KR20210026543A - A system of predicting biological activity for compound with target protein using geometry images and artificial neural network - Google Patents

Abstract

The present invention relates to a system for predicting the activity of a protein-binding component based on artificial neural network models. According to the present invention, the artificial neural network models predicting whether or not the compound is active with respect to a target protein are combined and then learning and prediction are performed. The system includes: a data management module storing protein-related compound activity data and three-dimensional shape or structure data on the protein and the compound; an individual neural network module provided with its own individual neural network module, extracting a descriptor from the three-dimensional shape or structure data on the protein or the compound, causing the individual neural network model to learn the extracted descriptor and the activity data, and storing result data for secondary learning output by application to the learned individual neural network model; an integrated neural network module provided with its own integrated neural network module, generating secondary learning data by integrating the result data for secondary learning generated at each individual neural network module, and causing the integrated neural network model to learn the generated secondary learning data; and an activity prediction module performing activity prediction by using the individual neural network model and the integrated neural network model with respect to a query protein and a query compound. In the system described above, protein-compound binding information learning and prediction are performed using artificial neural network models, and thus data bias can be reduced and accuracy can be enhanced even with respect to a difficult prediction question.

Description

{A system of predicting biological activity for compound with target protein using geometry images and artificial neural network}

The present invention learns the binding information between a protein and a compound using artificial intelligence in order to efficiently process the active substance discovery step, which is an early stage in the development of a new drug, to see if a given protein and compound can bind to show activity. It relates to a system for predicting the activity of a protein-binding compound based on a plurality of artificial neural network models to determine and predict.

In addition, in order to combine a plurality of artificial neural network models that predict whether a compound is active against a target protein, the present invention outputs probability distribution data on the degree of reaction between a protein and a compound through an individual artificial neural network model. And, it relates to a system for predicting the activity of a protein-binding compound based on a plurality of artificial neural network models, which constitutes a separate integrated artificial neural network using this as learning data.

The development of new drugs using computers is widely used in the drug discovery stage, and two methods are largely used to discover the hit substance of a compound. The first method is compound-based drug design, which involves finding or making compounds with similar chemical properties based on the fact that a compound binds to a specific structural protein. The second method, protein structure-based drug design, uses prior knowledge of the three-dimensional structure information of the protein to find compounds showing activity among compounds with similar shapes and sizes.

The protein structure-based drug design method can also be divided into three broad categories. The first is a method of quickly finding a compound having the most similar form to information in the form of a binding site or a binding pocket using a docking program [Patent Document 1]. The second is a method of combining elements or molecular fragments of a compound according to the size and size of a binding pocket. Third, it is a method of optimizing the known conformation among compounds showing activity in the binding cavity. Among these, a method in the form of virtual screening has been actively applied due to the development of computer technology [Patent Document 2].

From the technical aspect of virtual screening, in the method of finding compounds that exhibit activity on the target protein in the existing mechanical library of hundreds to tens of millions of compounds, the movement of methods using artificial neural networks (deep learning) is nowadays. It is actively. In 2015, Atomwise introduced AtomNet technology that predicts the binding affinity of molecules and the binding structure of a new drug target (protein) for the first time in the world.

However, when solving a high-difficulty prediction problem by constructing an artificial neural network with only a single model, the accuracy is often inferior due to the appearance of bias according to the data.To solve this problem, if various artificial neural networks are combined and predicted, the accuracy can be further improved. will be.

However, the method of combining models of various artificial neural networks is to perform learning after establishing an environment in which a plurality of individual artificial neural network models can be controlled by the artificial neural network integrated model. However, when the neural network's integrated model is actually applied, the following problems arise.

First, when the construction of each single artificial neural network itself is complex, each artificial neural network is constructed for each type, and in this case, development cost or maintenance cost increases rapidly in order to unified management of different artificial neural networks.

In addition, when learning by integrating each individual artificial neural network at the same time, learning itself may be impossible due to a load on the artificial intelligence server or lack of main memory space.

In addition, when training an integrated artificial neural network after each individual artificial neural network is completed, all training datasets that are introduced into each single artificial neural network must be randomly shuffled in the same order. This is because if the data is not shuffled randomly, the artificial neural network can make a biased prediction for the last learned data.

In addition, the developer implementing the integrated artificial neural network must be aware of the composition and data format of the input data (learning data set) of each individual artificial neural network, and must also acquire control over the data itself.

Therefore, there is a need for a technology that overcomes the above difficulties, reduces the enormous amount of time spent, and simplifies the entire process.

Korean Patent Application Publication No. 10-2018-0058648 (published on June 1, 2018) Korean Patent Application Publication No. 10-2019-0000167 (published on January 2, 2019)

http://dude.docking.org/ Connolly, M. L., "Analytical molecular surface calculation.", J. Appl. Cryst. 1983, 16, 548-558

An object of the present invention is to solve the above-described problems, and in order to combine a plurality of artificial neural network models that predict whether a compound is active against a target protein, proteins and compounds through individual artificial neural network models It is to provide a system for predicting the activity of protein-binding compounds based on a plurality of artificial neural network models, which constitutes a separate integrated artificial neural network that outputs probability distribution data for the response degree of and uses it as learning data.

In order to achieve the above object, the present invention relates to a system for predicting the activity of a protein-binding compound based on a plurality of artificial neural network models, and storing the activity data of the compound for the protein and the three-dimensional shape or three-dimensional structure data of the protein and the compound. A data management module; Equipped with its own individual neural network module, extract a descriptor from 3D shape data or 3D structure data of a protein or compound, train the individual neural network model with the extracted descriptor and active data, and apply it to the trained individual neural network model An individual neural network module that stores the output data for secondary learning; An integrated neural network that has its own integrated neural network module, generates secondary training data by integrating the secondary training result data generated from each individual neural network module, and trains the integrated neural network model with the generated secondary training data. module; And an activity prediction module that predicts activity for a query protein and a query compound using the individual neural network model and the integrated neural network model, wherein the individual neural network model is composed of at least two and one individual neural network module group It is characterized in that to form.

