CN115831246A

CN115831246A - Pharmaceutical chemical reaction synthesis and conversion rate prediction combined optimization method

Info

Publication number: CN115831246A
Application number: CN202211548092.8A
Authority: CN
Inventors: 侯静怡; 刘志杰; 阿力夫; 贺威; 唐宇鑫; 刘家炜
Original assignee: University of Science and Technology Beijing USTB
Current assignee: University of Science and Technology Beijing USTB
Priority date: 2022-12-05
Filing date: 2022-12-05
Publication date: 2023-03-21

Abstract

The invention discloses a combined optimization method for synthesis of a pharmaceutical chemical reaction and prediction of conversion rate, which comprises the following steps: acquiring a SMILES expression of a reactant, segmenting the SMILES expression of the reactant, and performing embedded expression on the segmentation to obtain segmentation characteristics of the SMILES expression of the reactant; carrying out hierarchical sequence coding on the word segmentation characteristics of the SMILES expression of the reactant; the two tasks of chemical reaction synthesis prediction and conversion rate prediction are combined, and are trained simultaneously, so that the combined optimization of chemical reaction synthesis and conversion rate prediction is realized. The method combines two tasks of chemical reaction synthesis prediction and conversion rate prediction, introduces a hierarchical sequence modeling technology, mutually guides model interpretable parameter optimization, and improves the training efficiency and performance of the model. And the invention introduces uncertainty estimation to deal with the interference brought to the model by uncertain data under the real condition.

Description

Pharmaceutical chemical reaction synthesis and conversion rate prediction combined optimization method

Technical Field

The invention relates to the technical field of pharmaceutical chemical reaction synthesis and conversion rate prediction, in particular to a combined optimization method of pharmaceutical chemical reaction synthesis and conversion rate prediction.

Background

With the advent of the big data era, the ability to analyze a large amount of existing chemical data through deep learning is highly desired, and models for predicting various aspects of chemical reactions are derived from the data, wherein chemical reaction synthesis prediction and conversion rate prediction are more critical and difficult problems in chemical reaction problems. Chemical reaction synthesis is a process of constructing a target product from a set of existing reactants and reagents, and since chemical reaction takes a lot of time and money, it is very important to predict the product of chemical reaction by machine. For the general process of chemical reaction synthesis reasoning, the model constructs the product by analyzing the chemistry of the reactants and combining the reaction conditions. In order to subsequently apply the extension to new chemical reactants, the model is required to have stronger adaptability to unusual chemical reaction types and even new classes. Chemical reaction synthesis prediction tasks, for which a recurrent neural network and a attention-based Transformer model are typically utilized, are to predict the products in an expression given the reactants in the expression of a pharmaceutical chemical reaction. The cyclic neural network emphasizes the long-time context dependence of the chemical reaction type, and the Transformer model mainly focuses on global structure information.

Chemical reaction conversion rate prediction is a problem of deducing the reactant conversion rate of a chemical reaction from a chemical reaction equation, wherein the conversion rate is the ratio of the amount of a converted reactant to the total amount of the reactant, and chemists tend to find a reaction path with higher conversion rate in consideration of economic and time factors. Similar to chemical reaction synthesis prediction, chemical reaction conversion prediction is generally based on a recurrent neural network and a Transformer.

At present, the two tasks are researched based on deep learning by adopting a basic network architecture and adopting a large amount of data to train a model so as to obtain a satisfactory result. However, in a real situation, in consideration of safety and reliability requirements of pharmaceutical chemistry production, there is a need to develop interpretable model structures and learning algorithm designs; in addition, the existing method is lack of a corresponding processing mechanism for solving the problems of high uncertainty of a chemical formula product synthesis result and conversion rate caused by measurement errors, experimental environment change and the like in real data.

After the chemical molecular formula is converted into the SMILES format, the existing sequence model algorithm can be naturally transplanted into the tasks of chemical reaction synthesis prediction and conversion rate prediction. Currently, chemical reaction synthesis prediction and conversion rate prediction are commonly used by using a model of a recurrent neural network and a Transformer based on an attention mechanism.

