Drug metabolite prediction method and system based on conditional generative adversarial network

Abstract

This invention belongs to the field of drug metabolism prediction technology, specifically relating to a method and system for predicting drug metabolites based on conditional generative adversarial networks (GANs). The method includes: extracting initial features of drug molecules and masking them, inputting them into a shared encoder to obtain a first encoding result; incorporating atomic position information into the first encoding result, decoding it, and then predicting the structure of the drug molecule; inputting reactant molecules into the shared encoder, inputting the obtained reactant features into a shared decoder, and predicting chemical reaction products; inputting substrate molecules into the shared encoder for encoding; fusing the substrate encoding result with the corresponding enzyme's amino acid sequence and inputting it into the shared decoder to predict metabolites; constructing a generator using the shared encoder and shared decoder, building a multi-task discriminator, and performing adversarial training on the generator and discriminator to obtain a prediction model for predicting metabolites. This invention improves prediction accuracy by fusing information from drug structure and enzymes to construct a prediction model.

Description

Technical Field

[0001] This invention belongs to the field of drug metabolism prediction technology, specifically relating to a method and system for predicting drug metabolites based on conditional generative adversarial networks. Background Technology

[0002] Metabolic processes can partially alter the structure of drugs. Drug molecules are modified by chemical modifications catalyzed by various drug-metabolizing enzymes, and the products are called metabolites.

[0003] While metabolic processes are protective mechanisms for eliminating xenobiotics, they can, for drugs, lead to reduced efficacy and safety issues. Several drugs have been withdrawn from the market due to hepatotoxicity, primarily caused by the formation of active metabolites. Identifying potential metabolites is crucial for anticipating adverse effects due to off-target activity. These metabolites also play a significant role in improving pharmacokinetics (PK), pharmacodynamics (PD), and safety in drug discovery and development. Therefore, drug metabolism studies are an essential component of drug development, providing insights into the suitability of compounds as drugs or suggesting ways to improve the metabolic properties of lead compounds through chemical modifications. Thus, analyzing metabolic processes has significant guiding implications for the development of effective drugs. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for predicting drug metabolites based on conditional generative adversarial networks (GANs), thereby improving the aforementioned problems. To achieve this objective, the technical solution adopted by this invention is as follows:

[0005] In a first aspect, this application provides a method for predicting drug metabolites based on conditional generative adversarial networks, including:

[0006] Initial features are extracted based on the structure of the drug molecule and masked. The masked initial features are then input into a shared encoder to obtain the first encoding result.

[0007] Atomic position information is incorporated into the first encoding result and input into the first decoder. The structure of the drug molecule is predicted based on the obtained first decoding result.

[0008] Reactant molecules are input into a shared encoder, and noise is introduced into the encoding result to obtain reactant features;

[0009] The reactant characteristics of the chemical reaction are input into a shared decoder, and the information of the chemical reaction products is predicted based on the obtained second decoding result.

[0010] Substrate molecules are input into a shared encoder to obtain substrate coding results;

[0011] The substrate encoding result of the metabolic reaction and the corresponding enzyme amino acid sequence are fused and input into a shared decoder. The metabolites are predicted based on the obtained third decoding result.

[0012] A generator is constructed by a shared encoder and a shared decoder, a multi-task discriminator is built, and the generator and discriminator are trained adversarially to obtain a prediction model. The prediction model is then used to predict drug metabolites.

[0013] Secondly, the present invention also provides a drug metabolite prediction system based on conditional generative adversarial networks, comprising:

[0014] The first module is used to extract initial features based on the structure of drug molecules and perform masking operations. The masked initial features are then input into a shared encoder to obtain the first encoding result.

[0015] The second module is used to incorporate atomic position information into the first encoding result, input it into the first decoder, and predict the structure of the drug molecule based on the obtained first decoding result.

[0016] The third module is used to input reactant molecules into the shared encoder, introduce noise into the encoding result, and obtain reactant features;

[0017] The fourth module is used to input the reactant characteristics of the chemical reaction into the shared decoder and predict the information of the chemical reaction products based on the obtained second decoding result;

[0018] The fifth module is used to input substrate molecules into the shared encoder to obtain the substrate encoding result;

[0019] The sixth module is used to fuse the substrate encoding result of the metabolic reaction with the corresponding enzyme amino acid sequence and input it into the shared decoder to predict the metabolites based on the obtained third decoding result.

[0020] The seventh module is used to construct a generator consisting of a shared encoder and a shared decoder, build a multi-task discriminator, perform adversarial training on the generator and discriminator to obtain a prediction model, and predict drug metabolites based on the prediction model.

[0021] The beneficial effects of this invention are as follows:

[0022] This invention proposes a molecular metabolite prediction method based on conditional generative adversarial networks (GANs). During the pre-training process of chemical reactions, a shared encoder and decoder are used. This helps the model learn the structure of substrate molecules and general chemical reaction rules, reducing the generation of invalid structures during the fine-tuning stage. In the fine-tuning stage, a pre-trained protein language model is used to encode enzyme information, obtaining enzyme features. Then, a conditionally gated feature fusion module is used to fully interact with the information of substrate molecules and enzymes. The constructed multi-discriminator set helps the model learn the rules governing the generation of metabolites after enzymes act on substrate molecules.

[0023] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart of the drug metabolite prediction method based on conditional generative adversarial networks, as described in this application.

[0026] Figure 2 This is a flowchart illustrating the molecular pre-training process in an embodiment of this application.

[0027] Figure 3 This is a flowchart illustrating the pre-training process for chemical reactions in an embodiment of this application.

[0028] Figure 4 This is a flowchart illustrating the metabolic reaction fine-tuning process in an embodiment of this application.