In addition, the present invention is a system for predicting the activity of a protein-binding compound based on a plurality of artificial neural network models, wherein the individual neural network module generates primary training data from the extracted descriptor and activity data, and the individual neural network model is a part of the primary training data. And generating result data for secondary training by applying another part of the primary training data to the trained individual neural network model.

In addition, in the present invention, in a system for predicting the activity of protein-binding compounds based on a plurality of artificial neural network models, the primary learning data includes identification information of proteins and compounds, descriptors for proteins and compounds, and activity of the corresponding compounds for the proteins. It includes a label value determined from the data, and the individual neural network module obtains the prediction result data by applying the descriptor of the corresponding data to the individual neural network model for each data of the primary training data, and obtains the prediction result data. A label of corresponding data of the learning data is assigned to generate result data for secondary learning, and identification information for proteins and compounds is included in the secondary learning result data.

In addition, the present invention is a system for predicting the activity of a protein-binding compound based on a plurality of artificial neural network models, wherein the output data of the individual neural network models is output as a probability value of a label value, and the integrated neural network module provides result data for secondary learning, It is characterized by combining and integrating based on a pair of proteins and compounds, but if the pairs of proteins and compounds match, the secondary learning result data of all individual neural network modules are combined into one data to generate secondary learning data.

In addition, the present invention is a system for predicting the activity of a protein-binding compound based on a plurality of artificial neural network models, wherein the individual neural network module group comprises at least one individual neural network module that generates a descriptor from 3D shape data, and a descriptor from 3D structure data. It characterized in that it comprises at least one individual neural network module for generating.

In addition, the present invention is a system for predicting the activity of a protein-binding compound based on a plurality of artificial neural network models, wherein the individual neural network module extracts one-dimensional data from three-dimensional shape data using Fourier transform and generates a descriptor or a homology. It is characterized by generating a descriptor by extracting a two-dimensional image from the three-dimensional structure data and generating a descriptor, or by projecting the three-dimensional shape data as a two-dimensional image.

In addition, the present invention is a system for predicting the activity of a protein-binding compound based on a plurality of artificial neural network models, wherein the 3D shape data is 3D data for a closed mesh surface as closed surface data, and the 3D structure The data are three-dimensional data on an atomic position in a chemical structure of a protein or a compound, and are characterized in that the data is composed of three-dimensional points.

As described above, according to the system for predicting the activity of protein-binding compounds based on a plurality of artificial neural network models according to the present invention, by learning and predicting binding information between proteins and compounds using a plurality of artificial neural network models, a prediction problem with high difficulty Also, the effect of reducing the bias according to the data and increasing the accuracy according to the data is obtained. In other words, regions that cannot be predicted by each individual artificial neural network can be classified with improved discrimination power.

In addition, according to the present invention, by independently learning each artificial neural network model, available resources of the entire system can be appropriately distributed, and development and maintenance costs can be significantly reduced.

According to the system according to the present invention, by learning the binding information of a protein and a compound using artificial intelligence and predicting its activity, it is possible to extract a large amount of active compound candidates from the structural information of the compound bound to the protein in a short time. Is obtained. Through this, it is possible to quickly select a compound to be tested experimentally, and significantly reduce the time and cost required to discover an effective substance.

1 is a block diagram of an entire system for implementing the present invention.
Figure 2 is a block diagram of the configuration of a system for predicting the activity of a protein-binding compound based on a plurality of artificial neural network models according to the present invention.
3 is a block diagram of a detailed configuration of a system for predicting activity of a protein-binding compound based on a plurality of artificial neural network models according to the present invention.
4 is a diagram illustrating the extraction of a substrate-binding site of a protein and a three-dimensional shape of the surface of a compound according to an embodiment of the present invention, (a) a compound, and (b) a view of the substrate-binding site of a protein.
5 is an exemplary diagram for calculating a Fourier coefficient vector by Fourier transform of a protein and a compound according to an embodiment of the present invention.
6 is an exemplary graph for a persistence diagram according to an embodiment of the present invention.
7 is an exemplary diagram for a homology image according to an embodiment of the present invention.
8 is a flowchart illustrating a process of generating a descriptor according to an embodiment of the present invention.
9 is a diagram illustrating a process of generating geometric images by applying spherical parametrization to a three-dimensional shape according to an embodiment of the present invention.
10 is a table illustrating integrated data for secondary learning data according to an embodiment of the present invention.
11 is a table showing a performance evaluation (based on a softmax value) for an individual neural network model of a homology method according to an experiment of the present invention.
12 is a table showing the performance evaluation (based on a softmax value) for an individual neural network model of the Fourier method according to the experiment of the present invention.
13 is a table showing the performance evaluation of the integrated neural network model according to the experiment of the present invention.

Hereinafter, specific details for the implementation of the present invention will be described with reference to the drawings.

In addition, in describing the present invention, the same portions are denoted by the same reference numerals, and repeated explanations thereof are omitted.

First, examples of the configuration of an entire system for implementing the present invention will be described with reference to FIG. 1.

1(a) and 1(b), the system for predicting the activity of protein-binding compounds based on a plurality of artificial neural network models according to the present invention may be implemented as a server system on a network or a program system on a computer terminal.

As shown in FIG. 1 (a), an example of the overall system for the implementation of the present invention is composed of an analysis terminal 10 and an active prediction system 30, and are connected to each other by a network 20. In addition, a database 40 for storing necessary data may be further provided.

The analysis terminal 10 is a general computing terminal such as PC, notebook, netbook, PDA, and mobile used by users such as new drug development researchers. The user transmits data, such as the 3D structure of the target protein and compound, to the activity prediction system 30 through the user terminal 10 or receives the activity prediction result value from the activity prediction system 30.

The activity prediction system 30 is connected to the network 20 as a conventional server to provide a service that supports the prediction of compound activity for a target protein using an artificial neural network. Meanwhile, the activity prediction system 30 may be implemented as a web server or a web application server that provides each service as a web page on the Internet. In addition, the activity prediction system 30 may be implemented as a cloud system, and may perform a learning or analysis function based on a cloud and provide an active prediction service.

The database 40 is a general storage medium for storing data required by the activity prediction system 30, and stores activity data such as data on a three-dimensional structure of a target protein or compound, and binding between a target protein and a compound.

