An antibody-drug conjugate activity prediction method and system based on virtual graphs and multi-scale features

Abstract

The application discloses an antibody conjugated drug binding prediction method based on a virtual graph and multi-scale features, proposes a drug carrier, a linker, an antibody heavy chain and a light chain, and an antigen target protein initial feature construction method, proposes a 1DCNN for extracting drug sequence features, designs a drug feature extraction method based on a graph virtual node, introduces a virtual node and a virtual edge into a molecular structure graph, takes a Graph Transformer as a graph feature extractor, takes a virtual node feature as a drug representation, then inputs protein and drug features into a feature fusion module, adds an attention mechanism and a gated skip connection mechanism in feature fusion, captures potential interactions while fusing feature information of different hidden layers, and realizes higher precision affinity prediction. The application can solve the technical problems that existing methods are difficult to extract structural features of antibody conjugated drugs and features of different components are difficult to fuse.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of biological medicine, and more particularly relates to an antibody-drug conjugate activity prediction method and system based on a virtual graph and multi-scale features. BACKGROUND

[0002] Antibody-drug conjugates (ADC) are a new cancer treatment method that combines monoclonal antibodies with highly cytotoxic payloads through chemical linkers, combining the targeting specificity of antibodies with the high toxicity of payloads, thereby ensuring safety while achieving efficient cell killing. Once the antibody-drug conjugate binds to the antigen, it enters the cell through receptor-mediated endocytosis and releases cytotoxic drugs to destroy cancer cells. However, endocytosis efficiency is affected by factors such as target type, binding affinity, and binding strength. If the activity of the antibody-drug conjugate is insufficient, it may lead to inefficient endocytosis or off-target release, thereby causing systemic toxicity problems. Therefore, the activity prediction of antibody-drug conjugates is a crucial problem in the field of drug discovery. However, traditional biological experiments are time-consuming and labor-intensive, so methods for predicting the activity of antibody-drug conjugates using deep learning have emerged.

[0003] Currently, there is only one method for predicting antibody-drug conjugates based on deep learning. This method uses a large language model to obtain feature vectors for the antibody heavy chain, antibody light chain, antigen, linker, and drug carrier, as well as the drug-antibody ratio. After concatenating these, it inputs them into a multi-layer perceptron to predict the activity of the antibody-drug conjugate.

[0004] However, the above-mentioned method for predicting the activity of antibody-drug conjugates has some non-negligible defects:

[0005] First, the above method only analyzes individual molecular features, and does not analyze from the perspective of the whole antibody-drug conjugate, making it difficult to capture the complex relationships between the various components of the antibody-drug conjugate.

[0006] Second, the above method does not fully learn the global features of the drug carrier and linker, making it difficult to fully extract high-dimensional features of the drug carrier and linker.

[0007] Third, the above method does not analyze the features of the drug and linker from multiple perspectives, making it difficult to achieve high-precision prediction of the activity of the antibody-drug conjugate. SUMMARY

[0008] In view of the above defects or improvement needs of the prior art, the present application provides an antibody conjugated drug binding prediction method and system based on virtual graphs and multi-scale features, which aims to solve the technical problems that the existing antibody conjugated drug activity prediction method only analyzes individual molecular features, does not analyze from the overall perspective of the antibody conjugated drug, and is difficult to capture the complex relationship between the components of the antibody conjugated drug, and the technical problems that it is difficult to fully extract the high-dimensional features of the drug carrier and linker due to insufficient learning of the global features of the drug carrier and linker, and the technical problems that it is difficult to achieve high-precision prediction of the activity of the antibody conjugated drug due to the lack of analysis of the features of the drug and linker from multiple angles.

[0009] To achieve the above-mentioned purpose, according to one aspect of the present application, an antibody conjugated drug binding prediction method based on virtual graphs and multi-scale features is provided, comprising the following steps:

[0010] (1) Obtain a pair of antibody conjugated drugs to be predicted, and preprocess the pair of antibody conjugated drug data to obtain a pair of preprocessed antibody conjugated drug data, which includes the simplified molecular linear input specification (SMILES) sequence of the drug carrier and linker, the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, the amino acid sequence of the antigen, the drug-antibody ratio (DAR) value, and the label of the antibody conjugated drug.

[0011] (2) For each of the amino acid sequences of the antibody heavy chain, the amino acid sequences of the antibody light chain, and the amino acid sequences of the antigen in the pair of preprocessed antibody conjugated drug data obtained in step (1), input the amino acid sequences of the antibody heavy chain, the amino acid sequences of the antibody light chain, and the amino acid sequences of the antigen into the protein large model ESM-2 to obtain the feature vectors corresponding to the amino acid sequences of the antibody heavy chain, the amino acid sequences of the antibody light chain, and the amino acid sequences of the antigen, respectively.

[0012] (3) For each SMILES sequence of the drug carrier and linker in the pair of preprocessed antibody conjugated drug data obtained in step (1), convert the SMILES sequence into a molecular structure graph, and add virtual nodes and feature encodings to the molecular structure graph to obtain the molecular structure graph corresponding to the SMILES sequence.

[0013] (4) For each SMILES sequence of the drug carrier and linker in the pair of preprocessed antibody conjugated drug data obtained in step (1), convert the SMILES sequence into a molecular fingerprint feature using the open source chemical library RDKit as the Morgan feature vector corresponding to the SMILES sequence.

[0014] (5) input the feature vectors corresponding to the amino acid sequences of all antibody heavy chains, the amino acid sequences of all antibody light chains, and the amino acid sequences of all antigens of the antibody-drug conjugate data pair obtained in step (2), the molecular structure graphs corresponding to the SMILES sequences of all drug carriers and linkers obtained in step (3), and the Morgan feature vectors corresponding to the SMILES sequences of all drug carriers and linkers obtained in step (4) into the pre-trained antibody-drug conjugate activity prediction model to obtain the activity prediction result of the antibody-drug conjugate data pair.

[0015] Preferably, the pre-processing of the antibody-drug conjugate data pair to be predicted in step (1) is specifically as follows:

[0016] First, the antibody-drug conjugate data pair to be predicted is screened to delete data with missing data to obtain the SMILES sequences of drug carriers and linkers, the amino acid sequences of antibody heavy chains, the amino acid sequences of antibody light chains, the amino acid sequences of antigens, the antibody-drug ratio, and the biological activity test data including IC50, EC50 and GI50.

