Drug chemical reaction type prediction method based on multi-level information fusion

Abstract

The application discloses a drug chemical reaction type prediction method based on multi-level information fusion, which comprises the following steps: extracting structure information of a molecular level in a drug chemical reaction; extracting reaction formula level sequence information; fusing the extracted structure information of the molecular level in the drug chemical reaction and the extracted reaction formula level sequence information; and finally obtaining a predicted drug chemical reaction category. The application not only considers the information of the whole reaction formula, but also pays attention to the structure information of local molecules, so that the model can reasonably select and complement each other between the whole information and the local information, and the accuracy of drug chemical reaction type prediction is improved.

Description

Technical Field

[0001] This invention relates to the technical field of drug chemical reaction type prediction, and in particular to a drug chemical reaction type prediction method, system, computer, and storage medium based on multi-level information fusion. Background Technology

[0002] Medicinal chemistry reactions are processes that transform reactants into one or more products. The medicinal chemistry synthesis toolbox currently offers hundreds of different types of reactions to choose from. Classifying medicinal chemistry reactions is crucial for reaction retrieval and the classification of new reactions. Previously, medicinal chemistry reactions were often classified based on manual rules and experience, a method that was not only time-consuming and labor-intensive but also prone to errors. Since the development of machine learning methods, data-driven machine learning prediction methods have become an inevitable trend in various fields, replacing manual experience.

[0003] Existing work on classifying drug chemical reaction types using machine learning methods can be broadly divided into two stages: traditional machine learning models based on the characteristics of drug chemical reactions and deep learning models based on Simplified Molecular Input Line Entry Specifications (SMILES). In the first stage, representative models utilize density functional theory (MFF) to calculate and encode the vibrational properties of each reactant, catalyst, and product molecule or atom, and then use machine learning models such as random forests and support vector machines to predict drug chemical reaction categories. In the second stage, drug chemical expressions are transformed into Simplified Molecular Input Line Entry Specifications (SMILES), and word segmentation is performed. Subsequent work using various natural language processing methods to predict drug chemical reaction categories has also achieved excellent results. In particular, Schwaller P et al., based on the task of predicting drug chemical reaction categories, fine-tuned and trained an English BERT model using SMILES sequence data from two datasets (Pistachio and USPTO) with a total volume of 180,000 sequences, resulting in the BERT model RXNFP, which is particularly capable of understanding SMILES sequences in the field of chemical reactions. This pre-trained model has played a significant role in all downstream tasks related to drug chemistry reactions.

[0004] While deep learning models based on SMILES sequences have achieved excellent performance, they fail to consider the unique characteristics of drug molecules. They merely represent drug molecules as specific symbols for specific molecule recognition, losing the inherent features of the molecules. Among the unique features of molecules, molecular structure is a crucial indicator. Molecules with the same structure but different chemical elements generally have similar reactive properties, while molecules with the same elements but different structures exhibit vastly different chemical properties. Therefore, molecular structure is a vital characteristic of reactive molecules, yet methods based on SMILES sequences have never considered molecular structure. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for predicting drug chemical reaction types based on multi-level information fusion. This method not only considers the overall information of the reaction formula but also focuses on the structural information of local molecules. It enables the model to make reasonable selections and mutual supplementation between overall and local information, thereby improving the accuracy of drug chemical reaction type prediction.

[0006] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0007] A method for predicting drug chemical reaction types based on multi-level information fusion includes:

[0008] Extracting structural information at the molecular level from medicinal chemical reactions;

[0009] Extract reaction hierarchy sequence information;

[0010] The molecular-level structural information and the reaction-level sequence information extracted from the drug chemical reaction are fused together to obtain the predicted drug chemical reaction category.

[0011] Furthermore, extracting structural information at the molecular level in medicinal chemical reactions includes:

[0012] Encoding the molecular-level structure in pharmaceutical chemical reactions:

[0013] Let R be a reaction equation in a pharmaceutical chemistry reaction, and let B be the set of reactants and products in that reaction equation, where B = {B1, B2, ..., B}. n}, where n is the number of molecules of reactants and products in the reaction equation;

[0014] Extract the structural information of each molecule in B.

[0015] Furthermore, extracting structural information for each molecule in B includes:

[0016] The adjacency matrix A of each molecule is obtained using the molecular data processing package Rdkit.i The embedding matrix H of each atomic node in the molecule i and degree matrix D i ; where the adjacency matrix A of each molecule is i The embedding matrix H represents the connections between atoms, and is the embedding matrix of each atomic node in the molecule. i The degree matrix D represents the characteristics of each atom. i This represents the probability that each atom is connected to other atoms;

[0017] Based on the adjacency matrix A of each molecule i The embedding matrix H of each atomic node in the molecule i and degree matrix D i Extract the structural information of each molecule.