On the other hand, the database 40 may be constructed by importing data from a library of natural products or chemical compounds that have already been built.

Specifically, the database 40 may include an individual neural network model 41, an integrated neural network model 42, and a secondary learning data storage 43 that stores result data for secondary learning. However, the configuration of the database 40 is only a preferred embodiment, and may be configured in a different structure according to the database construction theory in consideration of the ease and efficiency of access and search in developing a specific system.

Meanwhile, the activity prediction system 30 may be composed of a server-client system composed of a server and a client. That is, the descriptor generation, major learning or analysis functions of the active prediction system 30 are built in the server, and the user interface or simple preprocessing tasks for analysis can be built in the analysis terminal 10 as a client module. The division of work between the server and the client can be implemented in various forms according to the general server-client construction theory.

In addition, in the active prediction system 30, a learning function or a prediction function is built as an engine module, and a client service module installed in the analysis terminal 10 uses the engine module to train an artificial intelligence model with previously collected data. In addition, it is possible to provide a service for predicting the activity of a compound against a target protein through the learned model. In this case, the analysis terminal 10 may serve as another server.

In addition, as shown in FIG. 1 (b), another example of the overall system for the implementation of the present invention is configured with an activity prediction system 30 in the form of a program installed in the computer terminal 13. That is, each function of the activity prediction system 30 may be implemented as a computer program and installed in the computer terminal 10, and implemented as a program system on the computer terminal 10. A program installed in the computer terminal 10 may operate like a single program system 30. Meanwhile, data required by the activity prediction system 30 are stored and used in a storage space such as a hard disk of the computer terminal 10.

On the other hand, as another embodiment, the system for predicting the activity of a protein-binding compound based on a plurality of artificial neural network models may be implemented as a single electronic circuit such as an ASIC (on-demand semiconductor) in addition to being configured as a program and operating on a general-purpose computer. Alternatively, it may be developed as a dedicated computer terminal 10 that exclusively processes only predicting compound activity. This will be referred to as the active prediction system 30. Other possible forms may also be implemented.

Next, a configuration of a system for predicting activity of a protein-binding compound based on a plurality of artificial neural network models according to an embodiment of the present invention will be described with reference to FIG. 2.

As shown in FIG. 2, the system 30 for predicting activity of a protein-binding compound based on a plurality of artificial neural network models according to the present invention includes a data management module ( 31), an individual neural network module (33) that trains an individual neural network model (41) by generating a descriptor from 3D data and generates primary training data from the descriptor and active data and generates secondary training data (33), a number of individual neural networks An individual neural network model group 32 composed of a model 33, an integrated neural network module 34 that trains a secondary neural network model 42 using secondary training data, and a test target protein (or query protein) and a test target compound It is composed of an activity prediction module 35 that predicts activity for (or query compound) by using an individual neural network model 41 and an integrated neural network model 42.

First, the data management module 31 collects activity data of a compound for a protein, or forms and manages the three-dimensional shape or three-dimensional structure data of the protein and the compound.

The activity data is data indicating whether a specific compound is an active compound or an inactive compound for a specific target protein. In other words, it is already known whether the binding activity of a protein and a compound is active.

The three-dimensional shape data or three-dimensional structure data for a protein or compound is three-dimensional data representing the three-dimensional shape or structure of a protein or compound. In this case, the protein or compound is a target protein or compound belonging to the previously collected activity data.

Next, the individual neural network module 33 has its own neural network model (or individual neural network module) 41, extracts a descriptor from 3D shape data or 3D structure data of a protein or compound, and extracts the extracted descriptor and The result for secondary training by generating primary training data as active data, training an individual neural network model 41 as part of the primary training data, and applying another part of the primary training data to the trained individual neural network model 41 Generate data.

In addition, the individual neural network module 33 is composed of a plurality of units and is composed of one group, that is, an individual neural network module group 32. In particular, a plurality of individual neural network modules 33 may be provided according to a method of extracting a descriptor. That is, the method of extracting the descriptor is unique for each individual neural network module 33.

Meanwhile, preferably, the individual neural network module group 32 includes at least one individual neural network module 33 for generating a descriptor from 3D shape data, and at least one individual neural network module for generating a descriptor from 3D structure data ( 33). That is, by reflecting both the three-dimensional shape and the three-dimensional structure of proteins and compounds, it is possible to more accurately classify the activity prediction.

Preferably, the descriptor consists of one-dimensional data or two-dimensional image data. That is, the individual neural network module 33 extracts 1D data or 2D image data from 3D shape data or 3D structure data and uses it as a descriptor.

In addition, the individual neural network module 33 generates primary learning data from descriptors of proteins and compounds, and activity data of compounds for proteins. That is, primary learning data is generated by labeling a descriptor for a pair of a protein and a compound and activity data of a corresponding compound for the protein in the descriptor. That is, the activity data is used as a label.

The primary training data is divided into two groups, one group is used to train its individual neural network model 41, and the other group is used to generate result data for secondary training.

That is, the individual neural network module 33 trains its own individual neural network model 41 with some (or first group) of the primary training data. In this case, the first group may be divided into training data and verification data (or test data) therein.

In addition, the individual neural network module 33 applies the second group to the trained individual neural network model 41 to generate result data for secondary learning. That is, for each data of the first training data of the second group, the descriptor of the data is applied to the individual neural network model 41 to obtain the prediction result data, and the label of the data (primary training data) in the prediction result data It creates result data for secondary learning by giving (label of the descriptor).

In addition, information (or identification information) on proteins and compounds of the data (primary learning data) is included in the result data for secondary learning.

That is, the result data for secondary learning is composed of information on proteins and compounds (identification information such as ID), prediction result data, and labels.

Next, the individual neural network model 41 is a neural network model provided in the individual neural network module 33, and a DNN (Deep Neural Network) or a deep learning model, a cyclic neural network (RNN), a convolutional neural network (CNN), etc. are applied. I can. When the descriptor is a two-dimensional image, preferably, a convolutional neural network (CNN) is used.

The input data of the individual neural network model 41 is a descriptor of a protein and a descriptor of a compound, and the output data is a probability value of a label value (category value). That is, the output data is output as a probability value of each label value. The number of predicted result values of each individual neural network model 41, that is, the number of probabilities for an output label, is the same as the number of category values (Label). Preferably, the label value (or category value) of the output data is active and inactive.