[0017] Then, the label of the antibody-drug conjugate is obtained according to the following rules, which is specifically as follows: the biological activity test data including IC50, EC50 and GI50 are uniformly converted to nM unit at the cell level, if an antibody-drug conjugate has multiple biological activity values, the invalid value is discarded, and the remaining minimum biological activity value is reserved as the only biological activity value of the antibody-drug conjugate; for the antibody-drug conjugate that has been approved or is in clinical I, II, III phase, it is labeled as “positive sample”; for the antibody-drug conjugate in the experimental stage, if its IC50, EC50 or GI50 is less than or equal to 100 nM, it is labeled as “positive sample”, otherwise, it is labeled as “negative sample”.

[0018] Finally, the SMILES sequences of drug carriers and linkers, the amino acid sequences of antibody heavy chains and light chains and antigen proteins, the antibody-drug ratio, and the label of the antibody-drug conjugate are combined to obtain the pre-processed antibody-drug conjugate data pair.

[0019] Preferably, step (2) specifically comprises the following steps: first, converting the amino acid sequence of each antibody heavy chain, the amino acid sequence of each antibody light chain, and the amino acid sequence of each antigen into a token vector by the alphabet tool of the esm library; then, calling the pretrained.esm2_t33_650M_UR50D model of the esm library as the ESM-2 model, and inputting the token vector corresponding to the amino acid sequence of each antibody heavy chain, the token vector corresponding to the amino acid sequence of each antibody light chain, and the token vector corresponding to the amino acid sequence of each antigen into the ESM-2 model respectively to obtain the feature vector (with a length of 1280 dimensions) corresponding to the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, and the amino acid sequence of the antigen respectively.

[0020] Step (3) specifically comprises the following steps: first, converting the SMILES sequence into a molecular structure graph G=(V, E) using the open source chemical library RDKit, wherein V represents a set of all atoms in the SMILES sequence, and E represents a set of chemical bonds between atoms in the SMILES sequence; then, adding a virtual node vn to the molecular structure graph G, so that the virtual node is connected to all atoms in the molecular structure graph G and forms a virtual edge; subsequently, performing feature encoding on the molecular structure graph after adding the virtual node to obtain an M-dimensional atom feature vector and an N-dimensional bond feature vector, and the two vectors constitute the molecular structure graph corresponding to the SMILES sequence; wherein M is equal to 44, and N is equal to 10.

[0021] Preferably, the antibody conjugated drug activity prediction model comprises an encoder feature extraction module based on a graph transformer, a virtual graph structure feature extraction module based on a graph transformer, a principal component analysis (PCA) dimension reduction module, and a multilayer perceptron module.

[0022] The encoder feature extraction module based on the graph transformer comprises a linear preprocessing layer, three transformer layers, and a virtual node extraction layer, and the specific structure thereof is as follows:

[0023] The input of the preprocessing layer is the M-dimensional atomic feature vector and the N-dimensional chemical bond feature vector in the molecular structure graph, which is first subjected to linear regression operation by a linear regression function to obtain an intermediate 128-dimensional atomic feature vector and an intermediate 128-dimensional chemical bond feature vector, respectively; then, the connection relationship between the nodes and edges in the molecular structure graph is used to obtain the unit matrix, the degree matrix, and the adjacency matrix corresponding to the molecular structure graph; then, the symmetric normalized Laplacian matrix is obtained according to the unit matrix, the degree matrix, and the adjacency matrix corresponding to the molecular structure graph, and the Laplacian position feature vector of the symmetric normalized Laplacian matrix is obtained; then, the intermediate atomic feature vector obtained after the linear regression operation and the Laplacian position feature vector are added to obtain a processed 128-dimensional atomic feature vector, and the final output is a 128-dimensional atomic feature vector and a 128-dimensional chemical bond feature vector.

[0024] The input of the first transformer layer is the 128-dimensional atomic feature vector and the 128-dimensional chemical bond feature vector output by the preprocessing layer, which is first subjected to attention mechanism calculation operation by four independent attention heads in the transformer to obtain the attention score between two atoms; then, the message passing processing is performed using the attention score to aggregate the current features of adjacent atoms to obtain a 16-dimensional intermediate atomic feature vector and a 16-dimensional intermediate chemical bond feature vector; then, the intermediate atomic feature vector and the intermediate chemical bond feature vector after the message passing processing are subjected to residual sum processing to obtain a processed 128-dimensional atomic feature vector and a processed 128-dimensional chemical bond feature vector, and the final output is a 128-dimensional atomic feature vector and a 128-dimensional chemical bond feature vector.

[0025] The input of the second transformer layer is the 128-dimensional atomic feature vector and the 128-dimensional chemical bond feature vector output by the first transformer layer, which is subjected to attention calculation, message passing, and residual sum processing by the attention head to output a 128-dimensional atomic feature vector and a 128-dimensional chemical bond feature vector, respectively.

[0026] The input of the third transformer layer is the 128-dimensional atomic feature vector and the 128-dimensional chemical bond feature vector output by the second transformer layer, which is subjected to attention calculation, message passing, and residual sum processing by the attention head to output a 128-dimensional atomic feature vector and a 128-dimensional chemical bond feature vector.

[0027] The input of the virtual node extraction layer is the 128-dimensional atomic feature vector output by the third transformer layer, and the feature vector of the last atom in the atomic feature vector is obtained according to the atomic number index, which is taken as the feature vector of the drug carrier and the feature vector of the linker. The output is a 128-dimensional drug carrier feature vector and a 128-dimensional linker feature vector.

[0028] Preferably, the virtual graph structure feature extraction module based on the graph transformer comprises a virtual graph construction layer, a preprocessing layer, three transformer layers, and a virtual node extraction layer.