[0018] Furthermore, based on the adjacency matrix A of each molecule i The embedding matrix H of each atomic node in the molecule i and degree matrix D i The structural information of each molecule is extracted using a convolutional neural network. The extraction formula is as follows:

[0019]

[0020] in I is the identity matrix, W is the network parameters, and σ(·) is the activation function. This is the output of this layer and also the input of the next layer; the initial input is denoted as... The output of the last layer is denoted as F. i The set of molecular structure information in the entire reaction equation is denoted as F = {F1, F2, ..., F...} n}, that is, to obtain molecular structure information F at the molecular level.

[0021] Furthermore, extracting reaction hierarchy sequence information includes:

[0022] Convert the chemical reaction formula R into a SMILES sequence, denoted as S. r ;

[0023] Using the word segmentation mechanism in RXNFP to segment S r After word segmentation, the input is fed into the RXNFP model to extract features, which are then obtained as follows: d b ×(N r +1), where N r d represents the number of word segmentation nodes in the entire chemical reaction formula R. b The output feature dimension is represented by cls segmentation, which is used to obtain the overall features of the reaction.

[0024] The reaction hierarchy sequence information is denoted as X = X r .

[0025] Furthermore, a width learning system is used to fuse the extracted molecular-level structural information and the extracted reaction-level sequence information.

[0026] Furthermore, a width learning system is used to fuse the extracted molecular-level structural information F and the extracted reaction-level sequence information X in the drug chemical reaction to obtain the predicted drug chemical reaction category, including:

[0027] Fusion T:

[0028] T = [F; X]

[0029] T is mapped to Z through a feature layer:

[0030]

[0031] in, and It is obtained through random generation, where φ is the activation function, n is the number of node windows in the feature layer, and each window has k nodes; all the obtained features are denoted as Z. n =[Z1,Z2,…,Z n These features, after passing through an enhancement layer, yield the enhanced feature H. m The i-th enhanced feature is calculated as follows:

[0032]

[0033] in, and It is obtained through generation, where ξ is the activation function;

[0034] The conversion rate Y to be predicted is calculated using the following formula:

[0035] Y = [Z1, Z2, ..., Z n |H1,H2,…,H m W m =[Z n |H m W m

[0036] Among them, W m Through pseudo-inverse operation Calculated.

[0037] To achieve the above objectives, the present invention further provides a system for implementing a method for predicting drug chemical reaction types based on multi-level information fusion, which includes a molecular hierarchical structure information extraction module, a reaction hierarchical sequence information extraction module, and an information fusion module;

[0038] in,

[0039] The molecular-level structure information extraction module is used to extract molecular-level structural information in pharmaceutical chemical reactions;

[0040] The reaction hierarchical sequence information extraction module is used to extract reaction hierarchical sequence information;

[0041] The information fusion module is used to fuse the extracted molecular-level structural information and the extracted reaction-level sequence information in the drug chemical reaction to obtain the predicted drug chemical reaction category.

[0042] To achieve the above objectives, the present invention further provides a computer, including a memory, a processor, and computer instructions stored in the memory and executable on the processor, wherein the processor executes the instructions to implement the steps of the above-described method for predicting drug chemical reaction types based on multi-level information fusion.

[0043] To achieve the above objectives, the present invention further provides a storage medium storing computer instructions that, when executed by a processor, implement the steps of the above-described method for predicting drug chemical reaction types based on multi-level information fusion.

[0044] Compared with existing technologies, the principles and advantages of this solution are as follows:

[0045] This approach fully leverages the unique information of reacting molecules in pharmaceutical chemical reactions—molecular structure information—and, combined with existing SMILES sequence encoding methods, extracts features at both the molecular and reaction levels, innovatively proposing a multi-level information fusion-based method for predicting pharmaceutical chemical reaction types. In the modeling process, firstly, the molecular structure encoding is embedded using a graph convolutional neural network to extract molecular-level structural information. Next, the pharmaceutical chemical reaction formula is converted into SMILES format and used as input to a pre-trained model. Then, BERT-RXNFP, specifically designed for chemical reactions, is used to segment the SMILES sequence of the reaction formula and extract reaction-level features. Finally, the molecular-level structural information and the reaction-level sequence information are fused using a fully connected layer MLP to predict the pharmaceutical chemical reaction type.

[0046] The deep model used in this scheme can mine multi-level information, enabling structural and sequence information to be integrated and complement each other, which helps to improve the performance of drug chemical reaction category prediction. It is easy to implement in engineering and has significant theoretical and application value. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the services required in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart illustrating the principle of the drug chemical reaction type prediction method based on multi-level information fusion of the present invention.