[0029] Figure 5 This is a schematic diagram of the structure of a drug metabolite prediction device based on a conditional generative adversarial network according to an embodiment of this application;

[0030] Symbol explanation: 800 - Drug metabolite prediction device based on conditional generative adversarial network; 801 - Processor; 802 - Memory; 803 - Multimedia component; 804 - I / O interface; 805 - Communication component. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0032] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0033] Metabolism generally refers to the biotransformation of endogenous substances within cells, such as glucose, fatty acids, and amino acids. Metabolic processes involve enzyme-mediated chemical reactions that sustain life by providing energy to cells or eliminating potentially harmful compounds. Certain families of enzymes are responsible for eliminating xenobiotics, compounds not naturally present in the body, such as drugs, pesticides, and pollutants. When exogenous chemical drugs enter the body, they not only interfere with endogenous cellular metabolism but also undergo their own biotransformation; this process is called drug metabolism. Drug metabolism typically occurs in two phases in the liver. Phase I involves oxidation reactions, primarily mediated by the cytochrome P450 (CYP450) enzyme family, which usually involve introducing or exposing polar groups on drug molecules. These reactions prepare the molecule for the binding reactions in Phase II, mediated by transferases, which bind drug molecules to certain endogenous small molecules for two purposes: first, to inactivate potentially toxic compounds, and second, to increase their polarity to promote excretion from the body. Metabolic processes can partially alter the structure of drugs. Drug molecules are modified by chemical modifications catalyzed by various drug-metabolizing enzymes, and the products are called metabolites.

[0034] While metabolic processes are protective mechanisms for eliminating xenobiotics, they can, for drugs, lead to reduced efficacy and safety issues. Stage I reactions can form toxic metabolites, posing a risk of hepatotoxicity. In fact, several drugs have been withdrawn from the market due to hepatotoxicity, primarily due to the formation of active metabolites. Identifying potential metabolites is crucial for anticipating adverse effects that may result from off-target activity. These metabolites also play a significant role in improving pharmacokinetics (PK), pharmacodynamics (PD), and safety in drug discovery and development. Therefore, drug metabolism studies are an essential component of drug development, providing insights into the suitability of compounds as drugs or suggesting ways to improve the metabolic properties of lead compounds through chemical modifications. Thus, analyzing metabolic processes has significant guiding implications for the development of effective drugs.

[0035] Existing research has proposed some methods for predicting drug metabolites, such as those based on neural networks or graph sensing, but these techniques still have the following limitations:

[0036] (1) The generated model has invalid structure problems, which affect the chemical validity and usability of the prediction results. Metabolite prediction methods based on sequence generation often have problems such as syntax errors in the SMILES string or invalid chemical structures, such as incorrect valence states;

[0037] (2) Existing methods fail to fully consider the three-dimensional molecular structure and enzyme protein sequence information.

[0038] (3) Existing methods do not fully utilize the information exchange between enzymes and substrate molecules. Different enzymes have distinctly different active sites. Combined with preferences and reaction mechanisms, the dynamic interaction between enzyme active sites and substrate molecules helps models understand enzyme responses to substrates.

[0039] Example 1:

[0040] To address the problems in the prior art, this application provides a method for predicting drug metabolites based on conditional generative adversarial networks, including molecular pre-training, chemical reaction pre-training, and metabolic reaction fine-tuning, specifically including steps S100, S200, S300, S400, S500, S600 and S700.

[0041] S100. Extract initial features based on the structure of the drug molecule and perform a masking operation. Input the masked initial features into a shared encoder to obtain the first encoding result. The initial features include initial type features and initial paired features. The specific extraction method is as follows:

[0042] S110. Extract the initial type features of each atom in the drug molecule through embedding functions;

[0043] Based on SMILES of drug molecules, the RDKit tool is used to generate three-dimensional structural diagrams of the corresponding drug molecules, obtaining the three-dimensional structural information of the molecules, which includes two types of information: atom types and three-dimensional structure. Each atom... There is a type As atoms initial type features Its initial feature representation comes from an embedding matrix:

[0044]

[0045] in, Indicates the first i One atom, for The initial type characteristics, For an embedding function in lookup table form, the embedding matrix corresponding to the function is denoted as . , of which row vector That is, atomic type The corresponding embedding vector; each element of the embedding matrix is ​​randomly initialized as follows:

[0046]

[0047] in Represents the embedding matrix The The row vectors are sampled from a normal distribution with a mean of 0 and a standard deviation of 0.02. This represents a normal distribution with a mean of 0 and a standard deviation of 0.02.

[0048] S120. Based on the position information of each atom, the Gaussian embedding features of any atom pair are extracted using radial basis functions;

[0049] For modeling the three-dimensional structure of molecules, the standard Transformer architecture cannot model three-dimensional information due to the continuity of coordinates and the global rotation and translation invariance. Therefore, the Euclidean distance between atomic pairs is encoded as an additional edge feature to introduce geometric information. Specifically, radial basis functions are used, where atoms... and The three-dimensional coordinates are respectively and RBF can be represented as:

[0050]

[0051] in, Represents atoms and The Euclidean distance between them; Indicates atomic pairs Gaussian embedding features; express The center of a trainable Gaussian function, where K is a hyperparameter. There are K trainable parameters, which are in the form of Gaussian functions; This represents the scaling factor, which controls the bandwidth of the RBF and is usually a learnable parameter. This represents an exponential function.

[0052] S130. Use the first multilayer perceptron to map the Gaussian embedding features to obtain initial paired features;

[0053] To utilize paired Gaussian embedding features in a multi-head attention mechanism, a multilayer perceptron is used to map the Gaussian feature vectors, resulting in... Paired features under attention:

[0054] in, For this pair of atoms in Paired representations under each attention head; Indicates the first Paired features of the atomic pairs corresponding to each attention head; Represents a multilayer perceptron; Indicates atomic pairs Gaussian embedding features;

[0055] S140. Randomly select atoms of a preset proportion as the masked objects, and multiply the binary mask vector with the initial type features and initial pairwise features of the masked objects to obtain the initial features after masking.

