been continuous and that therefore the quantities m and ^ "between two measuring points

differentiable. If one wants to determine the mean flow ^{/; r} 'between two consecutive measurement times and t ₂ , then one can estimate it - initially apparently trivially - by:

In this case, the indexed variables denote the value at time ti or

is the amount of substance which, via the inlet, in the time interval considered

Reaction medium is supplied and VCD denotes the living cell density. The

Operator ^ symbolizes the difference of the corresponding quantities between t ₁ and

The above estimate (4.1) will be explained mathematically in the following. They are obtained by integration over the time interval where the individual

Measurements are interpolated linearly between the two sampling times, and by additional Taylor development: For linear interpolation applies to all

This is

For the remaining integral:

It will be a Taylor development of the logarithms around to the first

Order is carried out and delivers:

Substituting this expression for the above integral yields the estimate of flux according to Equation 4.1 above or Appendix). In the event that between two sampling times and a bolus for

Time is added, which contains the substance considered is so

assumed that the mean flow between and and between and

is equal to. It can then be easily shown that in this case the equation (4.1) can be applied, wherein the amount of substance that is supplied via the bolus, is included in the expression.

Thus, at the time of a current sampling from a training bioreactor, both the concentrations of extracellular metabolites in the previous sampling and the computationally based on the extracellular flows calculated since the last sampling are known and can be transferred together as reference value sets to the MLP to be trained, the As a result, based on the concentrations of the extracellular metabolites measured for the last sampling, the calculated extracellular body learns to predict that the smallest possible deviation from the calculated extracellular fluxes exists.

The trained MLP or trained neural network can now be stored and for one-step predictions of extracellular metabolite concentrations at any chosen future time , for example, next

Day, starting from the current concentrations at the time

be used. In summary, the idea of training the neural network is based on the fact that the concentrations of extracellular metabolites are easily measurable and from which, at least in retrospect, extracellular flows can be empirically determined. By using the measured at a certain time

Extracellular metabolite concentrations as input parameter values and extracellular flows as calculated over the time interval between this current time point and a future time point based on the extracellular metabolite concentration difference as output parameter values can be used to train a neural network or in other machine-learning algorithms. that he can at least predict the extracellular fluxes for one future time. The determination of the cell density allows a transfer of the total concentration difference in the medium to the individual cells of the cell culture contained in the medium.

According to preferred embodiments, the uptake / release rate of extracellular metabolite (reaction term) is calculated from the difference between the total of the change in concentration of the substance in the bioreactor minus the substance fed to the bioreactor (convection term). It has been found that the response term in extracellular metabolites is determined primarily by the uptake or release into or by cells, so that the measured

Concentration changes of the extracellular metabolites can be essentially equated with the uptake rates or release rates of these extracellular metabolites in the cells or from the cells. However, in some embodiments which provide for substantial delivery or withdrawal of certain extracellular metabolites during operation of the bioreactor, a correction calculation may be made whereby, in calculating the uptake rates of these extracellular metabolites, those portions of the concentration changes are subtracted from the measured changes in concentration Supply or removal of extracellular metabolites are due to the culture medium or from the culture medium. However, this correction does not necessarily have to be made in the fed batch process. Even in the fed-batch process, the incoming and outgoing streams are continuous. In the case of bolus feeding or sampling during fermentation, gradients are obtained which are continuous over a long distance and thus differentiable piece by piece are. Thus, the assumption of a differentiable course is the

Concentrate changes also justified for fed-batch reactors.

Preferably, however, a correction of the measured cell density during training and / or in the prediction of the extracellular flows by means of the trained MLP is made. Except for the batch fermentation seems at the

Fermentation approaches a similar, approximately linear relationship between the LDFI concentration and the cell density difference to exist.

According to embodiments of the invention, a measured total cell density is recorded continulatively in a fermenter and displayed in a plot. In parallel, the cell density is calculated. The prediction may be calculated, for example, by a trained MLP trained according to embodiments of the invention, the measured cell concentration being another

Output parameter value was used. This "predicted" cell density is also plotted in the plot. According to embodiments of the invention, therefore, a first temporal profile of the measured cell density of a cell culture of a particular cell type is determined empirically and also a second temporal profile of one based on the metabolic model and the extracellular

Metabolite concentrations predicted cell density. Thus, a plot is obtained which includes a temporal profile of the measured and predicted cell densities and their deviation from each other.

It turns out that there is often a considerable difference between the measured cell density and the cell densities predicted by the MLP, especially towards the end of a cell culture. The discrepancy between measured and predicted cell density is referred to below as the "density discrepancy profile" and can optionally also be plotted in the plot. The

Density discrepancy profile represents the temporal profile of the appearance of lysed cells, which are not detectable as cells but still have an influence on the concentrations of extracellular metabolites. An empirical function is then created, eg using a linear regression line over the density discrepancy profile characterized by two parameters a ₀ and a _± , which linearly relates the density of lysed cells according to the density discrepancy profile to the LDH measured at a particular time

Concentration in the medium. This function allows an approximate conversion of the LDH concentration in the density difference and thus in the lysed cells. The corrected cell density is therefore the sum of the measured cell density and that based on the measured LDH concentration by means of the linear function

calculated number of lysed cells. In other words, the measured cell densities 4Gΐi are then supplemented by the density difference with the aid of this conversion:

where _M denotes the LDH concentration at the h-th time, and so on

was chosen that is a time when one

Fermentation was started.

With these corrected cell density values, according to preferred embodiments, the corrected biomass flows are calculated. This has the advantage that

the proportion of inconsistent flux distributions as a result of MFA is significantly reduced.

FIG. 2 shows, by way of example, the process of obtaining information in a plurality of stages, with reference to various devices and data sources.