The individual neural network model 41 is divided into a regression type that predicts a numerical value of data and a classification type that predicts a specific category of data. Preferably, the output of the individual neural network model 41 is configured in the form of a classification.

In particular, the individual neural network model 41 calculates and outputs a softmax value capable of expressing predictability for each data (or label value indicating classification). The softmax function to obtain the softmax value is as follows.

[Equation 1]

Here, x _i is the final output value, and k+1 is the number of categories. e is an Euler constant or natural constant.

That is, the result of the softmax function is a one-dimensional vector of the same size as the number of categories (or label values) to be predicted, and is a kind of probability distribution that becomes 1.0 when all of the values are added.

Next, the integrated neural network module 34 has its own neural network model (or the integrated neural network module) 42, and the secondary learning by integrating the result data for secondary learning generated by each individual neural network module 33 Data is generated, and the integrated neural network model 42 is trained with the generated secondary training data.

That is, the integrated neural network module 34 combines and integrates the secondary learning result data generated by each individual neural network module 33 based on a pair of proteins and compounds. That is, if the pair of protein and compound is the same, the result data for secondary learning of different individual neural network modules are combined into one data and integrated. In particular, for one pair of protein and compound, secondary training data is generated by combining the secondary training result data of all individual neural network modules.

As described above, the result data of the individual neural network model 41 is output as a result value of the softmax function. Accordingly, the integrated neural network module 34 collects the result values of the softmax function output from each individual neural network model 41, and combines the softmax result values to form secondary learning data.

In addition, the secondary learning data is comprised of the identification information of each protein and compound. Therefore, the secondary learning data is composed of identification information of each protein and compound, prediction result values of all individual neural network models 41 for the protein and compound, and activity data (or label value) of the protein and compound.

In addition, the integrated neural network module 34 trains the integrated neural network model 42 using secondary learning data generated by integration. In this case, the secondary learning data may be divided into learning data and verification data (or test data) therein.

Next, the integrated neural network model 42 is a neural network model provided in the integrated neural network module 34, to which a deep neural network (DNN) or a deep learning model, a recurrent neural network (RNN), a convolutional neural network (CNN), etc. are applied. I can. One-dimensional data is used as the descriptor or input data.

Preferably, the integrated neural network model 42 is configured using a Fully-connected Neural Network (FNN). The number of elements constituting each data input to the integrated neural network model 42 (the number of features) can be obtained as follows.

Number of input data features = (number of individual neural networks) × (number of labels)

Next, the activity prediction module 35 predicts the activity of the test target protein (or query protein) and the test target compound (or query compound) using the individual neural network model 41 and the integrated neural network model 42.

That is, the activity prediction module 35 requests 3D data on the query protein and the query compound to the data management module 31, and transmits or transmits it to the individual neural network module 33.

In addition, the activity prediction module 35 requests the individual neural network model 32 to primarily predict the activity result by applying the 3D data of the query protein and the query compound to the learned individual neural network model 41. At this time, the individual neural network model 32 generates a descriptor from 3D data, and applies the generated descriptor to the individual neural network model 41 to calculate a first-order prediction result.

At this time, the activity prediction module 35 requests prediction of the query protein and the query compound to all individual neural network modules 33, and transmits or transmits all prediction results to the integrated neural network module 34.

In addition, the activity prediction module 35 applies the integrated neural network model 42 to the first prediction result of the query protein and the query compound, and requests the second prediction. At this time, the integrated neural network module 34, when acquiring all the first prediction results of all individual neural network models 41, integrates a plurality of first prediction results, and applies the integrated data to the integrated neural network model 42, The final quadratic prediction is output.

Meanwhile, the individual neural network modules 33 and the integrated neural network module 34 may be constructed as systems independent of each other. In this case, preferably, data may be exchanged through the database 40 such as the data storage 43 for secondary learning.

Next, a configuration of the data management module 31 according to an embodiment of the present invention will be described in more detail with reference to FIG. 3.

As shown in FIG. 3, the data management module 31 includes an active data collection unit 311 for collecting active data, and a 3D data forming unit 312 for generating 3D formation data or 3D structure data. do.

First, the activity data collection unit 311 collects activity data of each compound for each target protein.

The activity data is data indicating whether a specific compound C _j is an active compound or an inactive compound with respect to a specific target protein (or target protein) P _i. That is, the activity data consists of {(P _i , C _j , A _ij ) }. At this time, A _ij has an activation or deactivation value (binary value).

Meanwhile, the activity data may be expressed as an activity value indicating the degree of activity of the compound with respect to the target protein. In this case, it may be classified as active or inactive based on a predetermined reference value (or threshold).

Activation (active) means that a specific compound C _{j is bound to the target protein P i} , and inactive (inactive) means that it is not.

Preferably, the activity data collection unit 311 may collect activity data from a data set of previously established activity data.

As an example, a dataset provided by DUD-E (A Database of Useful Decoys: Enhanced) [Non-Patent Document 1] is used. The DUD-E dataset contains a total of 22,146 active compounds (average 217 active compounds per target protein) for a total of 102 target proteins, and 5-60 for each active instead of the inactive compound. It provides a decoy compound made in dogs. Of these, 4 target proteins were removed for reasons such as incompatibility, and data on the remaining 98 proteins were used in the experiment of the present invention. The purpose of the DUD-E dataset is as a benchmark dataset.

The decoy compound is a compound having a structure that is likely to be an inactive compound in theory, and reflects the fact that it is difficult to collect data on an inactive compound in reality. In other words, it is compound data constructed by standard data designers for the purpose of distinguishing against active compounds. The decoy compound is practically used for an inert compound.

Next, the 3D data forming unit 312 forms 3D shape data or 3D structure data for a protein or compound. In this case, the protein or compound is a target protein or compound belonging to the previously collected activity data.

First, a method of forming 3D shape data will be described. The three-dimensional shape data is closed surface data, preferably, three-dimensional data for a closed mesh surface.