[0056] Construct a matrix from the initial type characteristics of all atoms contained in the molecule. Construct a matrix from the initial pairwise features of all atomic pairs. To achieve the pre-training objective of structure reconstruction, this invention introduces a random masking strategy before inputting the data into the shared encoder. Specifically, 15% of the atoms are randomly selected from the atom set as the objects to be masked. Accordingly, a binary mask vector is defined. ,in This indicates the number of atoms in the current molecule. Represents atoms Masked This indicates that the atom Retaining this, the masked features are:

[0057]

[0058]

[0059] in, This indicates multiplication of the primary elements; , Represents atoms and The corresponding mask vector; Representation matrix The OK; Represents the type feature matrix after masking The OK; Representation matrix The Line number List; Represents the pairwise feature matrices after masking The Line number List; Used to control atoms and Whether the paired features are preserved, if any endpoint atom is masked (when or If the atom pair is zero, then the pairwise characteristic of the corresponding atom pair is set to zero; only if the atom pair is zero. and When all features are preserved, the paired features of the atom pairs are preserved. That is, when an atom is masked, its own feature vector and the paired features of all atom pairs associated with that atom are set to zero; only the paired features of the unmasked atoms and their atom pairs are passed as valid input to the shared encoder.

[0060] S150. Input the masked initial type feature matrix and the initial pairwise feature matrix into the shared encoder to obtain the first encoding result, as follows:

[0061] The initial features after masking are constructed into an initial type feature matrix and an initial pairwise feature matrix, which are then input into the shared encoder. Since the masked result obtained in step S140 in this embodiment is already in matrix form, there is no need to construct additional matrices; the masked feature matrix is ​​directly used. and The shared encoder serves as the initial input; it is used to extract molecular features. Specifically, the shared encoder consists of multiple attention blocks, and the feature matrix is ​​updated through these attention blocks.

[0062] Let the first In the layered Attention Block, the atom type feature matrix of the molecule is: Its pairwise characteristic matrix of atoms is As shown below:

[0063]

[0064]

[0065] in, Indicates the first Layer type feature matrix, Indicates the first Atoms in layer Attention Block Type characteristics, Indicates the number of atoms in a molecule. This represents the transpose of a matrix. Indicates the first Pairwise feature matrices of the layer Indicates the first Atoms of the layer and Paired features between;

[0066] That is, the feature matrix is ​​passed as input to the first Attention Block, and the output of each Attention Block is used as the input to the next Attention Block, until the last Attention Block.

[0067] The Attention Block update process is as follows:

[0068] S151. Construct query, key, and value vectors based on the type feature matrix, calculate the attention between atoms based on the query vector and key vector, and correct it using pairwise feature matrices to obtain the corrected attention;

[0069] For the layer attention head First, the molecular type feature matrix is ​​used. Construct a query ,key Sum value vector The calculation method is as follows:

[0070]

[0071] in, For the first Learnable parameters for each attention head. Then, the pairwise feature matrices of the atomic pairs are used. Each pair of features in , take it as the first Atoms in the attention head To the atom The attention bias term is used to adjust the attention score, and the specific calculation is as follows:

[0072]

[0073] in, Indicates the first Layer, First In the attention head, atoms For atoms Attention weights; Indicates the first Layer, First The query vector in each attention head corresponds to an atom. The query representation; Indicates the first Layer, First The key vector in each attention head corresponds to an atom. The key represents; This represents the dimension of the key vector in each attention head; Indicates the first Layer, First Atom pairs in the attention head The paired features are used as bias terms to enhance the model's perception of paired features; express function.

[0074] To enhance the adaptability of bias terms to inter-layer contexts, an inter-layer update mechanism for bias terms is further introduced, the update method of which is as follows:

[0075]

[0076] in, Indicates the first Atoms in the layer To the atom In the Bias features under attention heads; Indicates the first Atoms in the layer To the atom In the Bias features under attention heads; Indicates the first Layer, First The query vector in each attention head corresponds to an atom. The query representation; Indicates the first Layer, First The key vector in each attention head corresponds to an atom. The key is represented by .

[0077] S152. Calculate the output vector for each attention head based on the corrected attention and value vectors;

[0078] Ultimately, atoms In the The output corresponding to each head is represented as follows:

[0079]

[0080] in, Represents atoms In the Layer, First The output vector under each attention head; Indicates the first Layer, First The value vector in each attention head corresponds to an atom. The value represents; This indicates the number of atoms contained in the current molecule; Indicates the first Layer, First In the attention head, atoms For atoms Attention weights;

[0081] S153. Concatenate and map the outputs of each attention head in each layer to obtain the updated atomic features of each layer;

[0082] No. Atoms of the layer The update indicates the following:

[0083]

[0084] in, Indicates the first Layer Atom Updated atomic features; express The linear transformation weight matrix of the layer; Indicates the number of attention heads; Represents atoms In the Layer, First H The output vector under each attention head; H For the total number of attention heads;

[0085] S154. The updated atomic features are processed by residual connection and normalization to obtain the first encoding result.

[0086] To enhance the model's expressive power, residual connections and layer normalization are introduced. The final atomic representation update process is as follows:

[0087]

[0088] in, This represents atoms after normalization and residual connection. The intermediate features are represented; Indicates the first Layer Atom Updated atomic features; express Atoms in the layer Type characteristics; This represents the layer normalization operation. A feedforward neural network is used for... Perform a nonlinear transformation, and then perform residual connection and normalization again:

[0089]

[0090] in, Represents atoms In the Feature representation in layers; and These are the learnable parameters in the feedforward network; Indicates the activation function; This represents atoms after normalization and residual connection. The intermediate features are represented.

[0091] Type feature matrix of the molecule after masking Pairwise characteristic matrix of atoms via shared encoder In The output features after the Attention Block layer are represented as follows:

[0092]

[0093] in, This indicates the atom type characteristics of the molecules output by the shared encoder; This represents the pairwise characteristics of atomic pairs in the molecule output by the shared encoder. and This is the first encoding result; For shared encoders; This represents the type feature matrix of the molecule after masking. This represents the pairwise characteristic matrix of the atomic pairs after masking.