Cells whose metabolic state is to be predicted at a future point in time are kept in a bioreactor 208, for example in a fed-batch fermentor. The fermentor may include or be operatively coupled to some sensors, for example, cell density sensors. Instead or in addition, samples of the cell culture may be taken regularly and transferred to one or more analytical devices 250. There the measurement of the concentration of extracellular metabolites takes place. The measured cell density, the measured extracellular metabolite concentrations as well as optionally the time and amount of the external supply of metabolites (eg glucose boli) are forwarded to a data processing system 252. This system 252 includes a metabolic model of the cell as well as an MLP that has been trained to predict extracellular fluxes based on currently measured concentrations of extracellular metabolites for the currently used cell type. The measurements that were received and measured at a particular time are transmitted as input to the trained MLP, which subsequently predict extracellular fluxes at a future date. In the course of a subsequent MFA, the intracellular fluxes for the future time are predicted based on the extracellular flows and the stoichiometric equations of the model. The complete model including the predicted extracellular and intracellular fluxes may be considered

"Snapshot" image 254 of the metabolic state of a cell can be output and / or stored via a graphical user interface.

If the bioreactor 208 includes a training cell culture, ie a cell culture from which data are regularly collected over a longer period of time to generate a training data set, then the data processing system 252 is additionally configured to extracellular flow based on a plot of the measured extracellular metabolite concentrations for the future time interval (ie the time interval from the current time to the next time for which a prediction of the metabolic state is to be made) to be calculated and passed to the MLP as the output parameter value during the training.

Figure 3 shows a block diagram of a system for predicting the metabolic status of a cell which may be used to monitor and / or control one or more bioreactors.

The system 200 may be a data processing system

act of various types. For example, it may be a desktop computer, a server computer, a notebook or a portable portable device of a user. The system 200 may be a control module that is part of a bioreactor or bioreactor system having multiple bioreactors, or that is coupled to such a bioreactor or bioreactor system. In the embodiment shown here, the system is coupled to three bioreactors 204, 206, 208 and configured to monitor in real time the metabolic state of the cells in the particular cell culture, as well as for a future time, for example, 12 hours or 24 hours in the future to predict. Each of the bioreactors may have one or more measuring devices or sensors for determining the cell density and / or metabolite concentration, or a mechanism that allows a sample name, so that the metabolite concentration by others Devices can be made by analysis of the sample. Preferably, the bioreactors 204-208 have various actuators such as

Example valves, pumps, metering units for boluses, stirrers, etc., which are coupled to the system and may receive and execute control commands from the system if necessary.

The system 200 includes one or more processors 202 and a first interface 210 for receiving measurement data from the one or more

Bioreactors. The interface 210 may serve as a direct interface to the

Bioreactors or as an interface to analyzers in which samples are analyzed by the bioreactors, or a graphical user interface that allows a user to manually or otherwise input the obtained measurement data.

In addition, the system 201 includes or is coupled to volatile or nonvolatile storage media 212. The storage medium may be, for example, memory, hard disk, or storage of a cloud service or a network storage or combinations of the aforementioned types of storage. The storage medium includes a metabolic model 214 of the cells held and propagated in the bioreactors, for example a model as depicted in FIG.

The storage medium comprises a trained MLP 218, for example a trained neural network, which is designed to predict one or more extracellular flows at a future time, based on the measured concentrations of several extracellular metabolites received at a particular time.

In addition, the storage medium includes program logic 220 configured to pass the received measurements to the MLP 218 to perform a prediction of extracellular flows. In addition, the program logic is designed to perform in real time a metabolic flow analysis (MFA) based on the metabolic model 214 and the predicted extracellular fluxes for future intracellular fluxes

predict. The program logic 220 may be in any one of

Programming language implemented, eg C ++, Java, Matlab or in form several program modules in different or the same programming language that are interoperable with each other.

Optionally, the storage medium may have multiple reference values and

Reference value ranges include. These reference ranges indicate acceptable or desirable intracellular flows of various intracellular metabolites. By real-time matching of the predicted intracellular flows with the reference values 216, the program logic 220 can determine whether the cells in one or more of the bioreactors are heading for an undesirable metabolic state, and optionally

Countermeasures by issuing appropriate control commands, which counteract the predicted trend, take to the respective reactor via a second interface 222. Alternatively or additionally, in this case, a warning can be sent to a user via a user interface 224, for example a display device, for example an LCD display. About the

Display device, the user can be informed about the predicted extracellular and intracellular flows and also about any predicted deviations of these rivers from desirable reference areas.

FIG. 5 shows the calculation of intracellular flows at several

successive times during the operation of a bioreactor. The upper plot 502 shows a profile of the concentration of an extracellular metabolite which was determined at six measurement points (one measurement per day). The middle plot 504 shows that a trained MLP predicts one or more extracellular flows from these measurements for a future time, respectively.

A comparison of the positions of the points in the upper and middle plots shows that the points in time of the collection of the measured data and the times at which the extracellular rivers were predicted in each case were about half a day

apart. That is, for example, when daily at 12:00 noon, extracellular metabolite concentrations are measured, these data are used to predict extracellular flows at midnight. The lower plot 506 shows this by performing an MFA on each of these

In future, predicted extracellular fluxes would be augmented by predicted intracellular fluxes. FIG. 6 shows various flows depicted in the metabolic model of FIG. The flows were based on the measured change in metabolite concentrations over a time interval and the measured

Cell density calculated and show that the rivers for different metabolites have a very different and sometimes characteristic course.

FIG. 7 illustrates the successful use of the method for generating biological findings. The plot of the glucose flow top left shows that in the profile of the glucose flow at about 0.65 the glucose flow almost stops. Effects of this lack of glucose can be observed with alanine, series and glycine: alanine becomes larger rates in the case of the limitations

which probably serves to increase the supply of pyruvate. In addition, there is a stronger conversion of serine to glycine, which is associated with increased ammonium formation.

A comparison of the flow patterns of glucose and product reveals certain

Similarities: both an interim decline of the rivers at an early fermentation time, as well as a later burglary, as a short-term no

Glucose was more present in the medium, can be seen in both courses. This slump is missing in the product flows of fermenters without glucose limitation.

Likewise one sees the effect of the glucose Boli, which the sensitivity of the

Glucose flow compared to the glucose concentration in the medium shows, even in product formation. Obviously, there is a very close relationship between these rivers. At the end of the fermentation, the amount of product was highest in those bioreactors without glucose limitation. The availability of glucose seems to be essential for effective product formation, a deficiency should be strictly avoided.