[0049] Figure 2 This is a connection block diagram of the drug chemical reaction type prediction system based on multi-level information fusion of the present invention. Detailed Implementation

[0050] The present invention will be further described below with reference to specific embodiments:

[0051] like Figure 1 As shown in this embodiment, the method for predicting drug chemical reaction types based on multi-level information fusion includes the following steps:

[0052] S1. Extract structural information at the molecular level in pharmaceutical chemical reactions to achieve the goal of learning the independent characteristics of specific molecules in the reaction.

[0053] Generally, for the classification of pharmaceutical chemical reactions, the overall reaction formula determines the reaction characteristics, and the reaction classification result can be predicted based on the characteristic model. However, reactant molecules also contain molecular-specific information, especially molecular structure, which is an important characteristic of drugs. Therefore, the molecular structure is encoded first. Specifically, if a certain reaction formula in a pharmaceutical chemical reaction is denoted as R, and the set of reactants and products in that reaction formula is denoted as B, where B = {B1, B2, ..., B...} n}, where n is the number of molecules of reactants and products in the reaction equation;

[0054] Next, extract the structural information of each molecule in B:

[0055] The adjacency matrix A of each molecule is obtained using the molecular data processing package Rdkit. i The embedding matrix H of each atomic node in the molecule i and degree matrix D i; where the adjacency matrix A of each molecule is i The embedding matrix H represents the connections between atoms, and is the embedding matrix of each atomic node in the molecule. i The degree matrix D represents the characteristics of each atom. i This represents the probability that each atom is connected to other atoms;

[0056] Based on the adjacency matrix A of each molecule i The embedding matrix H of each atomic node in the molecule i and degree matrix D i The structural information of each molecule is extracted using a convolutional neural network. The extraction formula is as follows:

[0057]

[0058] in I is the identity matrix, W is the network parameters, and σ(·) is the activation function. This is the output of this layer and also the input of the next layer; the initial input is denoted as... The output of the last layer is denoted as F. i The set of molecular structure information in the entire reaction equation is denoted as F = {F1, F2, ..., F...} n}, that is, to obtain molecular structure information F at the molecular level.

[0059] S2. Extract reaction hierarchy sequence information, specifically including:

[0060] Convert the chemical reaction formula R into a SMILES sequence, denoted as S. r ;

[0061] Using the word segmentation mechanism in RXNFP to segment S r After word segmentation, the input is fed into the RXNFP model to extract features, which are then obtained as follows: d b ×(N r +1), where N r d represents the number of word segmentation nodes in the entire chemical reaction formula R. b The output feature dimension is represented by cls segmentation, which is used to obtain the overall features of the reaction.

[0062] The reaction hierarchy sequence information is denoted as X = X r ;

[0063] This step involves fine-tuning the pre-trained model to embed the sequence relationships of reactants, products, and reaction conditions in drug chemical reactions into vectors, further preparing for subsequent prediction of drug chemical reaction categories.

[0064] S3. Using a width learning system, the extracted molecular-level structural information F and the extracted reaction-level sequence information X from the drug chemical reaction are fused to obtain the predicted drug chemical reaction category. The specific process includes:

[0065] Fusion T:

[0066] T = [F; X]

[0067] T is mapped to Z through a feature layer:

[0068]

[0069] in, and It is obtained through random generation, where φ is the activation function, n is the number of node windows in the feature layer, and each window has k nodes; all the obtained features are denoted as Z. n =[Z1,Z2,…,Z n These features, after passing through an enhancement layer, yield the enhanced feature H. m The i-th enhanced feature is calculated as follows:

[0070]

[0071] in, and It is obtained through generation, where ξ is the activation function;

[0072] The conversion rate Y to be predicted is calculated using the following formula:

[0073] Y = [Z1, Z2, ..., Z n |H1,H2,…,H m W m =[Z n |H m W m

[0074] Among them, W m Through pseudo-inverse operation The parameter W is calculated by using the labels of the training data. m This parameter can be used to calculate the labels of the test set, which is the predicted conversion rate that the task needs to achieve.

[0075] In this embodiment, the range of n is 15-20, and the range of m is 15-20.

[0076] Specifically, this embodiment also provides a system for implementing a method for predicting drug chemical reaction types based on multi-level information fusion, such as... Figure 2As shown, it includes a molecular hierarchical structure information extraction module, a reaction hierarchical sequence information extraction module, and an information fusion module;

[0077] The module includes a molecular-level structure information extraction module, which extracts molecular-level structure information in drug chemical reactions; a reaction-level sequence information extraction module, which extracts reaction-level sequence information; and an information fusion module, which fuses the extracted molecular-level structure information and the extracted reaction-level sequence information to obtain the predicted drug chemical reaction category.

[0078] Specifically, this embodiment also provides a computer, which includes a memory, a processor, and computer instructions stored in the memory and executable on the processor. When the processor executes the instructions, it implements the steps of the above-described method for predicting drug chemical reaction types based on multi-level information fusion.