[0029] The input of the virtual graph construction layer is the 1280-dimensional feature vector corresponding to the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, and the amino acid sequence of the antigen, and the drug carrier feature vector and the linker feature vector output by the encoder feature extraction module based on the graph transformer. First, a virtual graph construction operation is performed using the open source library DGL to obtain a virtual ADC graph G'=(V',E'), where V' represents a set of all nodes in the virtual ADC graph, and E' represents a set of edges connecting the nodes in the virtual ADC graph. Then, a virtual node vn is added to the obtained virtual ADC graph, so that the virtual node is connected to all nodes in the virtual ADC graph and forms a virtual edge. Subsequently, feature encoding operation is performed on the virtual ADC graph after adding the virtual node, and the 1280-dimensional feature vector corresponding to the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, and the amino acid sequence of the antigen, and the 128-dimensional drug feature vector of the drug carrier and the linker are inserted into the corresponding node position in the virtual ADC graph to obtain the node feature vector of the virtual ADC graph. The edges of different types are 0-1 encoded, and the DAR value is added to the edge feature between the linker and the antibody heavy chain and light chain to obtain a 7-dimensional edge feature vector of the virtual ADC graph. Finally, the output is the virtual ADC graph corresponding to the antibody conjugated drug and its 7-dimensional edge feature vector.

[0030] The input of the preprocessing layer is a 1280-dimensional amino acid sequence of an antibody heavy chain, an amino acid sequence of an antibody light chain, an amino acid sequence corresponding to a feature vector of an antigen, and a virtual ADC graph and its edge feature vector output by the virtual graph construction layer, which first performs linear regression operation on the 1280-dimensional feature vector and the 7-dimensional feature vector respectively through a linear regression function to obtain a 128-dimensional intermediate node feature vector and a 128-dimensional edge feature vector respectively; then, the connection relationship between the nodes and edges in the virtual ADC graph is used to obtain the unit matrix, the degree matrix and the adjacency matrix corresponding to the ADC virtual graph; then, the symmetric normalized Laplacian matrix is obtained according to the unit matrix, the degree matrix and the adjacency matrix corresponding to the virtual ADC graph, and the Laplacian position feature vector of the symmetric normalized Laplacian matrix is obtained; then, the intermediate node feature vector obtained after the linear regression operation and the Laplacian position feature vector are added to obtain a processed 128-dimensional node feature vector, and the final output is a 128-dimensional node feature vector and a 128-dimensional edge feature vector.

[0031] The input of the first transformer layer is the 128-dimensional node feature vector and the 128-dimensional edge feature vector output by the preprocessing layer, which first performs attention mechanism calculation operation on the node feature vector and the edge feature vector through four independent attention heads in the transformer to obtain the attention score between two atoms; then, the attention score is used for message passing processing to aggregate the current features of adjacent atoms to obtain a 16-dimensional intermediate node feature vector and a 16-dimensional intermediate edge feature vector respectively; then, the intermediate node feature vector and the intermediate edge feature vector after the message passing processing are subjected to residual sum processing to obtain a processed 128-dimensional node feature vector and a 128-dimensional edge feature vector, and the final output is a 128-dimensional node feature vector and a 128-dimensional edge feature vector.

[0032] The input of the second transformer layer is the 128-dimensional node feature vector and the 128-dimensional edge feature vector output by the first transformer layer, which performs attention calculation, message passing and residual sum processing on the node feature vector and the edge feature vector through attention heads respectively, and the output is a 128-dimensional node feature vector and a 128-dimensional edge feature vector.

[0033] The input of the third transformer layer is the 128-dimensional node feature vector and the 128-dimensional edge feature vector output by the second transformer layer, which performs attention calculation, message passing and residual sum processing on the node feature vector and the edge feature vector through attention heads respectively, and the output is a 128-dimensional node feature vector and a 128-dimensional edge feature vector.

[0034] The input of the virtual node extraction layer is the 128-dimensional node feature vector output by the third transformer layer, which obtains the feature vector of the last node in the node feature vector according to the node sequence index, and uses it as the feature vector of the virtual ADC graph. The output is a 128-dimensional feature vector of the virtual ADC graph.

[0035] Preferably, the input of the PCA dimension reduction module is a 1280-dimensional amino acid sequence of the antibody heavy chain, a 1280-dimensional amino acid sequence of the antibody light chain, a 1024-dimensional Morgan feature vector corresponding to the feature vector of the antigen amino acid sequence, a 1024-dimensional Morgan feature vector of the drug carrier and the linker, and a 1-dimensional DAR value. First, these features are spliced into a 5889-dimensional vector, and then the 5889-dimensional vector is subjected to dimension reduction and linear regression operation to obtain a 128-dimensional feature vector and output.

[0036] The multi-layer perception module includes a splicing layer and a multi-layer perception layer, and the specific structure is as follows:

[0037] The input of the splicing layer is the 128-dimensional vector obtained by the virtual graph structure feature extraction module based on the graph transformer and the 128-dimensional vector obtained by the PCA dimension reduction module. The two are spliced, and the output is a 256-dimensional feature vector.

[0038] The multi-layer perception layer includes three linear layers, and the specific structure is as follows:

[0039] The input of the first linear layer is the 256-dimensional feature vector obtained by the splicing layer, which is processed by the activation function on the feature vector, and the output is a 1024-dimensional feature vector;

[0040] The input of the second linear layer is the 256-dimensional feature vector obtained by the first linear layer, which is processed by the activation function on the feature vector, and the output is a 128-dimensional feature vector;

[0041] The input of the third linear layer is the 128-dimensional feature vector obtained by the second linear layer, which is processed by the activation function on the feature vector, and the output is a 1-dimensional ADC activity prediction value.

[0042] Preferably, the antibody conjugated drug activity prediction model is obtained by the following steps:

[0043] (5-1) Obtain a plurality of antibody conjugated drug data pairs to be predicted, and preprocess all the antibody conjugated drug data pairs to be predicted to obtain a plurality of preprocessed antibody conjugated drug data pairs, including SMILES sequences of drug carriers and linkers, amino acid sequences of antibody heavy chains, amino acid sequences of antibody light chains, amino acid sequences of antigens, DAR values, and label values of antibody conjugated drugs, and all the preprocessed antibody conjugated drug data pairs are divided into a training set and a test set according to a ratio of 4:1;

[0044] (5-2) For each of the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, and the amino acid sequence of the antigen in the training set obtained in step (1), input the amino acid sequences into the protein large model ESM-2 to obtain the feature vector corresponding to each of the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, and the amino acid sequence of the antigen.