FIG. 8 illustrates another successful use of the method for

Generation of biological knowledge.

The so-called Laktatshift refers to the often observed in the cultivation of cell cultures effect that the lactate flow changes the sign of the positive to the negative. Numerous explanations for the lactic shift exist in the literature. Mulukutla et al. For example, postulate that the lactate shift is the result of regulatory mechanisms that are due to the increasing Lactatinhibition be set in motion. This biological hypothesis was checked by determining the lactate flow in several bioreactors over a number of measurements by means of the method according to the invention and for this the extracellular lactate concentrations were also measured. The corresponding results are shown in FIG. 9 for four bioreactors. It was found that by the time a reversal of the net flux direction of the lactate was observed, the lactate concentration in the different bioreactors was different. The mechanism postulated so far in the literature does not seem to be responsible for the lactate shift. Rather, it seems that the shift depends on the glutamine metabolism, as shown in Figure 6.2 of the appendix. A comparison of the courses of glutamine and lactate flows in several different

Fermenter types shows that the Laktatshift always with a complete

Consumption of glutamine is associated.

This observation has also been previously reported in the literature describing a particular CHO clone in which the shift only occurred when glucose was consumed (Zagari F, Jordan M, Stettier M, Broly H, Wurm FM (2013): "Lactate metabolism shift in CHO cell culture: the role of mitochondrial oxidative activity", New Biotechnology, Vol.30, No. 2). According to the invention, it is thus possible, by the repeated prediction of intracellular flows and by the comparison of these flows with other intracellular flows and / or

Concentrations of Extracellular Metabolites

Successfully testing and possibly rejecting cellular metabolism as well as identifying metabolic peculiarities of individual cell clones. Based on the results of the analysis it is obvious that most of the lactate formed in the common CHO clones originates from glutaminolysis. Interestingly, after the lactate shift, glutaminolysis first stops: there is no longer any glutamine uptake; rather, it is formed to a small extent, increasing its concentration in the bioreactor. The cells now seem to have adjusted their metabolism exclusively to the substrate glucose. It is striking that a second phase of glutamine uptake can be observed as soon as glucose reaches very low concentration ranges. At the same time, a short phase of renewed lactate production can be seen. FIG. 9 shows plots with lactate fluxes and glutamine concentrations of four bioreactors of different types, which compares the courses of

Glutamine concentration and lactate flux allowed in a given cell culture. The glutamine concentrations are characterized by solid lines and the lactate flows by dotted lines. The lactate fluxes and concentrations of the individual rivers and metabolites were previously normalized by the

To increase comparability. The arrows each indicate a reversal of the sign of the lactate flow.

FIG. 10 shows time courses of intracellular flows as calculated by means of descriptive modeling by MFA based on the metabolic model in each case for the current time (that is to say descriptively, not predictively).

In Section 2.2.2 of the Appendix, various methods are presented to treat redundant metabolic models. For the calculation of the flows, the optimization problem equation 2.9 in Appendix) was solved, thus obtaining a weighted least squares solution. This allowed the extracellular fluxes determined by the solution to the optimization problem to deviate from the fluxes directly calculated from the experimental data.

Equation (2.10) of the appendix was used to calculate the extracellular, balanced flows that were intracellular via equation (2.5) of the

Appendix determined, ie

For the formulation of the covariance matrix It is preferable to include the results of the model validation, which showed which standard deviations led to a largely consistent model and which measurements might be considered unreliable. Subsequently, the calculated flows were visualized: On the one hand, their time courses were plotted separately for metabolites, on the other hand, the entire flux distribution was mapped at selected times.

Figure 11 shows snapshots of intracellular and extracellular flows at various time points during the culture of a cell culture of CFIO cells.

In addition to the progressions of individual material flows over time can also

Snapshots of the total intracellular and extracellular flux distribution at the individual time intervals are considered. Figure 11 shows four metabolic states of cells from the fourth fed-batch fermentation. They come from the second, eleventh, 14th and 21st time interval. The four pictures show flow distributions during a fed-batch fermentation. As a reference the Gluk39osefluss serves in the second time interval. The product flow was multiplied by 10,000.

In the second time interval, ie in an early fermentation phase (phase I in the division shown in Fig. 11A), glucose is still present in excess, is rapidly transported into the cells and enters the Krebs cycle, and glutamine is also taken up Conversion to glutamate and further along with pyruvate via river v_8 in metabolites of the citrate cycle and in alanine Lactate and ammonia are released in larger quantities into the medium.

In the eleventh time interval (Phase III), shown in Figure 11B, the lactate shift has already occurred, i. Lactate is taken from the medium, metabolized to pyruvate, and then enters the Krebs cycle. The glucose consumption is reduced, glutamine is no longer absorbed. There is no more ammonium delivery. Biomass production has decreased, but product formation has increased.

Compared to the start of fermentation, the citrate cycle is almost unchanged, but the anaplerotic reaction of malate to pyruvate is reduced.

In the 14th time interval (also phase III), shown in Figure 11C, is the

Lactate concentration in the medium is very low, so even the intake

is restricted. Alanine is taken up and converted into pyruvate. Ammonium is increasingly released into the medium. Product formation has increased even further.

At the end of the fermentation (phase IV), shown in FIG. 11D, glucose is added in a bolus manner. Therefore, there is again an increased intake, the flow through the Citrate cycle is particularly strong. All other reactions are extremely reduced.

FIG. 12 shows the highly correlated course of intracellular flows, which were calculated for the current time by means of descriptive MFA and were predicted by means of a combination of MLP and MFA for a future point in time. A "descriptive MFA for the current time point" calculates the extracellular flows of the metabolic model from the measured concentration changes of the extracellular metabolites over the just past time interval and uses these extracellular flows as input to the MFA to calculate the intracellular flows for the current time point. On the other hand, "predictive" identification of intracellular fluxes is based on predicting the future MLP extracellular MLP and using the predicted extracellular fluxes as input to the NSA to predict intracellular flows at the future point in time.

Figure 13 shows some of the extracellular metabolites whose concentrations are used as input to predict extracellular flows.