Preferably, the 3D data forming unit 32 forms a three-dimensional shape of a target protein or compound using the Connolly surface [Non-Patent Document 2]. Specifically, a Connolly surface for a protein or compound is obtained, and three-dimensional shape data is generated from the Connolly surface. That is, as shown in FIG. 4, the three-dimensional shape of each protein is extracted from the three-dimensional data of the protein and the compound through the Connolly surface generation method. Preferably, three-dimensional position information of the surface is extracted through the Connolly surface generation method, and the position information is represented by three-dimensional coordinates of vertices forming a triangle of a mesh structure. That is, the set of vertices by the mesh structure represents a three-dimensional shape.

Connolly's surface refers to a surface that represents the extent to which a solvent can be accessed based on the van der Waals radius of each atom constituting the molecule. In other words, the Connolly surface represents the shape of the space occupied by proteins or compounds.

Preferably, in the case of a protein, the 3D data forming unit 32 extracts only the surface of the substrate binding site of the protein, and extracts a three-dimensional shape of the extracted surface. That is, since the entire shape of the protein is not required, but only the shape of the substrate-binding site complementary to the compound is required, only the surface of the substrate-binding site of the protein may be separately extracted.

That is, in a state in which a protein and a compound (ligand) are bound, information on a binding-site or binding site of a protein that comes within a certain range (a predetermined distance) is extracted based on the surface element coordinates of the compound. In addition, 3D shape data is generated while maintaining the closed curved surface at the extracted binding site.

Preferably, the 3D data forming unit 32 converts the extracted surface into a triangular mesh, integrates and transforms information on the surface and vertex to generate a three-dimensional shape.

Next, a method of forming 3D structure data will be described. The three-dimensional structure data is three-dimensional data on the position of atoms in the chemical structure of a protein or compound, and is data composed of three-dimensional points or points.

Preferably, the 3D data forming unit 312 forms 3D structure data of a chemical bond structure from a mole file of a protein or compound. Each point in the three-dimensional structure data represents an atomic position in three dimensions. Also, preferably, the atomic position in the state in which the protein and the compound are bound is extracted as 3D point data.

In this case, the mole file uses data provided (created) in advance, or data generated through a virtual combination. That is, when a compound or protein has a crystal structure through an actual experiment, location information of each element of the compound or protein is provided in the form of a mole file. In this case, use the provided mole file.

In addition, a chemical bond simulation tool (eg, an autodocking program) is used to virtually bind a compound and a protein, and form 3D point data from the combined virtual 3D chemical structure. Particularly, after performing a virtual combination, optimal positioning (optimal combination state) through a scoring function is obtained, and the location information at this time is stored as a Mol file.

In particular, the 3D data forming unit 312 extracts 3D structure data (or 3D point data) of the protein and the compound, respectively, from the 3D bonded structure of the protein and the compound. Preferably, separate three-dimensional structure data is extracted for each atom of the protein or compound. In addition, more preferably, 3D structure data may be extracted for only the main atom, for each atom, and 3D point data for the main atom and all other atoms.

As an example, proteins are composed of four atoms: C (carbon), N (nitrogen), O (oxygen), and S (sulfur). Therefore, 4 3D structural data are extracted for each atom of the protein. The compound may contain major atoms of C (carbon), N (nitrogen), and O (oxygen), as well as various other atoms. Therefore, the compound extracts three-dimensional point data for each atom for the major atoms of C (carbon), N (nitrogen), and O (oxygen), and extracts the three-dimensional structure data of the main atom and all other atoms. . Therefore, in total, eight three-dimensional structure data are extracted.

In addition, preferably, when extracting the 3D structure data of the protein, the 3D data forming unit 312 extracts the 3D structure data by extracting only the atomic positions near the site of the protein binding to the compound. That is, 3D structure data is formed by extracting only the positions of atoms located within a predetermined distance from the bonding site with the compound.

Next, a detailed configuration of the individual neural network module 33 according to an embodiment of the present invention will be described in more detail with reference to FIG. 3.

As shown in FIG. 3, the individual neural network module 33 includes a descriptor extracting unit 331 for extracting a descriptor from 3D data, an individual model learning unit 332 for learning the individual neural network model 41, and a secondary It consists of a result data generation unit 333 that generates result data for learning.

First, the descriptor extraction unit 331 generates 1D data or 2D image data from 3D data of a protein or compound. Each individual neural network module 33 has a different method of calculating the descriptor.

Specifically, it is classified into three methods, such as a method using Fourier transform, a method using homology, and a method using geometric image transformation. The Fourier transform method extracts a descriptor of one-dimensional data from three-dimensional shape data. In addition, the homology method generates a descriptor of a two-dimensional image from three-dimensional structure data. In addition, the geometric image conversion method extracts a descriptor of a 2D image from 3D shape data.

First, the Fourier transform method will be described.

Next, the Fourier transform (FT) is a method of expressing a function or signal as a sum of frequency components constituting the function. The transformed function becomes a complex function of frequency, and its absolute value represents the amount of frequency components constituting the original function.

First, the descriptor extraction unit 331 converts each coordinate constituting the 3D shape data extracted for a protein or compound into a spherical coordinate system through spherical harmonics.

That is, the coordinates (x,y,z) of the 3D shape data are converted to the coordinates (θ,φ,r) of the spherical coordinate system.

Next, the descriptor extracting unit 331 may express the basis function and its coefficients as a sum of the basis function and its coefficients as shown in the following equation, if Fourier transform is used for the coordinate data (θ, φ, r) of the spherical coordinate system.

[Equation 2]

Here, Y _l,k denotes the basis function, and c _l,k denotes the Fourier coefficient for the basis function Y _l,k. l represents the degree and k represents the order. L represents the size of the degree. The larger L, the smaller the error can be approximated.

That is, for all spherical coordinates (θ,φ,r), it is expressed as r = f(θ,φ). In this case, when Fourier transform is applied, r = f(θ,φ) is the basis function Y _l,k ( It can be expressed as a sum of weights (weights by Fourier coefficients) of θ,φ).