[0045] (5-3) For each of the SMILES sequence of the drug carrier and the linker in the training set obtained in step (1), convert the SMILES sequence into a molecular structure graph, and add a virtual node and a feature code in the molecular structure graph to obtain a molecular structure graph corresponding to the SMILES sequence, which includes an M-dimensional atomic feature vector and an N-dimensional chemical bond feature vector, where M equals 44 and N equals 10.

[0046] (5-4) For each of the SMILES sequence of the drug carrier and the linker in the training set obtained in step (5-1), use the open source chemical library RDKit to convert the SMILES sequence into a molecular fingerprint feature representation to obtain a Morgan feature vector corresponding to the SMILES sequence of the drug carrier and the linker.

[0047] (5-5) For each of the SMILES sequence of the drug carrier and the linker in each of the antibody-conjugated drug data pairs in the training set obtained in step (5-1), input the M-dimensional atomic feature vector and the N-dimensional chemical bond feature vector in the molecular structure graph corresponding to the SMILES sequence obtained in step (5-3) into the mapping layer of the encoder feature extraction module for mapping processing to obtain a 128-dimensional intermediate atomic feature vector of each (i.e., the i-th) atom and a 128-dimensional chemical bond feature vector e ij (0) between the i-th atom and the j-th atom , where i ∈ [1, the total number of atoms in the molecular structure graph corresponding to the SMILES sequence], j ∈ [1, the total number of atoms in the molecular structure graph corresponding to the SMILES sequence], and i ≠ j.

[0048] (5-6) For each of the SMILES sequence of the drug carrier and the linker in each of the antibody-conjugated drug data pairs in the training set obtained in step (5-1), input the intermediate atomic feature vector of the i-th atom obtained in step (5-5) into the preprocessing layer of the encoder feature extraction module for Laplacian feature position embedding processing to obtain a processed 128-dimensional atomic feature vector

[0049] (5-7) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the attention score corresponding to the i-th atom and the j-th atom obtained in step (5-6) and the chemical bond feature vector e between the i-th atom and the j-th atom obtained in step (5-5) ij (0) The first transformer layer of the input encoder feature extraction module is inputted to calculate the attention score between the i-th atom and the j-th atom

[0050] (5-8) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the attention score corresponding to the i-th atom and the j-th atom obtained in step (5-7) The first transformer layer of the input encoder feature extraction module is inputted to perform message passing processing to obtain the updated intermediate atom feature vector corresponding to the i-th atom and the intermediate chemical bond feature vector corresponding to the i-th atom and the j-th atom

[0051] (5-9) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the intermediate atom feature vector corresponding to the i-th atom obtained in step (5-8) and the intermediate chemical bond feature vector corresponding to the i-th atom and the j-th atom The first transformer layer of the input encoder feature extraction module is inputted to splice the intermediate atom feature vectors and the intermediate chemical bond feature vectors obtained by all attention heads in the first transformer layer, and add the spliced results to the original atom feature vector and the chemical bond feature vector inputted to the transformer layer to obtain the processed intermediate atom feature vector and the intermediate chemical bond feature vector corresponding to the i-th atom

[0052] (5-10) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the intermediate atom feature vector corresponding to the i-th atom obtained in step (5-9) and the intermediate chemical bond feature vector are respectively subjected to residual sum processing to obtain the processed atom feature vector ​and the bond feature vector corresponding to the i-th atom and the j-th atom

[0053] (5-11) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the atomic feature vector corresponding to the i-th atom obtained in step (5-10) and the bond feature vector corresponding to the i-th atom and the j-th atom The second transformer layer of the input encoder feature extraction module is processed in the same way as steps (5-7) to (5-10) above to obtain the feature vector corresponding to the i-th atom in the final output of the second transformer layer and the bond feature vector corresponding to the i-th atom and the j-th atom By analogy, after the processing of the third transformer layer, the atomic feature vector corresponding to the i-th atom in the final output of the third transformer layer is obtained and the bond feature vector corresponding to the i-th atom and the j-th atom

[0054] (5-12) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the atomic feature vector corresponding to the i-th atom obtained in step (5-11) The virtual node extraction layer of the input encoder feature extraction module is searched to obtain the 128-dimensional drug carrier feature vector and the 128-dimensional linker feature vector corresponding to the i-th atom.

[0055] (5-13) The feature vector corresponding to each antibody heavy chain amino acid sequence, each antibody light chain amino acid sequence, and each antigen amino acid sequence obtained in step (5-2) and each drug carrier feature vector and each linker feature vector obtained in step (5-12) are input into the graph construction layer of the structure feature extraction module to obtain a virtual ADC graph.

[0056] (5-14) The virtual ADC graph obtained in step (5-13) is input into the linear preprocessing layer of the structure feature extraction module for dimension mapping processing to obtain a 128-dimensional node feature vector and a 128-dimensional edge feature vector.

[0057] (5-15) The 128-dimensional node feature vector and the 128-dimensional edge feature vector obtained in step (5-14) are input into the three transformer layers of the graph transformer-based virtual graph structure feature extraction module for feature updating processing to obtain updated 128-dimensional node feature vectors and 128-dimensional edge feature vectors.

[0058] (5-16) The updated node feature vector and edge feature vector obtained in step (5-15) are input into the virtual node extraction layer for a lookup operation to obtain a 128-dimensional feature vector of the virtual ADC graph.

[0059] (5-17) The DAR value obtained in step (5-1), the feature vectors corresponding to the amino acid sequences of each antibody heavy chain, the amino acid sequences of each antibody light chain, and the amino acid sequences of each antigen obtained in step (5-2), and the Morgan feature vectors corresponding to the SMILES sequences of each drug carrier and linker obtained in step (5-4) are spliced to obtain a 5889-dimensional feature vector, which is then input into a PCA dimension reduction module for dimension reduction processing to obtain a 128-dimensional feature vector.

[0060] (5-18) The feature vector of the virtual ADC graph obtained in step (5-16) and the 128-dimensional feature vector obtained in step (5-17) are spliced to obtain a 256-dimensional fusion feature vector x, which is then input into a multi-layer perceptron module for training processing to obtain an activity prediction value p t .

[0061] (5-19) The label value in the antibody-drug conjugate data pair obtained in step (5-1) and the activity prediction value p The loss function is calculated, and the drug target binding affinity prediction model is trained using the loss function until the antibody-drug conjugate activity prediction model converges, thereby obtaining a preliminarily trained antibody-drug conjugate activity prediction model.