For each extracellular metabolite flux (indicated in the top row of the table) that is to be predicted by the trained MLP and passed as output parameter values during training, in the column below, the input metabolites are of decreasing relevance to the flow estimation -Metabolites "listed. Bio "refers to the measured

Biomass, which is preferably specified in the form of a cell density, and "TZD" the value corrected by the LDH concentration. The PMI calculation used for

Determination of this relevance, the values from all training fermentations are In addition, the table contains the results of the cross-validation, in which the number of input variables, which was determined by hidden variables H and the iteration number. The input metabolites 1504, their concentrations as

Input parameter values entering the neural network are highlighted in yellow.

The data from several (for example 3) fed-batch fermentations as well as those from a batch batch served as a training data set. The data obtained from other fermenters of the same cell type formed the test data set. In addition, cross validations can be performed with other training / test record splits.

However, the list shown in the table does not necessarily conform to a biologically intuitive order: for example, glucose concentration plays a minor role, although it would be obvious that the most important substrate of CFIO cells has a major impact on some material closures. However, the PMI-based arrangement described below has only a limited biological significance, which can be attributed in particular to the following facts: the progress of the metabolite concentrations are to some extent via the metabolism

correlated with each other. Therefore, it is possible that many others may lose relevance to the estimate after selecting a metabolite, since much of the information contained in them is already described by the first metabolite. However, the selected metabolite may not be the one that, mechanistically, actually has an impact on the material flow, but is only strongly correlated with it. In fact, it could be observed that for a sample white change in the data selection used to calculate the PMI values, some arrangements were different. In some cases, for example, the position of glutamate and glutamine were exchanged, which are closely linked via the metabolism. However, today's selection based on biological intuition, as has been done in some publications, has sometimes yielded much worse predictions. This may be because the biological relationships are often not always known and the redundant information is selected. In addition, the study of literature for the selection of suitable input parameter values based on presumed biological relevance takes a long time. First, for each metabolite flow, the potential inputs were sorted by relevance using Partial Mutual Information (PMI), see Equation 2.19 in the Appendix). The PMI was used by the

Discretization according to equation 2.22 in the appendix). The core density estimator used is based on city block function, for the calculation of

Residuals became the Nadaraya-Watson estimator (see Equation 2.27 in

Appendix). For this a training data set was used. This can, for example, as in the present case, the measurement data of 8

Cell culture projects in 8 different bioreactors. The order of entries for the estimation of the j-th extracellular metabolite flux was carried out according to the following algorithm:

1. Grasp in the crowd all potential input variables together, here all current concentrations of extracellular metabolites. The amount contains all already selected input variables (it is empty at the beginning). Let be the output variable, ie the jth extracellular metabolite flow between the current and the future sampling time.

2. Calculate an approximation for the partial trans information between each

potential input in and ^ taking into account the elements in

based on the given, standardized data set. By the

Standardization had e.g. all input variables and the output variable the mean 0 and the standard deviation 0.5. This has distorting effects on the relevance of the sizes due to different

Orders of magnitude eliminated. Preferably, this standardization prior to the training of the neural network with respect to the measured values of the

Training data set as well as when entering current measured values in

Test method using a trained MLP performed on the currently obtained measurements.

3. Note the variable in which has the highest PMI value. Add them and remove them

4. Repeat steps 2 and 3 until ^{L is} empty.

According to embodiments of the invention, the PMI for the parameter X, e.g. a particular extracellular metabolite, with respect to the parameter Y, e.g. another extracellular metabolite calculated as follows:

and

Where g is the density function of the marginal or common distributions. The residuals contain only the information of X and Y that are not yet contained in U. The larger the value for G, the stronger the dependency.

An approximate, discrete version of expression 2.19) is:

Pairs of samples of the X and Y and the associated, -distributed residual pairs are. The im

Normally, unknown density functions can be approximated using kernel density estimators, which also use information from the N samples. These estimators provide - in simplified terms - a continuous density function, the in its course similar to the histogram of the samples. It arises from a weighted superposition of N core functions. These are in turn

Density functions, the bell-shaped and symmetrical around each of the

Sample values are. Among other publications, the Gaussian kernel and the city block function have been used to calculate the PMI.

In general terms, the estimate is the density of a q-dimensional

Random vector X with the samples using the

core function with bandwidth m:

With the city block function as a core function, this results in:

The common density distribution of two random vectors X and Y can be determined using a product kernel estimator

The choice of bandwidth has a significant impact on the quality of the estimate. The larger the bandwidth, the smoother, but also more detailed is the density approximation. In several studies and present, the choice is yours proved successful.

For the calculation of the redundancies according to Equations 2.20 and 2.21, the conditional expected value must generally be used for two random vectors X and U are estimated. The Nadaraya-Watson estimator can be used for this. This is based on the previously applied principles and can be derived as follows:

In the approximate step, the densities were approximated using kernel density estimators. The last equal sign results from the fact that

is.

According to embodiments of the invention, the selection is made on the principle of a wrapper: For each output variable, 70 networks with the 1 to 7 (according to PMI) most relevant inputs and with 1 to 10 hidden neurons were trained in initially 1000 iterations. For this purpose, a training data set was formed from a part of the entire training data set, which consists of the 8 monitored training fermentations. The remaining data served as test data. At each training, the value of the test error was calculated over the number of iterations recorded and its minimum value, and the associated iteration number were determined. Subsequently, the comparison of the 70 networks based on the total minimum test error. This resulted in the determination of the combination of the input variables for the estimation of the respective metabolite flow, as well as the corresponding iteration number and the number of hidden neurons.

The list of selected input parameter values, also referred to herein as "inputs" or "input variables," for various output parameter values

("Outputs", "output variables") is marked accordingly in FIG.

According to preferred embodiments, the selection of those takes place

extracellular metabolites whose concentration is to be used as input parameter values for training or feeding the trained MLP ("input metabolites"), according to purely statistical criteria, individually for each output metabolite, ie individually for each extracellular flow; should be predicted.