Preferably, as shown in FIG. 5, the Fourier coefficients c _l,k obtained through the Fourier transform are one-dimensional vectors (c _0,0, c _1,-1, c _l,0, c _l,1, c _2,-2, c _2,-1, c _2,0, c _2,1, c _2,2, c _{3,-3, ...,} c _L,L ) It is created with. In other words, through the above transformation, the three-dimensional structures of proteins and compounds were transformed into one-dimensional vectors by Fourier coefficients.

More preferably, the Fourier coefficients are first ordered from a number having a small degree (l), and then a number having a small degree (k).

Next, a description will be given of a descriptor extraction method using homology.

In the homology-based method, a descriptor is calculated using homology for the three-dimensional structure data of a protein or compound.

Three-dimensional structure data is three-dimensional data on the atomic or molecular structure of a protein or compound. That is, the three-dimensional structure data is a set of finite three-dimensional point data (or point data). Each point represents the position of an atom.

That is, homology is applied to the 3D structure data, a 2-dimensional persistence diagram is extracted, and the persistence diagram is converted into a 2-dimensional image and used as a descriptor.

Specifically, first, persistence information is extracted by applying homology to 3D structure data. That is, after giving a real variable r value, this value is sequentially increased from 0 to change the connection state of the point set.

The connection state of the two points is judged as follows. Draw a three-dimensional sphere with a radius r with each point as the origin. For each r, the following is judged.

If two spheres intersect, the two points are considered to be connected. If not, it is regarded as not connected. When two points are connected, a side connecting the two points is created.

If three sides are connected to each other to form a triangle, a triangle is made with these sides.

If the four triangles are connected to each other to form a tetrahedron, a solid tetrahedron is formed using this as a face.

For each r through the above process, from the given point data, a simple complex consisting of a 0-dimensional complex (point), a 1-dimensional complex (side), a 2-dimensional complex (triangle), and a 3-dimensional complex (tetrahedron). ).

_{That is, calculate the homology H r} (M) for the complex M at a given r.

Next, create a homology group for each dimensional complex. That is, a 0-dimensional homology group, a 1-dimensional homology group, and a 2-dimensional homology group are created.

And for each dimension homology group, the number of constructors is calculated. A homology group is a group created by a finite number of constructors. The constructors or constructors of a homology group are newly created or destroyed according to the constituent form of the complex of the corresponding dimension.

The homology group is composed of complexes without boundaries by forming a cycle at that dimension. That is, in the 0-dimensional, each point forms a cycle, and the 0-dimensional homology group is composed of each point. A one-dimensional homology group consists of one-dimensional complexes (sides) forming a ring. In addition, the two-dimensional homology group is composed of two-dimensional complexes (triangles) connected in the form of a closed curved surface with an inner space. The constructor corresponds to each cycle in the dimension, and the number of constructors represents the number of cycles in the dimension.

Meanwhile, one-dimensional (lower-dimensional) composites (sides) constituting a two-dimensional (upper-dimensional) composite (triangle) form a ring. However, since a homology group is formed as a quotient group by a set of lower-dimensional complexes constituting a higher-dimensional complex, the lower-dimensional complexes are congruenced into one complex. Therefore, even if one-dimensional complexes (sides) forming a two-dimensional complex (triangle) form a ring, they are excluded from the one-dimensional homology group.

On the other hand, the creation time (r _S ) for a new constructor in a specific r is recorded, and when the constructor disappears, the expiration time (r _E ) for the constructor is recorded.

The persistence diagram is a diagram of all these records. The data required to represent the persistence diagram (persistence information) is shown as follows.

Persistence information = {(d, (r _S , r _E ))}

Where d is the dimension of the constructor. r _S and r _E are the creation and destruction times of the corresponding constructor, respectively. In other words, r is sequentially increased. When r = r _{S, the} corresponding constructor is created, and _{when r = r E, the} corresponding constructor is destroyed.

As an example, persistence information may be obtained as follows.

Persistence information = ((0,(0 ,0.8)),… ,(1,(0.4,0.9))}

That is, one persistence information (0, (0, 0.8)) indicates that the corresponding constructor is 0-dimensional, was created when r = 0, and destroyed when r = 0.8.

Next, persistence information is represented by a persistence diagram.

The x-axis and y-axis of the persistence diagram represent the creation and extinction times, respectively. And each persistence information is displayed as a dot (point) on the persistence diagram. At this time, the location of the point (point) corresponds to (x,y) (r _S ,r _E ), and the value of the point is expressed as a dimension.

Preferably, the dimension, which is the value of the point, is displayed in color. As an example, red represents a 0-dimensional constructor, green represents a 1-dimensional constructor, and blue represents a 2-dimensional constructor.

An example of a persistence diagram is shown in FIG. 6.

Next, a two-dimensional homology image is created from a persistence diagram and used as a descriptor.

Preferably, Gaussian mapping is applied to the persistence diagram, and values are differentially assigned according to the density of the points.

An example of a homology image is shown in FIG. 7.

Next, a descriptor extraction method using geometric image transformation will be described.

The geometric image conversion method generates a two-dimensional geometric image from three-dimensional shape data of a protein or compound, and extracts the two-dimensional geometric image as a descriptor.

As shown in FIG. 8, the descriptor extracting unit 331 is a spherical parameterization step (S31) of mapping 3D shape data to a spherical surface, and an Otalic parameter correcting shape data on the spherical surface through oralic mapping. Authalic Parameterization Step (S32), Octahedron Parameterization Step (S33) Mapping the Corrected Spherical Shape Data on the Surface of the Octahedron (S33), Rectangle Parameterization Mapping the Shape Data on the Octahedron to a 2D Rectangle By performing a square parameterization step (S34) and a step of generating a two-dimensional geometric descriptor (S35), a two-dimensional geometric image is calculated as a descriptor.

9 is a diagram showing a form of calculating a final 2D geometric image by using information on vertices mapped to a spherical mesh.

First, 3D shape data is mapped onto a spherical surface using Spherical Parameterization (S31).

In this case, the 3D shape data is 3D data on the shape or surface of the protein or compound formed by the 3D data forming unit 312. In particular, the 3D shape data is surface data expressed in a triangular mesh, and is vertex data of each triangle. That is, the 3D shape data is composed of 3D coordinates of the vertices.

Therefore, each vertex of the 3D shape data is mapped onto a spherical surface.