[0062] (5-20) The test set obtained in step (5-1) is tested on the preliminarily trained drug target binding affinity prediction model in step (5-19) until the value of the loss function reaches the optimal value, thereby obtaining a trained drug target binding affinity prediction model.

[0063] Preferably, the specific calculation formula in step (5-7) is:

[0064]

[0065] wherein are the linear transformation matrices of the feature vectors of the i th atom and the j th atom in the first transformer processing process, is the linear transformation matrix of the feature vector of the chemical bond between the i th atom and the j th atom in the first transformer layer, d h is the feature vector dimension of the linear transformation matrix .

[0066] Step (5-8) is calculated by the following formula:

[0067]

[0068] wherein N i represents the adjacent atom set composed of all adjacent atoms of the i-th atom in the SMILES sequence, is the linear transformation matrix of the adjacent atom feature vector of the i-th atom of the first transformer layer, ω ij is the attention score after the Softmax function activation.

[0069] In step (5-8), the message passing for the chemical bond is directly assigned to the intermediate chemical bond feature vector i.e. i.e.

[0070] Step (5-10) is calculated by the following formula:

[0071]

[0072] wherein Norm represents normalization, and represents a learnable parameter, and ReLU represents an activation function.

[0073] Preferably, step (5-18) is calculated by the following formula:

[0074] p t = Linear(ReLU(Linear(ReLU(Linear(x))))

[0075] In the three linear layers constituting the multi-layer perception module, the input dimension of the first linear layer is 256, the output dimension is 1024, the input dimension of the second linear layer is 1024, the output dimension is 128, the input dimension of the third linear layer is 128, the output dimension is 1, and x represents a 256-dimensional fusion feature vector.

[0076] The loss function of step (5-19) is calculated by the following formula:

[0077]

[0078] wherein α t is a balance factor adjusting the weight of positive and negative samples, and γ is a focusing parameter adjusting the influence of samples difficult to classify.

[0079] According to another aspect of the present application, there is provided an antibody-drug conjugate binding prediction method based on a virtual graph and multi-scale features, comprising:

[0080] The first module is configured to obtain an antibody-drug conjugate data pair to be predicted, and pre-process the antibody-drug conjugate data pair to be predicted to obtain a pre-processed antibody-drug conjugate data pair, which includes a simplified molecular-input line-entry system (SMILES) sequence of a drug carrier and a linker, an amino acid sequence of an antibody heavy chain, an amino acid sequence of an antibody light chain, an amino acid sequence of an antigen, a drug-antibody ratio (DAR) value, and a label of the antibody-drug conjugate.

[0081] The second module is configured to input, for each of the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, and the amino acid sequence of the antigen in the pre-processed antibody-drug conjugate data pair obtained by the first module, the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, and the amino acid sequence of the antigen into a protein large model ESM-2 to obtain a feature vector corresponding to the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, and the amino acid sequence of the antigen, respectively.

[0082] The third module is configured to, for each of the SMILES sequence of the drug carrier and the linker in the pre-processed antibody-drug conjugate data pair obtained by the first module, convert the SMILES sequence into a molecular structure graph, and add a virtual node and a feature code to the molecular structure graph to obtain a molecular structure graph corresponding to the SMILES sequence.

[0083] The fourth module is configured to, for each of the SMILES sequence of the drug carrier and the linker in the pre-processed antibody-drug conjugate data pair obtained by the first module, convert the SMILES sequence into a molecular fingerprint feature using an open source chemical library RDKit as a Morgan feature vector corresponding to the SMILES sequence.

[0084] The fifth module is configured to input, into a pre-trained antibody-drug conjugate activity prediction model, the feature vectors corresponding to all of the amino acid sequences of the antibody heavy chain, the amino acid sequences of the antibody light chain, and the amino acid sequences of the antigen in the antibody-drug conjugate data pair obtained by the second module, the molecular structure graphs corresponding to all of the SMILES sequences of the drug carrier and the linker obtained by the third module, and the Morgan feature vectors corresponding to all of the SMILES sequences of the drug carrier and the linker obtained by the fourth module, to obtain an activity prediction result of the antibody-drug conjugate data pair.

[0085] Overall, the above technical solutions conceived by the present application can achieve the following beneficial effects compared with the prior art:

[0086] (1) The present application adopts steps (2) to (5), proposes the construction of an antibody conjugate drug virtual graph, simulates the real structure of the antibody conjugate drug, and fully captures the interaction between the antigen, the antibody, the drug carrier and the linker;

[0087] (2) The present application adopts steps (2) to (5), proposes the idea of a virtual node, takes the output virtual node feature as the final drug representation feature, and connects the virtual node with the real atomic node through a virtual edge, so that all atomic high-dimensional feature information and global information can be automatically learned in the graph convolution;

[0088] (3) The present application adopts steps (2) to (5), analyzes the feature vectors of the various components of the antibody conjugate drug from multiple scales, and thus can improve the prediction accuracy of the activity of the antibody conjugate drug. BRIEF DESCRIPTION OF DRAWINGS

[0089] Figure 1 is the overall flowchart of the drug target affinity prediction method based on the graph virtual node of the present application;

[0090] Figure 2 is a network structure diagram of the drug target affinity prediction model used in the method of the present application. DETAILED DESCRIPTION

[0091] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0092] The basic method of the present application is that, in order to overcome the above defects or improve the needs of the prior art, the present application provides an antibody conjugate drug binding prediction method and system based on a virtual graph and multi-scale features, proposes an initial feature construction method for a drug carrier, a linker, an antibody heavy chain and a light chain, an antigen target protein, proposes to extract drug sequence features using 1DCNN, designs a drug feature extraction method based on a graph virtual node, introduces virtual nodes and virtual edges in the molecular structure graph, uses GraphTransformer as a graph feature extractor, uses virtual node features as drug representation, and then inputs the protein and drug features into a feature fusion module. The attention mechanism and the gated skip connection mechanism are added in the feature fusion to capture the potential interaction and fuse the feature information of different hidden layers, so as to realize higher precision affinity prediction.