[0094] S200: Integrate the atomic position information into the first encoding result, input it into the first decoder, and predict the structure of the drug molecule based on the obtained first decoding result;

[0095] S210. Obtain the preset dimension information of the atomic position;

[0096] The dimension of the atomic position is a hyperparameter, which can take values ​​such as 512 or 1024;

[0097] S220. Based on the position of the atom in the molecule, perform sine or cosine mapping to obtain the encoded value of the atom in each dimension;

[0098]

[0099] in, Indicates the first Each position (i.e., an atom) (position) in the The encoded value of the dimension; The dimension representing the positional encoding is a hyperparameter. Natural number index, used to represent The case where it is an even or odd number;

[0100] S230. Combine the encoded values ​​of the atom in all dimensions into a positional encoded vector. ;

[0101] S240. Add the position encoding vector to the atomic type feature in the first encoding result to obtain the initial input of the atomic type feature of the first decoder;

[0102]

[0103] in, Represents atoms The initial input of the atomic type features in the 0th layer of the first decoder; Represents atoms Atom type features obtained by encoding with a shared encoder; For atoms Location encoding vector;

[0104] S250, first decoder feature decoding;

[0105] The first decoder uses the same Attention Block as the shared encoder. Let , and Together they serve as the input to the first decoder; It is composed of all the atoms in the molecule. The initial input matrix is ​​formed; during the decoding process, after... After layer decoding, the atomic type embedding feature matrix of the molecule is obtained as follows: The pairwise characteristic matrix of the atoms in the molecule is :

[0106]

[0107] in, To decode and obtain the atomic type embedding feature matrix of the molecule, To decode and obtain the paired feature matrix of the molecule, Indicates the first decoder; Represents the initial input matrix; This represents the pairwise characteristics of atomic pairs in the molecule output by the shared encoder.

[0108] S260, prediction of atom type and paired features;

[0109] Finally, the atom types of the decoded molecules are embedded into the feature matrix. The pairwise characteristic matrix of atomic pairs Input to In this process, the prediction of pairwise characteristics of atom types and atom pairs is performed:

[0110]

[0111]

[0112] in, Indicates the predicted atom type, This indicates the pairwise characteristics of the predicted atomic pairs; To embed the feature matrix for the atom types of the molecule obtained from decoding; To decode and obtain the paired feature matrix of atomic pairs in the molecule; For predicting atom types, a multilayer perceptron, A multilayer perceptron for predicting the paired features of atomic pairs.

[0113] The total loss predicted in this section is the loss between the predicted result and the pairwise features of the true atom type and atom pairs, which are calculated using the cross-entropy loss function and the mean squared error loss function, respectively:

[0114]

[0115]

[0116]

[0117] in, This represents the loss function for molecular pre-training; and These represent the loss functions for predicting atom type and predicting the paired features of atom pairs, respectively. and These represent the atoms of the molecule. Atom types and atom pairs The true values ​​of paired features; and Each represents an atom Atom types and atom pairs The predicted values ​​of the paired features.

[0118] The above describes the molecular pre-training process, such as... Figure 2 As shown, the process involves masking the input molecule, encoding it with a shared encoder, adding positional encoding to the encoding result, decoding it with a decoder, and reconstructing the input molecule. This part is used to enhance the shared encoder's ability to extract molecular features, providing a unified characterization basis for subsequent chemical reaction pre-training and metabolic reaction fine-tuning.

[0119] S300. Input the reactant molecules into the shared encoder, introduce noise into the encoding result, and obtain the reactant characteristics;

[0120] In a chemical reaction, let the initial type characteristics of the reactant molecules be... The initial pairing characteristics of the corresponding atomic pairs are: The encoding process is completely consistent with the molecular pre-training stage in steps S100-S200, both using a shared encoder for structural feature extraction; represented as:

[0121]

[0122] in, This indicates the type characteristics of reactant molecules after encoding in a chemical reaction; The paired characteristics of atoms encoded by reactant molecules in a chemical reaction; For shared encoders; Characteristics of the initial type of reactant molecules; This represents the initial pairing characteristics of the atoms in the reactant molecules;

[0123] To improve the robustness of the generated molecules, standard Gaussian noise is introduced; the reactant molecules can be represented as:

[0124]

[0125] in, This indicates the type characteristics of reactant molecules after the introduction of noise; Noise from reactant molecules; This indicates the type characteristics of reactant molecules after encoding in a chemical reaction;

[0126] S400: Input the reactant characteristics of the chemical reaction into the shared decoder, and predict the information of the chemical reaction products based on the obtained second decoding result;

[0127] The shared decoder architecture follows that of the first decoder, but its parameters are different. Let... , and Together they serve as input to the shared decoder; represented as:

[0128]

[0129] in, This indicates the atomic type characteristics of the product molecule obtained by the decoder. This represents the paired characteristics of the atomic pairs in the product molecules obtained from decoding; Indicates a shared decoder; This indicates the type characteristics of reactant molecules after the introduction of noise; This refers to the paired characteristics of atomic pairs encoded by reactant molecules in a chemical reaction.

[0130] The decoded result , Enter them separately , In this process, the pairwise characteristics of all atomic types and atomic pairs in the product molecule are predicted:

[0131]

[0132]

[0133] in, This represents all the atom types in the product molecules of the predicted chemical reaction. This indicates the paired characteristics of all atomic pairs in the predicted product molecule; It is a multilayer sensor used to predict the atomic types of product molecules. It is a multilayer perceptron used to predict the paired characteristics of atomic pairs in product molecules.

[0134] When constructing the training samples for this model, data filtering is required first to select data where one reactant corresponds to only one product. These data are then used to construct the training samples, ensuring a one-to-one correspondence between reactants and products.

[0135] The loss predicted in this section is the loss between the predicted result and the pairwise features of the true atom type and atom pairs, using the cross-entropy loss function and the mean squared error loss function:

[0136]

[0137]

[0138]

[0139] in, The loss function representing the pre-training of chemical reactions; and These represent the loss functions for predicting the atomic types and pairwise characteristics of product molecules in a chemical reaction, respectively. The first molecule of the product The actual type label of each atom; Indicates the product molecule The true value of the paired characteristics of atomic pairs; The first molecule of the product Predicted type label for each atom; Indicates the product molecule Predicted values ​​of the pairwise characteristics of atomic pairs.

[0140] The above describes the pre-training process for chemical reactions, such as... Figure 3 As shown, the process involves encoding the input reactant molecules using a shared encoder, adding noise to the encoding result, decoding it using a shared decoder, and outputting the product molecules. The main purpose of chemical reaction pre-training is to learn the mapping from reactants to products. During this process, the shared encoder and shared decoder are trained simultaneously for fine-tuning of metabolic reactions.