The recurrent neural network is a neural network with short-term memory capability, and the neurons can receive information of other neurons and information of the neurons to form a network structure with loops. The model is naturally adapted to process sequence information by recursively computing the output at the next time step based on the output at the previous time step, and finally outputting sequence data of arbitrary length, training the network with linear sequences of chemical formulae under the SMILES rule, and guiding the model to find the products of the chemical reaction. However, the common recurrent neural network can only model a time sequence process, and it is difficult to extract hierarchical structure information of chemical reactions, and information of different levels such as functional groups and chemical bonds in chemical reaction formulas is very critical to expression of chemical formulas. The Transformer is a model based on a multi-head attention mechanism, the structure of the model abandons the traditional CNN and RNN, and the whole network structure is completely formed based on the attention mechanism. In the problem of chemical reaction synthesis prediction, a Transformer-based method is not modeled by hidden information transfer, but directly calculates the relationship between each sequence feature, can synchronously complete information transfer between chemical molecule sequences, focuses on global feature representation of the sequences, and also ignores local hierarchical information expression.

In summary, the prior art ignores that the two tasks of chemical reaction synthesis prediction and conversion rate prediction can provide interrelated information for chemical reaction synthesis prediction and conversion rate prediction, namely, the two tasks have high correlation with chemical bond breakage and functional group recombination, and the two tasks are actually beneficial to guiding model parameters to the expected interpretable target optimization through combined learning. In addition, the prior art has not considered how to process uncertain data.

Disclosure of Invention

The invention provides a combined optimization method for synthesis of a medicinal chemical reaction and prediction of a conversion rate, which aims to solve the technical problems that the existing prediction method of the medicinal chemical reaction ignores two tasks of synthesis prediction and conversion rate prediction of the chemical reaction and can provide correlated information, and the existing prediction method of the medicinal chemical reaction has poor interpretability and does not consider uncertain factors in the chemical reaction.

In order to solve the technical problems, the invention provides the following technical scheme:

in one aspect, the invention provides a method for jointly optimizing synthesis and conversion rate prediction of a medicinal chemical reaction, which comprises the following steps:

acquiring a SMILES expression of a reactant of a medicinal chemical reaction, segmenting the SMILES expression of the reactant, and performing embedded expression on the segmentation to obtain segmentation characteristics of the SMILES expression of the reactant;

carrying out hierarchical sequence coding on the word segmentation characteristics of the SMILES expression of the reactant to obtain hierarchical sequence coding characteristics of the reactant so as to improve the weight of atoms around chemical bonds which change in the chemical formula;

combining two tasks of medicinal chemical reaction synthesis prediction and conversion rate prediction, and simultaneously training the two tasks to realize the combined optimization of the medicinal chemical reaction synthesis and the conversion rate prediction; the synthesis prediction task predicts a product based on the hierarchical sequence coding characteristics of the reactants; the conversion rate prediction task predicts the conversion rate based on the hierarchical sequence coding features of the reactants and the hierarchical sequence coding features of the products output by the synthesis prediction task.

Further, hierarchically and sequentially encoding the participle characteristics of the SMILES expression of the reactant comprises:

the participle features of the SMILES expression for the reactant are hierarchically sequence coded using ON-LSTM.

Further, hierarchically and sequentially coding the word segmentation characteristics of the SMILES expression of the reactant, the method further comprises the following steps:

carrying out hierarchical coding ON SMILES expressions of reactants by utilizing three layers of ON-LSTM, and learning hierarchical coding characteristics by adopting a middle layer; the hierarchical coding features are obtained by weighted addition of intermediate layer hidden variable features.

Further, when the hierarchical coding features are obtained by weighted addition of the intermediate-layer hidden variable features, the weight is quantified as the probability of chemical bond breakage in the pharmaceutical chemical reaction.