[0093] The present application aims to solve the technical problems of the existing method that the structural features of antibody conjugated drugs are difficult to extract and the features of different components are difficult to fuse.

[0094] The present application has the advantage of improving the prediction of the effective performance of antibody conjugated drugs, which can be used for a large number of drug screening.

[0095] In order to make the purpose, technical method and advantages of the present application more clear and explicit, the present application will be further described in detail below in combination with the drawings.

[0096] As shown in the accompanying drawings, Figure 1 The present application provides an antibody conjugated drug binding prediction method based on virtual graph and multi-scale features, comprising the following steps:

[0097] (1) Obtain the antibody conjugated drug pair to be predicted, and pre-process the antibody conjugated drug data pair to be predicted to obtain the pre-processed antibody conjugated drug data pair, which includes the SMILES sequence of the drug carrier and the linker, the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, the amino acid sequence of the antigen, the drug-to-antibody ratio (DAR value) and the label of the antibody conjugated drug.

[0098] In this step, the pre-processing of the antibody conjugated drug data pair to be predicted to obtain the pre-processed antibody conjugated drug data pair is specifically as follows:

[0099] Firstly, the antibody conjugated drug data pair to be predicted is screened, and the data with data missing problem is deleted to obtain the SMILES sequence of the drug carrier and the linker, the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, the amino acid sequence of the antigen, the drug-to-antibody ratio, and the biological activity test data, including IC50, EC50 and GI50;

[0100] Then, the label of the antibody drug conjugate is obtained according to the following rules, and the process is as follows: the biological activity test data including IC50, EC50 and GI50 are uniformly converted into nM units at the cell level, if an antibody drug conjugate has multiple biological activity values, the invalid value is discarded, and the remaining minimum biological activity value is reserved as the only biological activity value of the antibody drug conjugate; for the antibody drug conjugate approved or in clinical I, II, III phase, it is labeled as a "positive sample"; for the antibody drug conjugate in the experimental stage, if its IC50, EC50 or GI50 is less than or equal to 100 nM, it is labeled as a "positive sample", otherwise, it is labeled as a "negative sample".

[0101] Finally, the obtained SMILES sequence of the drug carrier and the linker, the amino acid sequence of the antibody heavy chain and the antigen protein, the antibody drug ratio, and the label of the antibody drug are combined into the preprocessed antibody drug conjugate data pair.

[0102] (2) For each amino acid sequence of the antibody heavy chain, each amino acid sequence of the antibody light chain, and each amino acid sequence of the antigen in the preprocessed antibody drug conjugate data pair obtained in step (1), the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, and the amino acid sequence of the antigen are input into the protein large model ESM-2 to obtain the feature vectors corresponding to the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, and the amino acid sequence of the antigen, respectively.

[0103] This step specifically converts the amino acid sequence of each antibody heavy chain, the amino acid sequence of each antibody light chain, and the amino acid sequence of each antigen into a token vector through the alphabet tool of the esm library, then calls the pretrained.esm2_t33_650M_UR50D model of the esm library as the ESM-2 model, and inputs the token vector corresponding to the amino acid sequence of each antibody heavy chain, the amino acid sequence of each antibody light chain, and the amino acid sequence of each antigen into the ESM-2 model to obtain the feature vectors corresponding to the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, and the amino acid sequence of the antigen, respectively (the length is 1280 dimensions).

[0104] The advantage of this step (2) is that the ESM-2 model is used for feature extraction, which can more accurately capture the hidden structural features in the protein sequence and better utilize the evolutionary information of the protein.

[0105] (3) For each SMILES sequence of the drug carrier and the linker in the preprocessed antibody-drug conjugate data pair obtained in step (1), the SMILES sequence is converted into a molecular structure graph, and a virtual node and a feature code are added to the molecular structure graph to obtain a molecular structure graph corresponding to the SMILES sequence.

[0106] Specifically, first, the SMILES sequence is converted into a molecular structure graph G=(V, E) using the open source chemical library RDKit, where V represents a set of all atoms in the SMILES sequence, and E represents a set of chemical bonds between atoms in the SMILES sequence; then, a virtual node vn is added to the molecular structure graph G, so that the virtual node is connected to all atoms in the molecular structure graph G and forms a virtual edge; subsequently, the molecular structure graph after adding the virtual node is feature coded (the initial feature vector of each atom in the molecular structure graph is determined according to its symbol, formal charge, atomic hybridization, atomic chirality, etc., and the initial feature vector of each chemical bond is derived from the properties of the chemical bond, such as the type of the bond, aromaticity or conjugation), to obtain an M-dimensional atomic feature vector and an N-dimensional chemical bond feature vector, which constitute the molecular structure graph corresponding to the SMILES sequence; where M is equal to 44 in the present application, and N is equal to 10 in the present application.

[0107] The advantage of the present application is that the SMILES sequence of the drug carrier and the linker is converted into a molecular structure graph, so as to fully consider the structural information of the molecule, and the idea of the virtual node is proposed, so that the model can extract the global features of the drug and improve the performance of the model.

[0108] (4) For each SMILES sequence of the drug carrier and the linker in the preprocessed antibody-drug conjugate data pair obtained in step (1), the SMILES sequence is converted into a molecular fingerprint feature (with a length of 1024) using the open source chemical library RDKit, as a Morgan feature vector corresponding to the SMILES sequence.

[0109] (5) The feature vectors corresponding to the amino acid sequences of all antibody heavy chains, the amino acid sequences of all antibody light chains, and the amino acid sequences of all antigens in the antibody-drug conjugate data pair obtained in step (2), the molecular structure graphs corresponding to the SMILES sequences of all drug carriers and linkers obtained in step (3), and the Morgan feature vectors corresponding to the SMILES sequences of all drug carriers and linkers obtained in step (4) are input into the pre-trained antibody-drug conjugate activity prediction model to obtain the activity prediction result of the antibody-drug conjugate data pair.

[0110] The antibody conjugate drug activity prediction model comprises an encoder feature extraction module based on a graph transformer, a virtual graph structure feature extraction module based on a graph transformer, a principal component analysis (PCA) dimension reduction module, and a multilayer perceptron module.