[0141] S500: Input the substrate molecule into the shared encoder to obtain the substrate coding result;

[0142] Let the initial atomic type characteristics of the substrate molecule be: The pairing characteristics of corresponding atomic pairs are The encoding process is completely consistent with the pre-training stage of the chemical reaction, both using a shared encoder for structural feature extraction. This is represented as:

[0143]

[0144] in, This indicates the atomic type characteristics encoded in the substrate molecule; The paired characteristics of the encoded atomic pairs of the substrate molecule; Indicates the initial atomic type characteristics of the substrate molecule. This indicates the initial pairing characteristics of corresponding atomic pairs in the substrate molecule; For shared encoders;

[0145] S600: The substrate encoding result of the metabolic reaction and the corresponding enzyme amino acid sequence are fused and input into the shared decoder. The metabolites are predicted based on the obtained third decoding result.

[0146] This step employs a conditional gating mechanism to effectively fuse substrate molecule features with the amino acid sequence features of enzymes involved in metabolic reactions, thereby enhancing the model's ability to model metabolic processes.

[0147] The amino acid sequence characteristics of the enzyme were obtained using a pre-trained protein language model, and represented as follows: .

[0148] S610. The second multilayer perceptron is used to map the substrate molecule features, and the third multilayer perceptron is used to map the amino acid sequence to obtain the substrate molecule feature matrix and the amino acid sequence feature matrix.

[0149] Since substrate molecule features and amino acid sequence features have different semantic information, the two feature matrices are first mapped through two separate multilayer perceptrons to represent them in their respective feature spaces.

[0150]

[0151]

[0152] in, and These represent the use of multilayer perceptrons. and The mapped substrate molecule feature matrix and amino acid sequence feature matrix; It is a multilayer sensor used to map the features of substrate molecules; It is a multilayer perceptron used to map amino acid sequence features; This indicates the atomic type characteristics encoded in the substrate molecule; This indicates the initial amino acid sequence characteristics of the enzyme.

[0153] S620. The substrate molecule feature matrix and amino acid sequence feature matrix are spliced ​​together, and the splicing result is input into the activation function to generate the gating weight matrix.

[0154]

[0155] in, This represents the matrix obtained by splicing the substrate molecule feature matrix and the amino acid sequence feature matrix. By splicing, the expressions of these two features are combined, providing input for subsequent gating mechanisms. This indicates a splicing operation. and This represents the mapped substrate molecule feature matrix and amino acid sequence feature matrix;

[0156] A gating weight matrix is ​​generated using a gating mechanism. The fusion strength of the features is adjusted. Specifically, a gated weight matrix is ​​first generated using MLP and ReLU activation functions:

[0157]

[0158] in, This is the gated weight matrix; This represents the matrix obtained by splicing the substrate molecule feature matrix and the amino acid sequence feature matrix.

[0159] S630. Perform Bernoulli sampling on the gate weight matrix to generate binarized gate weights;

[0160]

[0161] in, For binarized gating weights; Sampling for Bernoulli; This is the gated weight matrix;

[0162] S640. Based on the binarized gating weights, the amino acid sequence feature matrix is ​​screened, and the screened amino acid sequence feature matrix is ​​fused with the substrate molecule features to obtain the fused features.

[0163]

[0164] in, Indicates the characteristics of the screened amino acid sequences; This indicates element-major multiplication; A is the amino acid sequence feature matrix; For binarized gating weights;

[0165] Fusion characteristics:

[0166]

[0167] in, D represents the fusion feature; D represents the substrate molecule feature matrix; This indicates the amino acid sequence characteristics after screening.

[0168] S650: Add Gaussian noise to the fused features and input it to the shared decoder;

[0169] To enhance the diversity of molecules generated by the model, standard Gaussian noise is introduced; the fusion features after adding noise can be represented as:

[0170]

[0171] in, This represents the fusion characteristics after adding noise. Indicates noise; Indicates fusion characteristics;

[0172] Substrate molecule fusion features with added noise obtained by the conditionally gated feature fusion module Paired features of atomic pairs output by the shared encoder As input to the shared decoder. Represented as:

[0173]

[0174] in, This indicates the atomic type characteristics of the metabolite molecules obtained from decoding; This represents the paired characteristics of atomic pairs in the decoded metabolite molecules; Indicates a shared decoder; This refers to the fusion characteristics of substrate molecules after noise has been added.

[0175] Finally, the decoded metabolite molecules were analyzed separately. Input to and In this process, the prediction of atomic types and pairwise characteristics of metabolites is performed:

[0176]

[0177]

[0178] in, This indicates the types of atoms in the metabolite molecules in the predicted metabolic reaction. This indicates the predicted pairwise characteristics of atomic pairs in the metabolite molecule; Multilayer sensing for predicting the types of atoms in metabolite molecules; A multilayer sensor for predicting the pairwise characteristics of atomic pairs in metabolite molecules. This indicates the atomic type characteristics of the metabolite molecules obtained from decoding; This represents the paired characteristics of atomic pairs in the decoded metabolite molecules;

[0179] The loss is the difference between the predicted result and the pairwise features of the true atom type and atom pairs, and is calculated using the cross-entropy loss function and the mean squared error loss function, respectively:

[0180]

[0181]

[0182]

[0183] in, This indicates the loss during the fine-tuning phase of the metabolic response; This represents the loss predicted by the atomic type. This represents the loss for predicting the paired features of atomic pairs; The first molecule representing a metabolite The true type of each atom; The first molecule representing a metabolite The predicted type of each atom; Representing the molecules of metabolites True pairing characteristics of atomic pairs; Representing the molecules of metabolites Predicted pairwise features of atomic pairs.

[0184] The S700 consists of a generator composed of a shared encoder and a shared decoder, which is used to construct a multi-task discriminator. The generator and discriminator are then subjected to adversarial training to obtain a prediction model, which is used to predict drug metabolites.

[0185] The multi-task discriminator mainly consists of three triplet consistency discriminators (denoted as...). , , ) and an adversarial discriminator (denoted as ) This is used to improve the realism and scientific rigor of the model, while ensuring that the model can learn more accurate reaction patterns between substrate, enzyme, and product.