Further, the two tasks of the synthesis prediction and the conversion rate prediction of the pharmaceutical chemical reaction are combined, and are trained simultaneously to realize the combined optimization of the synthesis of the pharmaceutical chemical reaction and the conversion rate prediction, and the method comprises the following steps:

combining two tasks of the synthesis prediction and the conversion rate prediction of the pharmaceutical chemical reaction, simultaneously training the two tasks, and constructing a combined prediction model of the synthesis and the conversion rate of the pharmaceutical chemical reaction; it includes:

adopting ON-LSTM as a generation network, and completing the prediction of a product based ON the hierarchical sequence coding characteristics of reactants; the input of the previous moment during the network generation training directly adopts a true value to prevent error accumulation, and a product sequence generated by cluster search regression is adopted during inference, so that a compound inverse synthesis task is completed;

after the prediction result of the product is obtained, calculating the hierarchical sequence coding characteristics of the product; after the hierarchical sequence coding features of the products are obtained, splicing and fusing the hierarchical sequence coding features of the reactants and the hierarchical sequence coding features of the products to obtain the global features of the chemical reaction expression;

based on the obtained global characteristics of the chemical reaction expression, adopting a conversion rate prediction network to predict the conversion rate of the chemical reaction of the medicine; in the training process, the conversion rate prediction value is sampled from normal inverse gamma distribution by utilizing evidence deep learning, and the model is trained by adopting a multi-task loss function method, so that the model has uncertainty estimation capability while accurately predicting.

Further, the generation network is a three-layer ON-LSTM.

Further, calculating a hierarchical sequence encoding feature of the product of generating the network output, comprising:

extracting hidden variable characteristics of the intermediate layer of the generated network;

and obtaining the hierarchical coding characteristics of the product by weighted addition of the hidden variable characteristics of the intermediate layers of the generated network.

Further, conversion rate prediction values p-NIG (alpha, beta, gamma, nu) sampled from normal inverse gamma distribution; wherein, α, β, γ, and ν are parameters in distribution, γ is more than 0 and less than 1, β is more than 0, ν is more than 0, α is more than 1, γ and α, β, ν are respectively obtained by calculation of two independent multilayer perceptrons, according to constraint conditions, outputs of two networks are respectively through Sigmoid and Softplus activation functions, and the optimization target of model integral uncertainty estimation is:

L＝L _BM +kL _UE

wherein L is _BM Estimating the cross entropy loss of gamma and a true value for a point, wherein k is a hyperparameter;

L _UE ＝L _NLL +L _LMSE +C

wherein L is _NLL Is the negative logarithm of the marginal likelihood, L _LMSE And C is an evidence regularization term.

In yet another aspect, the present invention also provides an electronic device comprising a processor and a memory; wherein the memory has stored therein at least one instruction that is loaded and executed by the processor to implement the above-described method.

In yet another aspect, the present invention also provides a computer-readable storage medium having at least one instruction stored therein, the instruction being loaded and executed by a processor to implement the above method.

The technical scheme provided by the invention has the beneficial effects that at least:

the method has the advantages that the two tasks of chemical reaction synthesis prediction and conversion rate prediction are combined, the two tasks are trained simultaneously, the two tasks mutually guide and interact to update the network, the interpretability is strong, and the uncertainty data is robust. The method for jointly optimizing the medicinal chemical reaction synthesis and the conversion rate prediction has obvious application value in the scenes of reducing the research and development cost of the compound and improving the research and development efficiency of the compound, and lays a model foundation for realizing comprehensive and accurate reaction synthesis of the compound and wide application of the conversion rate prediction.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.

FIG. 1 is a schematic diagram of an implementation principle of a pharmaceutical chemistry reaction synthesis and conversion rate prediction joint optimization method provided by an embodiment of the invention;

FIG. 2 is a schematic flow chart of the implementation of the method for optimizing the combination of the synthesis of the pharmaceutical chemical reaction and the prediction of the conversion rate provided by the embodiment of the invention;

FIG. 3 is a molecular diagram of ciprofloxacin.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

First embodiment

Aiming at the problems that the existing prediction method of the medicinal chemical reaction is not strong in interpretability and uncertainty factors in the chemical reaction are not considered, the embodiment provides a medicinal chemical reaction synthesis and conversion rate prediction joint optimization method based on uncertainty estimation, and as shown in fig. 1, the method comprises three parts of chemical formula SMILES format conversion, hierarchical sequence modeling and joint optimization based on uncertainty; the method is characterized in that two interrelated tasks of medicinal chemical reaction synthesis prediction and conversion rate prediction are jointly learned, interpretable guide information of a model is given based on the tasks in the learning process, and an uncertainty estimation method is adopted to train the model.