[0111] The encoder feature extraction module based on the graph transformer comprises a linear preprocessing layer, three transformer layers, and a virtual node extraction layer, and the specific structure is as follows:

[0112] The input of the preprocessing layer is an M-dimensional atomic feature vector and an N-dimensional bond feature vector in a molecular structure graph. The linear regression function is used to perform linear regression operation on the M-dimensional atomic feature vector and the N-dimensional bond feature vector, respectively, to obtain an intermediate atomic feature vector of 128 dimensions and a bond feature vector of 128 dimensions. Then, the connection relationship between the nodes and edges in the molecular structure graph is used to obtain the unit matrix, the degree matrix, and the adjacency matrix corresponding to the molecular structure graph. Then, the symmetric normalized Laplacian matrix is obtained according to the unit matrix, the degree matrix, and the adjacency matrix corresponding to the molecular structure graph, and the Laplacian position feature vector of the symmetric normalized Laplacian matrix is obtained. Then, the intermediate atomic feature vector obtained after the linear regression operation and the Laplacian position feature vector are added to obtain a processed atomic feature vector of 128 dimensions. Finally, the output is an atomic feature vector of 128 dimensions and a bond feature vector of 128 dimensions.

[0113] The input of the first transformer layer is the 128-dimensional atomic feature vector and the 128-dimensional bond feature vector output by the preprocessing layer. The four independent attention heads inside the transformer are used to perform attention mechanism calculation operation on the atomic feature vector and the bond feature vector in sequence to obtain the attention score between two atoms. Then, the message passing processing is performed using the attention score to aggregate the current features of adjacent atoms to obtain an intermediate atomic feature vector of 16 dimensions and an intermediate bond feature vector of 16 dimensions. Then, the intermediate atomic feature vector and the intermediate bond feature vector after the message passing processing are subjected to residual sum processing to obtain a processed atomic feature vector of 128 dimensions and a processed bond feature vector of 128 dimensions. Finally, the output is an atomic feature vector of 128 dimensions and a bond feature vector of 128 dimensions.

[0114] The input of the second transformer layer is the 128-dimensional atomic feature vector and the 128-dimensional chemical bond feature vector output by the first transformer layer, which are respectively subjected to attention calculation, message passing and residual sum processing by the attention head, and the output is respectively a 128-dimensional atomic feature vector and a 128-dimensional chemical bond feature vector.

[0115] The input of the third transformer layer is the 128-dimensional atomic feature vector and the 128-dimensional chemical bond feature vector output by the second transformer layer, which are respectively subjected to attention calculation, message passing and residual sum processing by the attention head, and the output is a 128-dimensional atomic feature vector and a 128-dimensional chemical bond feature vector.

[0116] The input of the virtual node extraction layer is the 128-dimensional atomic feature vector output by the third transformer layer, which is indexed according to the atomic number to obtain the feature vector of the last atom in the atomic feature vector, which is taken as the feature vector of the drug carrier and the feature vector of the linker, and the output is a 128-dimensional drug carrier feature vector and a 128-dimensional linker feature vector.

[0117] The virtual graph structure feature extraction module based on the graph transformer includes a virtual graph construction layer, a preprocessing layer, three transformer layers and a virtual node extraction layer, and the specific structure is as follows:

[0118] The input of the virtual graph construction layer is the 1280-dimensional feature vector corresponding to the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain and the amino acid sequence of the antigen, and the drug carrier feature vector and the linker feature vector output by the graph transformer-based encoder feature extraction module. First, the open source library DGL is used for virtual graph construction operation to obtain a virtual ADC graph G'=(V',E'), wherein V' represents a set of all nodes in the virtual ADC graph, and E' represents a set of edges connected between the nodes in the virtual ADC graph. Then, a virtual node vn is added to the obtained virtual ADC graph, so that the virtual node is connected with all the nodes in the virtual ADC graph and forms a virtual edge. Subsequently, the virtual ADC graph after adding the virtual node is subjected to feature coding operation, and the 1280-dimensional feature vectors corresponding to the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain and the amino acid sequence of the antigen, and the 128-dimensional drug carrier and linker feature vectors are inserted into the corresponding node positions in the virtual ADC graph to obtain the node feature vectors of the virtual ADC graph. The edges of different types are subjected to 0-1 coding, and the DAR value is added to the edge feature between the linker and the antibody heavy chain and light chain to obtain a 7-dimensional edge feature vector of the virtual ADC graph. Finally, the output is the virtual ADC graph corresponding to the antibody conjugated drug and the 7-dimensional edge feature vector thereof.

[0119] The input of the preprocessing layer is a 1280-dimensional amino acid sequence of an antibody heavy chain, an amino acid sequence of an antibody light chain, an amino acid sequence corresponding to a feature vector of an antigen, and a virtual ADC graph and its edge feature vector output by the virtual graph construction layer, which first performs linear regression operation on the 1280-dimensional feature vector and the 7-dimensional feature vector respectively through a linear regression function to obtain a 128-dimensional intermediate node feature vector and a 128-dimensional edge feature vector respectively; then, the connection relationship between the nodes and edges in the virtual ADC graph is used to obtain the unit matrix, the degree matrix and the adjacency matrix corresponding to the ADC virtual graph; then, the symmetric normalized Laplacian matrix is obtained according to the unit matrix, the degree matrix and the adjacency matrix corresponding to the virtual ADC graph, and the Laplacian position feature vector of the symmetric normalized Laplacian matrix is obtained; then, the intermediate node feature vector obtained after the linear regression operation and the Laplacian position feature vector are added to obtain a processed 128-dimensional node feature vector, and the final output is a 128-dimensional node feature vector and a 128-dimensional edge feature vector.

[0120] The input of the first transformer layer is the 128-dimensional node feature vector and the 128-dimensional edge feature vector output by the preprocessing layer, which first performs attention mechanism calculation operation on the node feature vector and the edge feature vector through four independent attention heads in the transformer to obtain the attention score between two atoms; then, the attention score is used for message passing processing to aggregate the current features of adjacent atoms to obtain a 16-dimensional intermediate node feature vector and a 16-dimensional intermediate edge feature vector respectively; then, the intermediate node feature vector and the intermediate edge feature vector after the message passing processing are subjected to residual sum processing to obtain a processed 128-dimensional node feature vector and a 128-dimensional edge feature vector, and the final output is a 128-dimensional node feature vector and a 128-dimensional edge feature vector.