1st order of the input metabolites according to their "relevance":

First of all, all available input metabolites or at least all metabolites, which are present in the metabolic model as extracellular metabolites, are transferred to a "first quantity" and classified according to their relevance: It is determined by the PMI criterion which metabolite is the largest Statement A / predictive power for output (rate of extracellular

Metabolites whose extracellular flow is to be predicted - "output metabolites"). This metabolite is taken over into a "second set", which then sets the actual input parameter values. The first in this

Sorting step certain "relevance" or predictive power is not yet dependent on the metabolites in the second set.

2. Determination of the actual input metabolites:

Preferably, not all input metabolites are used as input in training the MLP or in using the MLP (this leads to very poor predictive power due to overfitting). It is therefore determined how many of the

most relevant input metabolites should be used. For this purpose, the Predictive power of the "x" most relevant input metabolites determined by a test data set, x is varied, and then the number x is selected with the best predictive power.

After the first sorting step and the transfer of the most relevant metabolite from the first into the "second amount" of the actual input metabolites (the concentration of which sets the input parameters of the MLP), the input metabolite within the metabolite remaining in the first set is repeatedly identified, the has the greatest predictive power with respect to the flow of a given output metabolite, taking into account the content of the second set. If the concentration profile of the metabolite with the highest predictive relevance within the remaining members of the first set correlates strongly with a metabolite already contained in the second set, this metabolite is usually not converted to the second set, as its predictive power is high However, its concentration profile does not provide any significant contribution over that of a metabolite already in the second set. Rather, its inclusion would only increase the amount of redundant information in the second set. Thus, if the metabolite in the first set can not be included in the second amount for these reasons, the metabolite with the next highest relevance score of the first set will not become an excessive increase in the information content redundancy

Concentration profiles of metabolites of the second amount leads, transferred from the first to the second amount.

Thus, a metabolite can be very meaningful for the rate of output of a given output metabolite, without being included in the second set. A takeover can be omitted, in particular, if the concentration profile of this metabolite correlates very strongly with that of a metabolite already contained in the second amount, so that it suffices to use only one of the two or its concentrations as input parameter value during MLP training and later when using the trained MLP. That If one chooses one of them as "relevant", the other one loses his

Meaning. This is recognized by the PMI criterion). This continues until all of the predictive power relevant and sufficiently independent input metabolites are included in the second set. For example, in the case of the Glu flow (4th column in the table shown in Figure 13), it has been found that better information can be obtained from information about the current biomass and the current glycine concentration than if only biomass were used as input used (too little

Information), or as if in addition to information about glutamine, aspartate, etc. fed into the neural network (too much information, overfitting).

FIG. 14 shows a histogram of the obtained Root Mean Square Error (RMSE) for intracellular flows obtained for several fed-batch fermentation runs: 12 fed batch reactors, each one CHO Cell clone, which differed from the cell clones of the 11 other reactors. The 12 cell clones were genetically engineered so that all 12 clones produce the same product, a bispecific antibody. Although the same DNA sequence in the

Transfection was used to generate the clones, the clones due to different insertion and / or different copy number of the integrated DNA sequences metabolic differences. Ten Fed batches were used to train the MLP (here: a neural network model), two Fed batches to test the model. These two Fed batches were used for the RMSE calculations.

The intracellular fluxes RMSE are calculated from a difference between the intracellular fluxes predicted by a combination of MLP and MFA and intracellular fluxes, which can be calculated from measured extracellular fluxes. RMSE is never negative, a value of 0 (almost never achieved in practice) would indicate a perfect fit of the predicted to the measured data. In general, a lower RMSE is better than a higher one. RMSE is the square root of the average squared error. The

Effect of each error on RMSE is proportional to the size of the squared error; bigger bugs have a disproportionate impact on RMSE.

The same type of medium and nutrient solutions were used for the 12 fed-batch reactors to make the bispecific antibody. The However, the medium and the nutrient solutions differed from the medium and the nutrient solutions for the bioreactors or cell cultures whose metabolic

Footprint is shown in Figure 5.9 of the appendix.

The MLP, here a neural network (NN) has been recalibrated to the new record, i. the NN, which has already been trained once on data obtained from the bioreactors shown in Figure 5.9 of the Appendix, was "retrained" or retrained for data from the 12 fed-batch reactors for production of the bispecific antibody. The models were made in Python with the help of

Machine Library - Scikit learn - created. The hyperparameters of various NN models were optimized using a raster search function. Models and

Hyperparameters that best fit the data were stored and used as a "re-trained" NN for future predictions of extracellular and intracellular flows of cell cultures in the 12 bioreactors.

FIG. 15 shows a histogram of the obtained RMSE for extracellular flows obtained for the 12 fed-batch fermentation runs mentioned in FIG. All RMSE values are normalized to the error obtained for the glucose metabolite. RMSE values for extracellular flows are calculated from a difference between measured external flows and extracellular fluxes predicted by a combination of MLP, especially a neural network, and MFA.

It was observed that the RMSEs of both the intracellular and extracellular flows obtained for the 12 fed-batch reactors for production of the bispecific antibody were in the same range as the RMSEs obtained for other cell cultures or other cell clones ( see Master thesis - "Appendix" - page 67, Figure 5.9, the protein product of this cell culture is an antibody fusion protein, the 12 cell cultures are CHO cell cultures, the RMSE of the external and internal flows (as measured by the predicted ) were between 10-35%.

Figure 16 shows a comparison of the predicted extracellular metabolite fluxes (black line) with two for two identical cell clones in different

Bioreactors (fed-batch and split-batch bioreactor) measured extracellular metabolite fluxes (two gray lines) for different metabolites (amino acids). at The cell clone in both bioreactors are recombinant CHO cells (monoclonal) that synthesize a particular antibody fusion protein. The curves show that the method according to embodiments of the

Invention is able to predict the flows of extracellular metabolites for both fed batch (fb) and split batch (b) very accurately. In detail, Figure 16 shows a comparison of the extracellular measured fluxes with the predicted extracellular fluxes in different combinations

Training data / test data set. The expression "fb / fb + b" is called fed-batch data set as training data and fed-batch + batch as test data record. In each case, a run from the test data record is shown.