[0186] The discriminant model is constructed as follows:

[0187] Constructing the first discriminator Second discriminator and the third discriminator And construct real triplet based on real drug metabolism data. The tripartite consists of three elements: a substrate molecule feature matrix, an amino acid sequence feature matrix, and a metabolite molecule feature matrix; that is... ;

[0188] in, A matrix representing the molecular characteristics of metabolites in a real metabolic reaction; and These represent the initial substrate molecule feature matrix and the enzyme amino acid sequence feature matrix, respectively.

[0189] By replacing individual elements in the real triplet, three types of negative samples are constructed, with each type of negative sample replacing the same element. The three types of negative samples correspond to the replacement of the substrate molecule feature matrix, the amino acid sequence feature matrix, and the metabolite molecule feature matrix, respectively.

[0190] By randomly replacing the substrate molecules, we obtain Construct triples As Input to the discriminator.

[0191] By randomly replacing the amino acid sequence of the enzyme, we obtain Construct triples As The input to the discriminator;

[0192] By randomly replacing metabolites, we obtain Construct triples As Input to the discriminator.

[0193] The three negative samples are combined with the real triples respectively, and the resulting three combined samples are used as the inputs of the first discriminator, the second discriminator, and the third discriminator respectively.

[0194] The goal of the triplet consistency discriminator is to maximize the discrimination probability of true samples while minimizing the discrimination probability of negative samples.

[0195] Construct the collaborative loss function for the first discriminator, the second discriminator, and the third discriminator;

[0196] The following collaborative loss function is adopted:

[0197]

[0198] in, For collaborative losses; Indicates the first N A triplet consistency discriminant; and These represent the labels of the positive and negative triplets in the triplet consistency discriminator, respectively. A triple representing a real metabolic response; This represents a true triplet of a metabolic response, where random substitutions are performed with the first... N The negative sample triplet is obtained by taking the elements corresponding to the triplet consistency discriminator.

[0199] Construct the adversarial discriminant and the corresponding adversarial loss function;

[0200] Adversarial discriminator This discriminator is used to distinguish between real product molecules and model-generated product molecules. By using this discriminator, the realism of generated metabolites can be improved. The adversarial loss function is:

[0201]

[0202] in, This indicates the loss against the discriminator; Characteristic representation of actual metabolite molecules; This indicates the molecular characteristics of the metabolites generated by the generator. and These represent the labels of positive and negative samples in the adversarial discriminator, respectively.

[0203] Discriminator training is performed based on the collaborative loss function and the adversarial loss function;

[0204] The total loss of the multi-task discriminator is:

[0205]

[0206] in, The total loss of the multi-task discriminator. To mitigate losses, This indicates the loss against the discriminator;

[0207] The above describes the process of fine-tuning metabolic responses, such as... Figure 4 As shown, the process involves inputting a substrate molecule, encoding it with a shared encoder, fusing the encoded result with the input amino acid sequence using conditionally gated features, decoding it with a shared decoder, and predicting the product molecule. The predicted product molecule and the real product molecule are input into an adversarial discriminator, and the real triplet consisting of the substrate molecule, amino acid sequence, and real product molecule, as well as the corresponding negative sample triplet, are input into a triplet consistency discriminator.

[0208] Through adversarial training of the generator and discriminator, a prediction model is obtained, and drug metabolites are predicted based on the prediction model.

[0209] This invention employs an encoder-decoder structure for pre-training tasks, performing molecular reconstruction and chemical reaction reconstruction separately. In molecular reconstruction, a shared encoder extracts the three-dimensional structural features of the molecule, which are then reconstructed using an independent decoder. In chemical reaction reconstruction, a shared encoder and a shared decoder are used to generate and reconstruct the structures before and after the reaction. This pre-training strategy helps the model learn the structural information of molecules and the general rules of chemical reactions, laying the foundation for improving the ability to generate effective structures in subsequent fine-tuning stages.

[0210] In the fine-tuning stage, a shared encoder is first used to encode the substrate molecule to obtain its structural features. Simultaneously, a pre-trained protein language model is used to encode the enzyme's amino acid sequence information to obtain the enzyme's amino acid sequence feature vector. Subsequently, conditional gating feature fusion is introduced. Using the substrate molecule as a condition, a gating mechanism is used to achieve the fusion of amino acid sequence and substrate features, thereby realizing full information interaction and efficient fusion between the enzyme and the substrate.

[0211] The final discriminator module consists of a triplet consistency discriminator and an adversarial discriminator. The triplet consistency discriminator uses real "substrate-enzyme-product" combinations as positive samples and constructs negative samples by randomly replacing any element in the triplet, thus driving the model to learn the molecular-level rules of enzyme action on substrates to produce specific metabolites. The adversarial discriminator is used to distinguish generated molecules from real products, improving the realism and discriminative ability of the model's generated results.

[0212] Example 2:

[0213] This embodiment provides a drug metabolite prediction system based on conditional generative adversarial networks, including:

[0214] The first module is used to extract initial features based on the structure of drug molecules and perform masking operations. The masked initial features are then input into a shared encoder to obtain the first encoding result.

[0215] The second module is used to incorporate atomic position information into the first encoding result, input it into the first decoder, and predict the structure of the drug molecule based on the obtained first decoding result.

[0216] The third module is used to input reactant molecules into the shared encoder, introduce noise into the encoding result, and obtain reactant features;

[0217] The fourth module is used to input the reactant characteristics of the chemical reaction into the shared decoder and predict the information of the chemical reaction products based on the obtained second decoding result;

[0218] The fifth module is used to input substrate molecules into the shared encoder to obtain the substrate encoding result;

[0219] The sixth module is used to fuse the substrate encoding result of the metabolic reaction with the corresponding enzyme amino acid sequence and input it into the shared decoder to predict the metabolites based on the obtained third decoding result.

[0220] The seventh module is used to construct a generator consisting of a shared encoder and a shared decoder, build a multi-task discriminator, perform adversarial training on the generator and discriminator to obtain a prediction model, and predict drug metabolites based on the prediction model.