Referring to the example of FIG. 9, it is the same as (a) mapping from 3D shape data to (b) spherical data.

Next, the shape data on the spherical surface is corrected through authentic parameterization (S32). That is, the data on the spherical surface is corrected so that the area of each mesh in the 3D shape data is preserved in the area of the mesh on the corresponding spherical surface. That is, it is corrected so that the area distortion ratio is minimized.

In other words, the vertex information of the original mesh of the 3D shape data mapped on the spherical surface is corrected to a spherical mesh by minimizing the area distortion ratio to be mapped. .

Next, the corrected shape data on the spherical surface is mapped onto the surface of the octahedron (S33). That is, through Octahedron Parameterization, shape data (or mesh) on a spherical surface is converted into an octahedral domain.

Preferably, the octahedron is an octahedron.

Referring to the example of FIG. 9, it is the same as (b) mapping from spherical data to (c) octahedral data.

Next, the shape data on the octahedron is mapped to a two-dimensional square (S34).

That is, the transformed octahedron is flattened and mapped to a two-dimensional square. Preferably, the two-dimensional square is formed into an N×N square. As an example, N is set to 128.

Meanwhile, the eight triangles on each surface of the octahedron are combined to form a two-dimensional geometric image. Preferably, eight triangles on each surface of the octahedron are converted into right-angled isosceles triangles, and these right-angled isosceles triangles are combined with each other to form a square or square.

In particular, for 2D images or square images, in order to prevent the image from being cut off at the edges and corners, copies of each triangle corresponding to the domain of the octahedron are rotated up and down, and each square is attached. Make it form.

Referring to the example of FIG. 9, (c) mapping from octahedral data to (d) square data is shown.

Next, a descriptor of a 2D geometric image is generated for the mapped 2D rectangle (S35).

In the previous steps, three-dimensional shape data, mapping data on a sphere, mapping data on a corrected sphere, mapping data on an octahedron, and mapping data on a two-dimensional rectangle have a one-to-one mapping relationship with each other. Accordingly, the 3D shape data and the 2D rectangle are mapped to each other, and in particular, each vertex of the 3D shape data is mapped to each vertex of the 2D rectangle.

A pixel value of each pixel of a 2D image or a square is set as a characteristic value (or geometric characteristic value) of a 3D shape at a location of 3D shape data mapped to a pixel location. That is, the pixel position (x',y') of the 2D image (or 2D square) is mapped to a point (x,y,z) of the 3D shape data. In the case of a pixel other than a vertex, a corresponding position can be found by interpolation or the like.

Preferably, the geometric characteristic value of the 3D shape data is one or more of the x-coordinate value of the 3D x-axis, the y-coordinate value of the y-axis, the z-coordinate value of the z-axis, and the main curvature value (maximum curvature value, minimum curvature value). use. Preferably, all five geometric feature values are used.

Meanwhile, each geometric characteristic value may be normalized to a numerical range (0 to 255) of the pixel.

Therefore, a two-dimensional geometric image having each geometric feature value as a pixel value is generated for each geometric feature value. Preferably, all K (eg, 5) two-dimensional geometric images are generated.

Next, the individual model learning unit 332 generates primary training data from the extracted protein and compound descriptors and activity data of the compound for the protein, and uses its own individual neural network model 41 as part of the primary training data. To learn.

In particular, when the descriptor of the protein and the descriptor of the compound are separately configured, the individual model learning unit 332 combines the descriptor of the protein and the descriptor of the compound with respect to the pair of the protein and the compound to form an entire descriptor.

For example, in the case of the Fourier method, the entire descriptor is a one-dimensional vector having 2n elements by combining two vectors each having n elements for a target protein and a corresponding compound.

In addition, in the case of the homology method, the entire descriptor is an image that combines a homology image of a target protein and a homology image of a corresponding compound.

In addition, in the case of the geometric transformation method, the entire descriptor is K two-dimensional geometric images having pixel values as geometric characteristic values (normalized characteristic values) of 3D shape data, and is an image that combines each of proteins and compounds. K is the number of geometric feature values (number of types).

In the following, for convenience of description, the entire descriptor is mixed with the descriptor.

In addition, the individual model learning unit 332 generates primary learning data by labeling the descriptor (or the entire descriptor) for the pair of the protein and the compound and the activity data of the corresponding compound for the corresponding protein to the corresponding descriptor. That is, the activity data is used as a label.

In addition, the individual model learning unit 332 divides the primary training data into two groups. One group (or the first group) is used to train its individual neural network model 41, and the other group (or the second group) is used to generate result data for secondary learning.

In addition, the individual model learning unit 332 trains its own individual neural network model 41 with some (or first group) of the primary training data. In this case, the first group may be divided into training data and verification data (or test data) therein.

Next, the result data generation unit 333 applies the second group to the trained individual neural network model 41 to generate result data for secondary learning. In other words, the descriptor among the first training data of the second group is applied to the individual neural network model 41 to obtain the prediction result data, and the label of the primary training data corresponding to the prediction result data (the label of the corresponding descriptor) is assigned. Generate result data for secondary learning.

That is, the result data for secondary learning is composed of information on proteins and compounds (identification information such as ID), prediction result data, and labels. In other words, in order to identify each prediction result data, there must be a protein ID and a compound ID serving as a data identification number for each data.

In particular, the individual neural network model 41 calculates and outputs a softmax value capable of expressing predictability for each data (or label value indicating classification). Therefore, the prediction result data is composed of a probability value (or softmax value) of active or inactive.

Next, a detailed configuration of the integrated neural network module 34 according to an embodiment of the present invention will be described in more detail with reference to FIG. 3.

As shown in Figure 3, the integrated neural network module 34 is a data integration unit 341 for generating secondary training data by integrating the result data for secondary learning generated by each individual neural network module 33, and, It consists of an integrated model learning unit 342 that trains the integrated neural network model 42 with training data.

First, the data integration unit 341 combines and integrates the generated secondary learning result data based on a pair of a protein and a compound. That is, if the pair of protein and compound is the same, the result data for secondary learning of different individual neural network modules are combined into one data and integrated. In particular, for one pair of protein and compound, secondary training data is generated by combining the secondary training result data of all individual neural network modules. Since the protein and compound are combined by a pair (or identification information thereof), the order of data output from each individual neural network module 33 does not necessarily have to be identical.