FIG. 17 shows several plots with two curves each, all for a fed batch

Bioreactor with a cell clone for the preparation of the antibody fusion protein were obtained. The curves consisting of dotted lines

("Descriptive MFA") were obtained by using measured concentrations of extracellular metabolites as input to metabolic flow analysis (MFA) to calculate various extracellular flows, each corresponding to one of the 11 plots. The solid line curves (\ "NN-MFA \") were obtained by using extracellular metabolite concentrations measured at a given time t0 as input to an MLP (eg, NN) to predict extracellular flows at a future time t1 future predicted extracellular fluxes were used as input for metabolic flux analysis (MFA). A comparison of the two curves of each of the 11 plots thus shows that the values predicted using MLP + MFA for a future date are very high

Have agreement with values obtained with a static MFA model using measured time-now concentrations of extracellular metabolites.

FIG. 18 shows several plots with two curves each, all for a fed batch

Bioreactor with a first cell clone ZK1 were obtained for the preparation of a bispecific antibody. The dotted line curves ("NN + Descriptive Extracellular MFA") were obtained by using extracellular metabolite fluxes measured at a given time tO as input to an MLP (eg, NN) to detect extracellular flows on one predict future time t1, and these future predicted extracellular fluxes were used as input to metabolic flux analysis (MFA). The outlined lines ("Descriptive MFA extracellular") were obtained by taking currently measured

Extracellular flows of extracellular metabolites were used as input to metabolic flow analysis (MFA) to calculate various extracellular flows, each corresponding to one of the plots shown in FIG. The extracellular fluxes were normalized for each plot and for each day in terms of the concentration of glucose. Here, instead of measured extracellular metabolite concentrations, current extracellular flows, which were calculated on the basis of the metabolite concentrations measured currently and at a past point in time, were used as MLP input. Thus, a metabolite concentration in the broader sense was used as input. It was observed that alternatively, when the measured metabolite concentrations were used as input, the predictive results of the MLP were ultimately substantially identical. Thus, it is ultimately possible to measure measured concentrations in the narrower sense as well as metabolite concentrations in the following sense

equally serve as input to the MLP.

A comparison of the two curves of each of the plots of Figure 18 thus shows that the values predicted by MLP + MFA for a future time point agree well with values obtained with a static MFA model using measured now-time concentrations of

receives extracellular metabolites. Deviations in detail, however, are possible.

The plots in Figure 18 compare a) the progress of the measured

(extracellular) fluxes with their MLP predicted counterparts. All flows are expressed as normalized values normalized to the measured glucose flow. Therefore, the glucose flow is always at 1 (not shown as a plot); and b) the progress of the MFA-calculated internal flows from measured external flows with the MFA-calculated internal flows from the NN-calculated external flows. Again, all flux against the

Glucose flow normalized.

For individual metabolites, the measured values differed a little from the predicted fluxes. On the one hand, it should be noted that the scaling is due to the glucose normalization corresponds to a very high "dissolution" and the deviations were rather small considering the total amount of the metabolite. In addition, a certain tendency of the overfitting data was observed, which is usually remediable by increasing the data set.

The plots of FIG. 18 generated for 11 different extracellular fluxes represent a "metabolic fingerprint" of the cell clone ZK1. It will be shown below that this fingerprint is distinctly different from that of other cell clones producing the same product (bispecific antibodies, but containing the same) May differ in protein coding sequence in aberrant copy number or elsewhere in the genome.

FIG. 19 shows several plots with two curves each, all for a different fed batch bioreactor BR2 with a second cell clone ZK2 for locating the

bispecific antibody were obtained. The curves were determined as already described for Figure 18, but based on data from the second cell clone ZK2. The plots shown in FIGS. 18 and 19 relate to the data of two bioreactors BR1, BR2 which were operated with the same type of medium and with the same nutrient solution and are essentially only

differ in that in the two bioreactors different cell clones ZK1 and ZK2 are kept, although both synthesize the same product (bispecific antibody), but in genetic terms

Have differences that may also affect the metabolism of the clones: The clones may have been generated, for example, by a method that does not fully control the position of integration of a new DNA sequence segment into the genome of the cells and / or the number the integrated sequence sections offers. Thus, although the starting lines used are genetically identical, in the course of integration of new genes or DNA sequences (e.g., in the case of random integration in the course of transfection), the

Gene sequences for the light chain (LC) and heavy chain (FIC) of the

Bispecific antibody in different copy numbers and different genome loci are inserted. This, in turn, can cause metabolic differences as well as differences in productivity and vitality of a particular cell clone, which, in accordance with embodiments of the invention, can be readily recognized by metabolic prediction using a combination of MLP and MFA differentiable. The second bioreactor BR2 was identical to the bioreactor BR1, which was used for the cell culture of the clone ZK1.

FIGS. 18 and 19 each show, for a particular cell clone ZK1, ZK2, that the combination of a neural network NN and a metabolic one

Flow analysis MFA to other and ultimately more accurate predictions comes as a purely on an MFA and the current is-time related prediction of extracellular flows. In addition, this combination has the advantage that the quality of the

Prediction by the measured concentrations of extracellular fluxes are verifiable again and again. A direct prediction of the intracellular flows would only allow a review with 13C-labeled metabolites, which is neither practical nor economical, especially for the flow rate or even for production fermentations. In addition, an amount of predicted extracellular fluxes calculated for a particular cell clone ZK1, ZK2, e.g. shown in Figure 18 for ZK1 and in Figure 19 for ZK2, a

This can be analyzed to identify individual cell clones that are unique in terms of a particular parameter, such as the "metabolic fingerprint" of a particular cell clone. Growth rate, rate of ammonia degradation, the rate of production of serine, etc. appear particularly advantageous. The metabolic "fingerprints" can also be used to make two or more different cell clones in terms of their

to select or characterize metabolic similarity.

FIG. 20 shows several plots with two curves each, all of which have been obtained for the bioreactor BR1 with the cell clone ZK1 and which each represent calculated intracellular flows. FIG. 20 thus corresponds to FIG. 18 with the difference that the calculated intracellular instead of the calculated extracellular flows are shown.