[0221] As an optional implementation, the initial features include initial type features and initial paired features; the first module includes:

[0222] The first unit is used to extract the initial type features of each atom in the drug molecule through an embedding function;

[0223] The second unit is used to extract Gaussian embedding features of arbitrary atom pairs based on the position information of each atom using radial basis functions.

[0224] The third unit is used to map the Gaussian embedding features using the first multilayer perceptron to obtain initial paired features;

[0225] The fourth unit is used to randomly select a preset proportion of atoms as the masked objects. The binary mask vector is multiplied by the initial type features and initial pairwise features of the masked objects to obtain the initial features after masking.

[0226] As an optional implementation, the first module further includes:

[0227] The fifth unit is used to construct an initial type feature matrix and an initial pairwise feature matrix from the masked initial features and input them into the shared encoder, which includes multiple attention blocks;

[0228] The sixth unit is used to construct query, key, and value vectors based on the type feature matrix, calculate the attention between atoms based on the query vector and key vector, and correct it using pairwise feature matrices to obtain the corrected attention;

[0229] The seventh unit is used to calculate the output vector under each attention head based on the corrected attention and value vector;

[0230] The eighth unit is used to concatenate and map the outputs of each attention head in each layer to obtain the updated atomic features of each layer;

[0231] The ninth unit is used to process the updated atomic features through residual connection and normalization to obtain the first encoding result, which includes atomic type features and paired features of atomic pairs.

[0232] As an optional implementation, the second module includes:

[0233] The tenth unit is used to obtain the preset dimensional information of the atomic positions;

[0234] The eleventh unit is used to perform sine or cosine mapping based on the position of the atom in the molecule to obtain the encoded value of the atom in each dimension;

[0235] The twelfth unit is used to combine the encoded values ​​of atoms in all dimensions into a positional encoded vector;

[0236] The thirteenth unit is used to add the position encoding vector to the atomic type features in the first encoding result to obtain the initial input of the atomic type features of the first decoder.

[0237] Example 3:

[0238] Corresponding to the above method embodiments, this embodiment also provides a drug metabolite prediction device based on conditional generative adversarial networks. The drug metabolite prediction device based on conditional generative adversarial networks described below can be referred to in correspondence with the drug metabolite prediction method based on conditional generative adversarial networks described above.

[0239] Figure 5 This is a block diagram illustrating a drug metabolite prediction device 800 based on a conditional generative adversarial network, according to an exemplary embodiment. Figure 5As shown, the conditional generative adversarial network (GAP)-based drug metabolite prediction device 800 includes a processor 801 and a memory 802. The GAP-based drug metabolite prediction device 800 may also include one or more of the following: a multimedia component 803, an input / output (I / O) interface 804, and a communication component 805. The processor 801 controls the overall operation of the GAP-based drug metabolite prediction device 800 to complete all or part of the steps in the aforementioned GAP-based drug metabolite prediction method. The memory 802 stores various types of data to support the operation of the GAP-based drug metabolite prediction device 800. This data may include, for example, commands for any application or method operating on the GAP-based drug metabolite prediction device 800, and application-related data such as contact data, sent and received messages, images, audio, video, etc. The memory 802 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0240] Multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals.

[0241] The received audio signal can be further stored in memory 802 or transmitted via communication component 805. The audio component also includes at least one speaker for outputting the audio signal. I / O interface 804 provides an interface between processor 801 and other interface modules, such as a keyboard, mouse, buttons, etc. These buttons can be virtual or physical. Communication component 805 is used for wired or wireless communication between the conditional generative adversarial network-based drug metabolite prediction device 800 and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof, is used; therefore, the corresponding communication component 805 may include a Wi-Fi module, a Bluetooth module, or an NFC module.

[0242] Example 4:

[0243] Corresponding to the above embodiment of the drug metabolite prediction method based on conditional generative adversarial networks, this embodiment also provides a readable storage medium. The readable storage medium described below can be referred to in conjunction with the drug metabolite prediction method based on conditional generative adversarial networks described above.

[0244] A readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described embodiments of the drug metabolite prediction method based on conditional generative adversarial networks.

[0245] Specifically, the readable storage medium can be a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, or any other readable storage medium capable of storing program code.

[0246] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0247] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting drug metabolites based on conditional generative adversarial networks, characterized in that, include: Initial features are extracted based on the structure of the drug molecule and masked. The masked initial features are then input into a shared encoder to obtain the first encoding result. Atomic position information is incorporated into the first encoding result and input into the first decoder. The structure of the drug molecule is predicted based on the obtained first decoding result. Reactant molecules are input into a shared encoder, and noise is introduced into the encoding result to obtain reactant features; The reactant characteristics of the chemical reaction are input into a shared decoder, and the information of the chemical reaction products is predicted based on the obtained second decoding result. Substrate molecules are input into a shared encoder to obtain substrate coding results; The substrate encoding result of the metabolic reaction and the corresponding enzyme amino acid sequence are fused and input into a shared decoder. The metabolites are predicted based on the obtained third decoding result. A generator is constructed by a shared encoder and a shared decoder, a multi-task discriminator is built, and the generator and discriminator are trained adversarially to obtain a prediction model. The prediction model is then used to predict drug metabolites.

2. The drug metabolite prediction method based on conditional generative adversarial networks according to claim 1, characterized in that, The initial features include initial type features and initial paired features; Initial features are extracted based on the structure of the drug molecule and then masked, including: The initial type features of each atom in the drug molecule are extracted using an embedding function; Based on the position information of each atom, the Gaussian embedding features of any atom pair are extracted using radial basis functions. The Gaussian embedding features are mapped using a first multilayer perceptron to obtain initial paired features; A predetermined proportion of atoms are randomly selected as the objects to be masked. The initial features after masking are obtained by multiplying the binary mask vector with the initial type features and initial pairwise features of the objects to be masked.