Specifically, the execution flow of the method is shown in fig. 2, and includes the following steps:

s1, obtaining a SMILES expression of a reactant of a medicinal chemical reaction, performing word segmentation on the SMILES expression of the reactant, and performing embedded expression on the word segmentation to obtain word segmentation characteristics of the SMILES expression of the reactant;

it should be noted that, the method of this embodiment firstly needs to convert the reactants and products of the chemical reaction from a molecular diagram to a molecular SMILES format, for example, for ciprofloxacin in fig. 3, the SMILES format is: n1CCN (CC 1) C (F) = C2) = CC (= C2C4= O) N (C3 CC 3) C = C4C (= O) O. And performing word segmentation on the SMILE expression of the reactant, performing embedded expression on the word segmentation to obtain word segmentation characteristics of the SMILE expression of the reactant, and connecting the embedded matrix with the model in the subsequent step to perform end-to-end learning.

S2, carrying out hierarchical sequence coding on the word segmentation characteristics of the SMILES expression of the reactant to obtain hierarchical sequence coding characteristics of the reactant so as to improve the weight of atoms around chemical bonds which change in the chemical formula;

it should be noted that, the method of this embodiment needs to perform sequence coding on the word segmentation features of the expression of the chemical formula SMILES. Because the sequence of atoms is generally not disturbed completely in the process of chemical reaction, but some chemical bonds are broken and functional groups are recombined, the information of the functional groups with broken chemical bonds needs to be learned in a self-adaptive manner, so that the information extraction capability of the model is improved, and the calculation efficiency is improved. For this purpose, this embodiment utilizes ON-LSTM (ordered nerves LSTM) to perform multi-level coding ON the ON-LSTM, to increase the weight of atoms around the chemical bond that changes in the chemical formula, and to express the chemical expression according to a certain semantic level. The ON-LSTM is a language model that can adaptively encode various hierarchical structures, such as complete sentence information, phrase hierarchical information, and word segmentation level information, and in fact, these hierarchical structures are implicit and have no explicit meaning, and after the model is trained based ON different tasks, its hierarchical semantics will exhibit characteristics related to the tasks. The ON-LSTM is a variant of the long-short memory network LSTM, and the input and the output of the ON-LSTM are the same as the LSTM.

Specifically, in this embodiment, the SMILES expression corresponding to the reactant chemical formula is hierarchically encoded by using three layers of ON-LSTM, and since the first layer and the third layer are affected by input and output and cannot learn a good hierarchical characteristic, this embodiment adopts an intermediate layer to learn the hierarchical encoding characteristic. Let the intermediate layer hidden variable be characterized by h _t The hierarchical coding features are obtained by weighted addition of the hidden variable features:

wherein, a _t According to master forget gate f in ON-LSTM for weighting _t And calculating to obtain:

wherein D represents the characteristic dimension of a forgetting gate. Weight a _t The larger the score, the more likely it is to be a segmentation point of high level semantics. In this embodiment, since the model is based on the study of the chemical reaction synthesis prediction and conversion rate prediction tasks, the weight is the possibility of chemical bond breakage in the chemical reaction of the drugAnd (5) sex quantization.

S3, combining the drug chemical reaction synthesis prediction task and the conversion rate prediction task, and simultaneously training the two tasks to realize the combined optimization of the drug chemical reaction synthesis and the conversion rate prediction; wherein the synthesis prediction task predicts a product based on hierarchical sequence coding features of the reactants; the conversion prediction task predicts conversion based on hierarchical sequence coding features of the reactants and hierarchical sequence coding features of the products output by the synthetic prediction task.

It should be noted that, because the chemical reaction synthesis process and the conversion rate of the chemical reaction both depend on the breakage of the chemical bond, the two interrelated tasks of the pharmaceutical chemical reaction synthesis and the conversion rate prediction are jointly learned, and the expression ability and the effectiveness of the model can be improved. The embodiment unsupervised locates the position of chemical bond breaking through the functional group hierarchy in the hierarchical chemical formula hierarchical sequence modeling, so as to find the key information influencing the synthesis and conversion rate of the chemical reaction, and the model has interpretability.