FIG. 21 shows several plots with two curves each, all of which have been obtained for the bioreactor BR2 with the cell clone ZK2 and which each represent calculated intracellular flows. FIG. 21 thus corresponds to FIG. 19 with the difference that the calculated intracellular instead of the calculated extracellular flows are shown. The totality of the plots in FIGS. 19 and 21 thus constitutes one

metabolic fingerprint, which can be used for metabolic characterization of a cell clone. By collecting data from a The multiplicity of these "fingerprints" for a multiplicity of different clones and fermentation conditions and detection of the vitality and / or productivity of the individual clones makes it possible, for example by means of correlation analyzes, to recognize advantageous metabolic fingerprints and to use these as metabolic fingerprints recognized as advantageous as a reference value to recognize beneficial clones in future cloning and to select them for a fermentation project.

List of reference numbers

Claims

A method of predicting the metabolic state of a cell culture of cells of a particular cell type comprising:

Providing (102) a metabolic model (402) of a cell of the

The metabolic model includes a variety of intracellular and extracellular metabolites, as well as a variety of intracellular (410) and extracellular (408) fluxes, wherein the metabolic model comprises stoichiometric equations that include at least a stoichiometric relationship between one of the intracellular (406) and specify one of the extracellular (404) metabolites;

at any of a plurality of times during the culture of the cell culture:

receiving (106) a plurality of measurements measured at that time, the measurements comprising concentrations of a plurality of extracellular metabolites of the metabolic model in the culture medium of the cell culture and a measured cell density of the cells in the cell culture;

o Input (108) the received measured values as

Input parameter values into a trained Machinearning program logic - MLP - (218);

o Prediction (110) of extracellular fluxes (408) of extracellular metabolites at one point in time by the MLP below

Using the received measurements, the

Time of arrival is a time subsequent to the reception of the measured values, whereby the extracellular flows

Uptake rates of extracellular metabolites into a cell and / or delivery rates of extracellular metabolites from a cell into the medium;

o Execution (112) of a metabolic flow analysis for the calculation of the intracellular flows at the time of the future under Use of the predicted extracellular fluxes of extracellular metabolites and the stoichiometric equations of the metabolic model.

2. The method of claim 1, further comprising generating the MLP by machine learning, the generation comprising:

Generating a training data set, the generation comprising:

o at any of a plurality of training times during the cultivation of at least one training cell culture of cells of the particular cell type:

^■ receiving a plurality this training time measured measurement values, wherein the measured values in the at least comprise concentrations of several extracellular metabolites of the metabolic model in the culture medium of at least a training cell culture as well as a measured cell density of the cells a training cell culture;

^■ receipt of the time of the current training time; and

^■ calculation of extracellular fluxes of extracellular metabolites as a function of the measurements received at that time and at the respective previous time, the extracellular flows being uptake rates of the extracellular metabolites into the cell and / or delivery rates of the extracellular metabolites into the medium;

Train the MLP, where the training includes:

o Enter the readings taken at each of the training times

are received as input parameter values into the MLP, and input of the extracellular flows occurring at the time following that training time for that subsequent one

Have been calculated as output parameter values associated with these input parameter values into the MLP; and

o performing a learning process by the MLP such that the MLP learns to predict the respective associated output parameter values based on the input parameter values; Storage of the trained MLP in a volatile or non-volatile storage medium.

3. The method of claim 2, wherein the training data set is generated such that at each of a plurality of training times, during the culture of multiple training cell cultures of cells of the particular cell type, the measurements and times are received and the extracellular fluxes of the extracellular metabolites are calculated be, with the

Cell cultures are cultured in bioreactors of different types, wherein the types of bioreactors include at least two different bioreactor types from the following bioreactor types: a fed-batch bioreactor, a batch bioreactor, a perfusion reactor, a chemostat and a split-batch

Bioreactor.

4. Method according to one of the preceding claims, wherein the MLP is a neural network or a system of several neural networks.

5. The method according to any one of the preceding claims, wherein it is in the

measured concentrations of the several extracellular metabolites each involves:

a volume indication of the metabolite,

is in particular a mass concentration or molar concentration, or

a value which is linear or at least approximately

correlated linearly with the volume-related content indication, in particular a corrected or normalized measured

Metabolite concentration or a measured flow of the extracellular metabolite.

A method according to any one of the preceding claims, wherein the MLP is a system of multiple sub-MLPs, each sub-MLP contained in the system being for the extracellular prediction Of a single extracellular metabolite were trained and used selectively to predict the extracellular flow of this single extracellular metabolite at the time of the future.

7. The method according to any one of the preceding claims, wherein the MLP for

Prediction of the extracellular flow of a single extracellular metabolite measured extracellular metabolite concentrations as input parameter values, wherein the multiple

extracellular metabolites that differ as input parameter values for at least two of the extracellular metabolites whose extracellular flux is to be determined.

8. The method of claim 7, further comprising:

Measurement of the concentrations of all input candidate metabolites over several time points, with the input candidate metabolites all

comprises extracellular metabolites that are metrologically available in a reference bioreactor with a cell culture of the particular type or that include all extracellular metabolites of the metabolic model;

for each of the extracellular metabolites whose extracellular flux is to be predicted, carrying out a selection procedure for the identification of the several extracellular metabolites identified as

Input parameter values are to be used to predict the extracellular flow of this single metabolite, the

The selection method with respect to this single metabolite comprises: a) defining a first amount of extracellular metabolites, the first set comprising all input candidate metabolites; b) calculating a first relevance score of each of the extracellular metabolites in the first set as a function of the measured concentrations of that metabolite, the first relevance score indicating the predictive power of the concentration of the particular extracellular metabolite with respect to the extracellular flow of that single extracellular metabolite; c) converting only the extracellular metabolite having the highest first relevance score from the first amount to a still empty second amount of extracellular metabolites removing that metabolite from the first amount; d) Calculation of a further relevance score of each of the extracellular metabolites in the first set as a function of the measured

Concentrations of this metabolite and the measured

Concentrations of all extracellular metabolites contained in the second set, where the further relevance score is the predictive power of the concentration of the respective input candidate metabolite in the

Indicating the extracellular flux of this single extracellular metabolite taking into account the metabolites already present in the second amount; e) transferring only that extracellular metabolite of the first quantity having the highest further relevance score into the second quantity by removing that metabolite from the first quantity, the transfer taking place only if the second quantity is absorbed by the uptake of this metabolite does not exceed a maximum limit for informative redundancy of the metabolites it contains with respect to the prediction of extracellular flow of that single extracellular metabolite; f) repeating steps d) and e) until no more metabolites can be transferred from the first to the second amount without the second amount exceeding the maximum limit for informative redundancy; and g) selectively using only the metabolites that have been converted to the second amount as input parameter values for predicting the extracellular flow of that single extracellular metabolite.