Preferably, the data integration unit 341 collects result values of the softmax function output from each individual neural network model 41, and combines the softmax result values to form secondary learning data. That is, when the softmax value output from each individual neural network model 41 is collected, it is integrated based on the data identification number (protein ID, compound ID, etc.).

In addition, the secondary learning data is comprised of the identification information of each protein and compound.

Therefore, the secondary learning data is composed of identification information of each protein and compound, prediction result values of all individual neural network models 41 for the protein and compound, and activity data (or label value) of the protein and compound. In particular, the predicted result value is a result value of a softmax function.

An example of secondary learning data is shown in the table of FIG. 10. In FIG. 10, there are two categories of labels, and each individual neural network model 41 exemplifies three cases. The number of label categories and the number of individual neural network models 41 can be applied variably.

Next, the integrated model learning unit 342 trains the integrated neural network model 42 by using the second learning data generated by integration. In this case, the secondary learning data may be divided into learning data and verification data (or test data) therein.

Next, the effects of the present invention will be described with reference to FIGS. 11 to 13.

As described above, the integrated neural network module 34 may independently collect the softmax value output from the individual neural network module 33 and integrate the collected data. Therefore, even if the development schedule of each individual neural network module 33 is different, it is sufficient to learn the integrated neural network model 42 as soon as the individual neural network module 33 produces data. In addition, after securing individual data, the integrated neural network model 42 may be trained by separating it from the individual neural network model 41 or the individual neural network module 33.

The activity prediction system according to the present invention has the following effects.

That is, the entire development period and individual artificial neural network learning time can be shortened.

In addition, it is possible to improve the efficiency of using the artificial intelligence server system. In other words, individual artificial neural networks can be trained in an available time without simultaneously learning them. In addition, system availability is improved because there is no need to load individual artificial neural networks into memory at the same time.

In addition, as shown in FIGS. 11 to 13, it was confirmed that the discrimination power for inactive can be improved (here, an example of applying two individual artificial intelligence models is shown).

The present invention implements a method of improving data predictability by configuring a separate integrated artificial neural network (ensemble) by integrating the output files after receiving the results of learning a plurality of individual artificial neural networks. The present invention is not limited to a descriptor technique expressing a three-dimensional structure of a protein and a compound, but can be applied to all fields integrating a classification-type artificial neural network.

In the above, the invention made by the present inventors has been described in detail according to embodiments, but the invention is not limited to the embodiments, and it goes without saying that various changes can be made without departing from the gist of the invention.

10: analysis terminal 20: network
30: active prediction system 31: data management module
311: active data collection unit 312: 3D data generation unit
32: individual neural network model group 33: individual neural network model
331: descriptor extraction unit 332: individual model learning unit
333: result data generation unit
34: integrated neural network model 341: data integration unit
342: integrated model learning unit 35: active prediction module
40: database 41: individual neural network model
42: integrated neural network model 43: data storage for secondary learning

Claims (7)

In a system for predicting the activity of a protein-binding compound based on a plurality of artificial neural network models,
A data management module that stores activity data of a compound for a protein, and data of a three-dimensional shape or three-dimensional structure of the protein and the compound;
Equipped with its own individual neural network module, extract a descriptor from 3D shape data or 3D structure data of a protein or compound, train the individual neural network model with the extracted descriptor and active data, and apply it to the trained individual neural network model An individual neural network module that stores the output data for secondary learning;
An integrated neural network that has its own integrated neural network module, generates secondary training data by integrating the result data for secondary training generated by each individual neural network module, and trains the integrated neural network model with the generated secondary training data. module; And,
Including an activity prediction module for predicting the activity of the query protein and the query compound using the individual neural network model and the integrated neural network model,
The individual neural network model is composed of at least two or more to form one individual neural network module group. A system for predicting the activity of a protein-binding compound based on a plurality of artificial neural network models.
The method of claim 1,
The individual neural network module generates primary training data from the extracted descriptor and active data, trains the individual neural network model as part of the primary training data, and applies another part of the primary training data to the trained individual neural network model. A system for predicting activity of protein-binding compounds based on a plurality of artificial neural networks, characterized in that generating result data for secondary learning.
The method of claim 1,
The primary learning data includes identification information of proteins and compounds, descriptors for proteins and compounds, and a label value determined from activity data of the corresponding compound for the protein,
The individual neural network module obtains prediction result data by applying a descriptor of the corresponding data to the individual neural network model for each data of the primary training data, and assigns a label of the corresponding data of the corresponding primary training data to the prediction result data. A system for predicting activity of protein-binding compounds based on a plurality of artificial neural networks, characterized in that generating result data for secondary learning and including identification information on proteins and compounds in the secondary learning result data.
The method of claim 1,
The output data of the individual neural network model is output as a probability value of a label value,
The integrated neural network module combines and integrates the result data for secondary learning based on the pair of protein and compound, but if the pair of protein and compound match, the result data for secondary learning of all individual neural network modules are combined into one data. A system for predicting the activity of protein-binding compounds based on a plurality of artificial neural network models, characterized in that for generating primary learning data.
The method of claim 1,
The individual neural network module group comprises at least one individual neural network module for generating a descriptor from 3D shape data, and at least one individual neural network module for generating a descriptor from 3D structure data. A system for predicting the activity of protein-binding compounds based.
The method of claim 1,
The individual neural network module extracts one-dimensional data from three-dimensional shape data using Fourier transform and generates it as a descriptor, or extracts a two-dimensional image from three-dimensional structure data using homology and generates a descriptor, or three-dimensional shape. A system for predicting the activity of a protein-binding compound based on a plurality of artificial neural network models, characterized in that a descriptor is generated by projecting data into a two-dimensional image.
The method of claim 1,
The 3D shape data is 3D data on a closed mesh surface as closed surface data, and the 3D structure data is 3D data on the atomic position in the chemical structure of a protein or compound. A system for predicting the activity of protein-binding compounds based on a plurality of artificial neural network models, characterized in that the data consists of.

Images