Based on the above, in this embodiment, the implementation process of S3 is as follows:

1. synthetic prediction of pharmaceutical chemistry

After hierarchical sequence coding is carried out ON reactants, an ON-LSTM is adopted as a sequence generation network model, and product prediction is completed based ON hierarchical sequence coding characteristics of the reactants; during training, the input of the previous moment directly adopts a true value to prevent error accumulation, and during deduction, a product sequence generated by clustering search regression is adopted to complete a compound inverse synthesis task; and after the prediction result of the product is obtained, calculating the hierarchical sequence coding characteristics of the product.

In particular, in this embodiment, the generation of the product is achieved using three layers of ON-LSTM as decoders, i.e., generating a network model. In the training process, the input of the decoder is the hierarchical coding characteristic x of the reactant _enc At each time of decoderInputting the moment, namely giving the SMILES expression characteristic of the participle level at the last moment, outputting the characteristic at the current moment, extracting the hidden variable characteristic of the middle layer of the decoder aiming at the next reaction conversion rate prediction, and calculating the hierarchical sequence coding y of the product in the same way as the calculation of the hierarchical coding characteristic of the reactant in the S2 _enc . Product sequences generated by cluster search regression were used in the tests.

2. Prediction of conversion rate of pharmaceutical chemical reaction

After the hierarchical sequence coding features of the products are obtained, the hierarchical sequence coding features of the reactants and the hierarchical sequence coding features of the products are spliced and fused to obtain a global feature g = [ x ] of a chemical reaction expression _enc ，y _enc ](ii) a After the expression global characteristics are obtained, based on the obtained expression global characteristics, adopting a conversion rate prediction network model to predict the conversion rate of the medicinal chemical reaction based on uncertainty; in the training process, the conversion rate prediction value is sampled from normal inverse gamma distribution by utilizing evidence deep learning, and the model is trained by adopting a multi-task loss function method, so that the model has uncertainty estimation capability while accurately predicting.

Specifically, in this embodiment, in order to estimate the uncertainty existing between the reaction process and the real-world environment, the proposed model is trained by sampling the conversion rate prediction value p from the normal inverse gamma distribution using evidence deep learning:

p～NIG(α，β，γ，ν)，

wherein (alpha, beta, gamma and delta) are parameters in distribution, gamma is more than 0 and less than 1, v, beta is more than 0, alpha is more than 1, gamma and alpha, beta and v are respectively calculated by two independent multilayer perceptrons, and the output of the two networks respectively passes through a Sigmoid and Softplus activation function according to constraint conditions. The optimization objective of the overall uncertainty estimation is:

L＝L _BM +kL _UE ，

L _UE ＝L _NLL +L _LMSE +C

wherein L is _NLL Is the negative logarithm of the marginal likelihood, L _LMSE And C is an evidence regularization term, namely a Mean Square Error (MSE) loss function of the Lipschitz correction.

In conclusion, the method combines two tasks of chemical reaction synthesis prediction and conversion rate prediction, introduces a hierarchical sequence modeling technology, and the two tasks mutually guide model interpretable parameter optimization, thereby improving the training efficiency and performance of the model. Moreover, the method of the embodiment also introduces uncertainty estimation to deal with the interference brought to the model by uncertain data under the real condition. Therefore, the method has obvious application value in the scenes of reducing the research and development cost of the compound and improving the research and development efficiency of the compound, and lays a model foundation for realizing comprehensive and accurate reaction synthesis of the compound and wide application of conversion rate prediction.

Second embodiment

The present embodiment provides an electronic device, which includes a processor and a memory; wherein the memory has stored therein at least one instruction that is loaded and executed by the processor to implement the method of the first embodiment.

The electronic device may have a relatively large difference due to different configurations or performances, and may include one or more processors (CPUs) and one or more memories, where at least one instruction is stored in the memory, and the instruction is loaded by the processor and executes the method.