[0079] Specifically, this embodiment also provides a storage medium storing computer instructions that, when executed by a processor, implement the steps of the above-described method for predicting drug chemical reaction types based on multi-level information fusion.

[0080] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, any changes made in accordance with the shape and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for predicting drug chemical reaction types based on multi-level information fusion, characterized in that, include: Extracting structural information at the molecular level from medicinal chemical reactions; Extract reaction hierarchy sequence information; The molecular-level structural information and the reaction-level sequence information extracted from the drug chemical reaction are fused together to obtain the predicted drug chemical reaction category. Extracting structural information at the molecular level in medicinal chemical reactions includes: Encoding the molecular-level structure in pharmaceutical chemical reactions: Let R be a reaction equation in a pharmaceutical chemistry reaction, and let B be the set of reactants and products in that reaction equation, where B = {B1, B2, ..., B}. n }, where n is the number of molecules of reactants and products in the reaction equation; Extract the structural information of each molecule in B; Extracting structural information for each molecule in B includes: The adjacency matrix A of each molecule is obtained using the molecular data processing package Rdkit. i The embedding matrix H of each atomic node in the molecule i and degree matrix D i ; where the adjacency matrix A of each molecule is i The embedding matrix H represents the connections between atoms, and is the embedding matrix of each atomic node in the molecule. i The degree matrix D represents the characteristics of each atom. i This represents the probability that each atom is connected to other atoms; Based on the adjacency matrix A of each molecule i The embedding matrix H of each atomic node in the molecule i and degree matrix D i Extract the structural information of each molecule; Based on the adjacency matrix A of each molecule i The embedding matrix H of each atomic node in the molecule i and degree matrix D i The structural information of each molecule is extracted using a convolutional neural network. The extraction formula is as follows: in I is the identity matrix, W is the network parameters, and σ(·) is the activation function. This is the output of this layer and also the input of the next layer; the initial input is denoted as... The output of the last layer is denoted as F. i The set of molecular structure information in the entire reaction equation is denoted as F = {F1, F2, ..., F...} n }, that is, to obtain molecular structure information F at the molecular level.

2. The method for predicting drug chemical reaction types based on multi-level information fusion according to claim 1, characterized in that, Extracting reaction hierarchy sequence information includes: Convert the chemical reaction formula R into a SMILES sequence, denoted as S. r ; Using the word segmentation mechanism in RXNFP to segment S r After word segmentation, the input is fed into the RXNFP model to extract features, which are then obtained as follows: Where N r d represents the number of word segmentation nodes in the entire chemical reaction formula R. b The output feature dimension is represented by cls segmentation, which is used to obtain the overall features of the reaction. The reaction hierarchy sequence information is denoted as X = X r .

3. The method for predicting drug chemical reaction types based on multi-level information fusion according to claim 2, characterized in that, A width learning system is used to fuse the molecular-level structural information and the extracted reaction-level sequence information from the drug chemical reactions.

4. The method for predicting drug chemical reaction types based on multi-level information fusion according to claim 3, characterized in that, A width learning system is used to fuse the extracted molecular-level structural information F and the extracted reaction-level sequence information X from the drug chemical reactions to obtain the predicted drug chemical reaction categories, including: Fusion T: T = [F; X] T is mapped to Z through a feature layer: in, and It is obtained through random generation, where φ is the activation function, n is the number of node windows in the feature layer, and each window has k nodes; all the obtained features are denoted as Z. n =[Z1,Z2,…,Z n These features, after passing through an enhancement layer, yield the enhanced feature H. m The i-th enhanced feature is calculated as follows: in, and ξ is the activation function obtained from the generation process; The conversion rate Y to be predicted is calculated using the following formula: Y=[Z1,Z2,…,Z n |H1,H2,…,H m ]W m =[Z n |H m ]W m Among them, W m Through pseudo-inverse operation Calculated.

5. A system for implementing the drug chemical reaction type prediction method based on multi-level information fusion as described in any one of claims 1-4, characterized in that, It includes a molecular hierarchical structure information extraction module, a reaction hierarchical sequence information extraction module, and an information fusion module; in, The molecular-level structure information extraction module is used to extract molecular-level structural information in pharmaceutical chemical reactions; The reaction hierarchical sequence information extraction module is used to extract reaction hierarchical sequence information; The information fusion module is used to fuse the extracted molecular-level structural information and the extracted reaction-level sequence information in the drug chemical reaction to obtain the predicted drug chemical reaction category.

6. A computer, comprising a memory, a processor, and computer instructions stored in the memory and executable on the processor, characterized in that, When the processor executes the instructions, it implements the steps of the method according to any one of claims 1-4.

7. A storage medium storing computer instructions, characterized in that, When executed by a processor, the computer instructions implement the steps of the method according to any one of claims 1-4.

Images