3. The drug metabolite prediction method based on conditional generative adversarial networks according to claim 2, characterized in that, The masked initial features are input into the shared encoder to obtain the first encoding result, including: The masked initial features are constructed into an initial type feature matrix and an initial pairwise feature matrix, which are then input into a shared encoder, which includes multiple attention blocks. Query, key, and value vectors are constructed based on the type feature matrix. Attention between atoms is calculated based on the query vector and key vector, and then corrected using pairwise feature matrices to obtain the corrected attention. Calculate the output vector for each attention head based on the corrected attention and value vectors; The outputs of each attention head in each layer are concatenated and mapped to obtain the updated atomic features of each layer. The updated atomic features are processed by residual connection and normalization to obtain the first encoding result, which includes atomic type features and paired features of atomic pairs.

4. The drug metabolite prediction method based on conditional generative adversarial networks according to claim 3, characterized in that, Atom position information is incorporated into the first encoding result, including: Obtain the preset dimensional information of the atomic positions; By performing sine or cosine mapping based on the position of the atom in the molecule, the encoded value of the atom in each dimension can be obtained; Combine the encoded values ​​of the atom in all dimensions into a positional encoded vector; The positional encoding vector is added to the atomic type features in the first encoding result to obtain the initial input of the atomic type features of the first decoder.

5. The drug metabolite prediction method based on conditional generative adversarial networks according to claim 1, characterized in that, The substrate encoding results include substrate molecular features and substrate paired features; The substrate encoding result of the metabolic reaction and the corresponding enzyme amino acid sequence are fused and then input into a shared decoder, including: A second multilayer perceptron is used to map the substrate molecule features, and a third multilayer perceptron is used to map the amino acid sequence, to obtain the substrate molecule feature matrix and the amino acid sequence feature matrix; The substrate molecule feature matrix and the amino acid sequence feature matrix are spliced ​​together, and the splicing result is input into the activation function to generate a gating weight matrix; Bernoulli sampling is performed on the gate weight matrix to generate binarized gate weights; Based on the binarized gating weights, the amino acid sequence feature matrix is ​​screened, and the screened amino acid sequence feature matrix is ​​fused with the substrate molecule feature matrix to obtain the fused feature. Gaussian noise is added to the fused features and then input to the shared decoder.

6. The drug metabolite prediction method based on conditional generative adversarial networks according to claim 1, characterized in that, Construct a multi-task discriminator and conduct adversarial training on the generator and discriminator, including: A first discriminator, a second discriminator, and a third discriminator are constructed, and a real triplet is constructed based on real drug metabolism data; the triplet includes three elements: substrate molecule feature matrix, amino acid sequence feature matrix, and metabolite molecule feature matrix. By replacing individual elements in the real triplet, three types of negative samples are constructed, with each type of negative sample replacing the same element. The three types of negative samples correspond to the replacement of the substrate molecule feature matrix, the amino acid sequence feature matrix, and the metabolite molecule feature matrix, respectively. The three negative samples are combined with the real triples respectively, and the resulting three combined samples are used as the inputs of the first discriminator, the second discriminator, and the third discriminator respectively. Construct the collaborative loss function for the first discriminator, the second discriminator, and the third discriminator; Construct the adversarial discriminant and the corresponding adversarial loss function; Discriminator training is performed based on the collaborative loss function and the adversarial loss function.

7. A drug metabolite prediction system based on conditional generative adversarial networks, characterized in that, include: The first module is used to extract initial features based on the structure of drug molecules and perform masking operations. The masked initial features are then input into a shared encoder to obtain the first encoding result. The second module is used to incorporate atomic position information into the first encoding result, input it into the first decoder, and predict the structure of the drug molecule based on the obtained first decoding result. The third module is used to input reactant molecules into the shared encoder, introduce noise into the encoding result, and obtain reactant features; The fourth module is used to input the reactant characteristics of the chemical reaction into the shared decoder and predict the information of the chemical reaction products based on the obtained second decoding result; The fifth module is used to input substrate molecules into the shared encoder to obtain the substrate encoding result; The sixth module is used to fuse the substrate encoding result of the metabolic reaction with the corresponding enzyme amino acid sequence and input it into the shared decoder to predict the metabolites based on the obtained third decoding result. The seventh module is used to construct a generator consisting of a shared encoder and a shared decoder, build a multi-task discriminator, perform adversarial training on the generator and discriminator to obtain a prediction model, and predict drug metabolites based on the prediction model.

8. The drug metabolite prediction system based on conditional generative adversarial networks according to claim 7, characterized in that, The initial features include initial type features and initial paired features; the first module includes: The first unit is used to extract the initial type features of each atom in the drug molecule through an embedding function; The second unit is used to extract Gaussian embedding features of arbitrary atom pairs based on the position information of each atom using radial basis functions. The third unit is used to map the Gaussian embedding features using the first multilayer perceptron to obtain initial paired features; The fourth unit is used to randomly select a preset proportion of atoms as the masked objects. The binary mask vector is multiplied by the initial type features and initial pairwise features of the masked objects to obtain the initial features after masking.

9. The drug metabolite prediction system based on conditional generative adversarial networks according to claim 8, characterized in that, The first module also includes: The fifth unit is used to construct an initial type feature matrix and an initial pairwise feature matrix from the masked initial features and input them into the shared encoder, which includes multiple attention blocks; The sixth unit is used to construct query, key, and value vectors based on the type feature matrix, calculate the attention between atoms based on the query vector and key vector, and correct it using pairwise feature matrices to obtain the corrected attention; The seventh unit is used to calculate the output vector under each attention head based on the corrected attention and value vector; The eighth unit is used to concatenate and map the outputs of each attention head in each layer to obtain the updated atomic features of each layer; The ninth unit is used to process the updated atomic features through residual connection and normalization to obtain the first encoding result, which includes atomic type features and paired features of atomic pairs.

10. The drug metabolite prediction system based on conditional generative adversarial networks according to claim 7, characterized in that, The second module includes: The tenth unit is used to obtain the preset dimensional information of the atomic positions; The eleventh unit is used to perform sine or cosine mapping based on the position of the atom in the molecule to obtain the encoded value of the atom in each dimension; The twelfth unit is used to combine the encoded values ​​of atoms in all dimensions into a positional encoded vector; The thirteenth unit is used to add the position encoding vector to the atomic type features in the first encoding result to obtain the initial input of the atomic type features of the first decoder.

Images