Third embodiment

The present embodiment provides a computer-readable storage medium, in which at least one instruction is stored, and the instruction is loaded and executed by a processor to implement the method of the first embodiment. The computer readable storage medium may be, among others, ROM, random access memory, CD-ROM, magnetic tape, floppy disk, optical data storage device, and the like. The instructions stored therein may be loaded by a processor in the terminal and perform the above-described method.

Furthermore, it should be noted that the present invention may be provided as a method, apparatus or computer program product. Accordingly, embodiments of the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, embodiments of the present invention may take the form of a computer program product embodied on one or more computer-usable storage media having computer-usable program code embodied in the media.

Embodiments of the present invention are described with reference to flowchart illustrations and/or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, embedded processor, or other programmable data processing terminal to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing terminal to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal to cause a series of operational steps to be performed on the computer or other programmable terminal to produce a computer implemented process such that the instructions which execute on the computer or other programmable terminal provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

It should also be noted that, in this document, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or terminal. Without further limitation, an element defined by the phrases "comprising one of \ 8230; \8230;" does not exclude the presence of additional like elements in a process, method, article, or terminal device that comprises the element.

Finally, it should be noted that while the above describes a preferred embodiment of the invention, it will be appreciated by those skilled in the art that, once the basic inventive concepts have been learned, numerous changes and modifications may be made without departing from the principles of the invention, which shall be deemed to be within the scope of the invention. Therefore, it is intended that the appended claims be interpreted as including preferred embodiments and all such alterations and modifications as fall within the scope of the embodiments of the invention.

Claims

1. A combined optimization method for the synthesis of a pharmaceutical chemical reaction and the prediction of conversion rate is characterized by comprising the following steps:

combining two tasks of the synthesis prediction and the conversion rate prediction of the pharmaceutical chemical reaction, and simultaneously training the two tasks to realize the combined optimization of the synthesis of the pharmaceutical chemical reaction and the conversion rate prediction; the synthesis prediction task predicts a product based on the hierarchical sequence coding characteristics of the reactants; the conversion rate prediction task predicts the conversion rate based on the hierarchical sequence coding features of the reactants and the hierarchical sequence coding features of the products output by the synthesis prediction task.

2. The method of claim 1, wherein the hierarchical sequence coding of the participle features of the SMILES expression of the reactant comprises:

the participle features of the SMILES expression of the reactant are hierarchically sequence coded using ON-LSTM.

3. The method of claim 1, wherein the hierarchical sequence coding of the participle features of the SMILES expression of the reactant comprises:

4. The method for joint optimization of synthesis of pharmaceutical chemical reaction and prediction of conversion rate as claimed in claim 3, wherein the weight is quantified for the probability of chemical bond rupture in pharmaceutical chemical reaction when the hierarchical coding features are obtained by weighted addition of intermediate-layer latent variable features.

5. The method for jointly optimizing the synthesis of pharmaceutical chemical reaction and the prediction of conversion rate according to claim 1, wherein the joint optimization of the synthesis of pharmaceutical chemical reaction and the prediction of conversion rate by combining the two tasks and simultaneously training the two tasks comprises:

adopting ON-LSTM as a generation network, and completing the prediction of a product based ON the hierarchical sequence coding characteristics of reactants; the input of the previous moment during the network training is directly true value to prevent error accumulation, and the product sequence generated by clustering search regression is adopted during deduction, so that the compound inverse synthesis task is completed;

6. The method for combinatorial optimization of synthesis of pharmaceutical chemistry reactions and prediction of conversion ratios according to claim 5, wherein the generating network is a three-layer ON-LSTM.

7. The method of claim 6, wherein calculating the hierarchical sequence coding features of the products of generating the network output comprises:

8. The method of claim 5, wherein the conversion ratio prediction value p-NIG (α, β, γ, v) is sampled from a normal inverse gamma distribution; wherein, alpha, beta, gamma and ν are parameters in distribution, gamma is more than 0 and less than 1, beta is more than 0, v >, alpha is more than 1, gamma and alpha, beta and ν are respectively obtained by calculation of two independent multilayer perceptrons, according to constraint conditions, the output of two networks respectively passes through Sigmoid and Softplus activation functions, and the optimization target of the model integral uncertainty estimation is as follows:

L＝L _BM +kL _UE

L _UE ＝L _NLL +L _LMSE +C