9. The method according to claim 8,

- where the first relevance score is calculated as the partial mutual information score (PMI score) between a metabolite of the first quantity and the single metabolite whose extracellular flux is to be predicted;

- where the second relevance score is calculated as the PMI score - between a metabolite of the first quantity and the single metabolite whose extracellular flux is to be predicted taking into account all the metabolites already present in the second quantity.

10. The method according to any one of the preceding claims, wherein the MLP for

Used to predict the extracellular flow of each extracellular metabolite, as an input parameter value of several extracellular metabolites, wherein the plurality of extracellular metabolites comprise at least one, preferably at least two, amino acids.

11. The method according to any one of the preceding claims, wherein the plurality of

Time intervals by time intervals of 10 minutes to 48 hours, preferably from 1 to 24 hours, are separated.

12. A method according to any one of the preceding claims, wherein the cell type is a genetically modified cell type which is maintained and / or propagated in a bioreactor for the purpose of obtaining a biomolecule.

13. The method according to any one of the preceding claims, wherein at each of the future times, a calculation of several or all intracellular flows of the metabolic model takes place.

14. The method according to any one of the preceding claims, further comprising:

Identification of one or more intracellular metabolites of the

a metabolic model whose calculated intracellular fluxes (410) deviate from a respective reference value or reference value range by more than the limit value;

- automatic identification of the intracellular flow, as

limiting factor for cell growth or production of a desired biomolecule acts.

A method of monitoring and / or controlling a bioreactor that includes the cell culture of cells of a particular cell type, comprising:

- Calculating the intracellular fluxes of the cells at a future time according to the method of any one of claims 1-14;

Comparing the predicted intracellular fluxes with reference or reference ranges for acceptable intracellular fluxes of the respective one or more intracellular metabolites;

- Output of a warning if a deviation of the calculated

intracellular flow of at least one of the intracellular metabolites exceeds its limit from its respective reference or reference value range; and or

- Sending a control command to the bioreactor to automatically initiate steps that change the state of the bioreactor or the medium contained therein so that the deviation is reduced, the automatic steps in particular a change in the amount and / or a change in the composition of a nutrient solution include.

16. The method of claim 15, further comprising:

Identifying the reaction within the metabolic model of the cells which acts as a limiting factor for cell growth or production of a desired biomolecule according to claim 14; and

automatically adding selectively those substances which alter the limiting factor intracellular flow to promote cell growth or production of the biomolecule or the quality of the biomolecule, or issue a request for such addition via a user interface.

17. A method of identifying a metabolically beneficial clone of cells of a particular cell type, comprising:

Culturing different cell cultures in multiple bioreactors, the different cell cultures containing different clones of cells of the particular cell type; - Calculation of the intracellular flow of one or more

intracellular metabolites separated at multiple time points for each of the cell clones according to any one of claims 1-16;

18. The method according to any one of the preceding claims, further comprising:

- calculating the actual extracellular flow of one or more of the extracellular metabolites from the extracellular metabolite concentrations measured at the current time and date;

Performing another metabolic flow analysis to calculate the current intracellular fluxes (410) at the current time using the calculated extracellular extracellular metabolite fluxes and stoichiometric equations of the metabolic model; and

Use of the calculated current intracellular fluxes as

Description of a current metabolic state of the cells of the cell culture.

19. The method according to any one of the preceding claims,

at each of the time points during culture of the cell culture, further receiving a lactate dehydrogenase LDH concentration measured in the medium of the cell culture;

- wherein the prediction of extracellular flows of extracellular

Metabolites at each of the future times by the MLP below

Using a corrected instead of the measured cell density, wherein the calculation of the corrected cell density for each of the times comprises:

o Calculation of the density of lysed cells in the medium of the cell culture as a function of the measured LDH concentration, the function being an empirically determined, heuristic and linear function representing the dependence of the LDH concentration in the medium on the number of lysed cells of this particular cell type; o Calculation of the corrected cell density as the sum of the measured cell density in the medium and the calculated density of the lysed cells.

20. System (200) for predicting the metabolic state of a cell culture of cells of a particular cell type comprising:

one or more processors (202);

a first interface (210) for receiving measured values from one

Bioreactor (204, 206, 208) containing the cell culture;

a volatile or non-volatile storage medium (212) with:

a metabolic model (402) of a cell of the particular cell type, the metabolic model including a plurality of intracellular metabolites and extracellular metabolites, and a plurality of intracellular (410) and extracellular (408) fluxes, wherein the metabolic model comprises stoichiometric equations comprising at least one specify stoichiometric relationship between one of the intracellular (406) and one of the extracellular (404) metabolites;

o a trained machinelearning program logic - MLP - (218);

o a program logic (220) adapted to perform a method at each of a plurality of times during the culture of the cell culture, comprising:

^■ Receiving (106) a plurality of measured values measured at this time via the first interface, wherein the measured values of several concentrations of extracellular metabolites of the metabolic model in the culture medium of the cell culture, as well as a measured cell density of the cells include the cell culture;

^■ Input (108) of the received measured values as

Input parameter values in the MLP;

^■ Prediction (110) of extracellular fluxes (408) of

extracellular metabolites at a future time point by the MLP using the received measurement values, the future time being a time subsequent to the time of receipt of the measurement values, wherein the extracellular fluxes record the rates of extracellular metabolites in a cell and / or delivery rates

extracellular metabolites from a cell into the medium; Performing (112) a metabolic flow analysis to calculate the intracellular fluxes on the

Future time using predicted extracellular flows and stoichiometric equations of the metabolic model.