[0121] The input of the second transformer layer is the 128-dimensional node feature vector and the 128-dimensional edge feature vector output by the first transformer layer, which performs attention calculation, message passing and residual sum processing on the node feature vector and the edge feature vector through attention heads respectively, and the output is a 128-dimensional node feature vector and a 128-dimensional edge feature vector.

[0122] The input of the third transformer layer is the 128-dimensional node feature vector and the 128-dimensional edge feature vector output by the second transformer layer, which performs attention calculation, message passing and residual sum processing on the node feature vector and the edge feature vector through attention heads respectively, and the output is a 128-dimensional node feature vector and a 128-dimensional edge feature vector.

[0123] The input of the virtual node extraction layer is the 128-dimensional node feature vector output by the third transformer layer, the feature vector of the last node in the node feature vector is obtained according to the node sequence index, and the feature vector of the virtual ADC graph is used as the output, and the output is a 128-dimensional feature vector of the virtual ADC graph.

[0124] The input of the PCA dimension reduction module is a 1280-dimensional amino acid sequence of an antibody heavy chain, an amino acid sequence of an antibody light chain, a feature vector corresponding to an amino acid sequence of an antigen, a 1024-dimensional Morgan feature vector of a drug carrier and a linker and a 1-dimensional DAR value, which is first spliced into a 5889-dimensional vector, and then the 5889-dimensional vector is dimensionally reduced and linearly regressed to obtain a 128-dimensional feature vector and output.

[0125] The multi-layer perception module includes a splicing layer and a multi-layer perception layer, and the specific structure is as follows:

[0126] The input of the splicing layer is a 128-dimensional vector obtained by the virtual graph structure feature extraction module based on the graph transformer and a 128-dimensional vector obtained by the PCA dimension reduction module, and the two are spliced, and the output is a 256-dimensional feature vector.

[0127] The multi-layer perception layer includes three linear layers, and the specific structure is as follows:

[0128] The input of the first linear layer is a 256-dimensional feature vector obtained by the splicing layer, which is processed by an activation function on the feature vector, and the output is a 1024-dimensional feature vector;

[0129] The input of the second linear layer is a 256-dimensional feature vector obtained by the first linear layer, which is processed by an activation function on the feature vector, and the output is a 128-dimensional feature vector;

[0130] The input of the third linear layer is a 128-dimensional feature vector obtained by the second linear layer, which is processed by an activation function on the feature vector, and the output is a 1-dimensional ADC activity prediction value;

[0131] Specifically, the antibody conjugate drug activity prediction model of the application is obtained by the following steps:

[0132] (5-1) Obtain a plurality of antibody conjugate drug data pairs to be predicted, preprocess all the antibody conjugate drug data pairs to be predicted to obtain a plurality of preprocessed antibody conjugate drug data pairs, including SMILES sequences of drug carriers and linkers, amino acid sequences of antibody heavy chains, amino acid sequences of antibody light chains, amino acid sequences of antigens, DAR values and label values of antibody conjugate drugs, and divide all the preprocessed antibody conjugate drug data pairs into a training set and a test set according to a ratio of 4:1;

[0133] (5-2) For each of the amino acid sequences of the antibody heavy chains, the amino acid sequences of the antibody light chains, and the amino acid sequences of the antigens in the training set obtained in step (1), input these amino acid sequences into the protein large model ESM-2 to obtain the feature vectors corresponding to each of the amino acid sequences of the antibody heavy chains, each of the amino acid sequences of the antibody light chains, and each of the amino acid sequences of the antigens.

[0134] (5-3) For each of the SMILES sequences of the drug carriers and linkers in the training set obtained in step (1), convert the SMILES sequence into a molecular structure graph, and add virtual nodes and feature encodings in the molecular structure graph to obtain the molecular structure graph corresponding to the SMILES sequence, which includes an M-dimensional atomic feature vector and an N-dimensional chemical bond feature vector, where M is equal to 44 in the present application, and N is equal to 10 in the present application.

[0135] (5-4) For each of the SMILES sequences of the drug carriers and linkers in the training set obtained in step (5-1), use the open source chemical library RDKit to convert the SMILES sequence into a molecular fingerprint feature representation to obtain the Morgan feature vector corresponding to the SMILES sequence of the drug carrier and linker.

[0136] (5-5) For each of the SMILES sequences of the drug carriers and linkers in each of the antibody-conjugated drug data pairs in the training set obtained in step (5-1), input the M-dimensional atomic feature vector and the N-dimensional chemical bond feature vector in the molecular structure graph corresponding to the SMILES sequence obtained in step (5-3) into the mapping layer of the encoder feature extraction module for mapping processing to obtain the 128-dimensional intermediate atomic feature vector of each (i.e., the i-th) atom and the 128-dimensional chemical bond feature vector e ij (0) between the i-th atom and the j-th atom

[0137] (5-6) For each of the SMILES sequences of the drug carriers and linkers in each of the antibody-conjugated drug data pairs in the training set obtained in step (5-1), input the intermediate atomic feature vector of the i-th atom obtained in step (5-5) into the pre-processing layer of the encoder feature extraction module for Laplacian feature position embedding processing to obtain the processed 128-dimensional atomic feature vector

[0138] (5-7) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the attention score corresponding to the i-th atom and the j-th atom obtained in step (5-6) is and the chemical bond feature vector e between the i-th atom and the j-th atom obtained in step (5-5) ij (0) The first transformer layer of the input encoder feature extraction module is inputted to calculate the attention score between the i-th atom and the j-th atom

[0139] The specific calculation formula in this step is:

[0140]

[0141] wherein are the linear transformation matrices of the feature vectors of the i-th atom and the j-th atom in the first transformer processing process, respectively, is the linear transformation matrix of the feature vector of the chemical bond between the i-th atom and the j-th atom in the first transformer layer, d h is the linear transformation matrix of the feature vector of the linear transformation matrix

[0142] (5-8) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the attention score corresponding to the i-th atom and the j-th atom obtained in step (5-7) is The first transformer layer of the input encoder feature extraction module is inputted for message passing processing to obtain the updated intermediate atom feature vector corresponding to the i-th atom and the intermediate chemical bond feature vector corresponding to the i-th atom and the j-th atom

[0143] Specifically, for the message passing of atoms, the current features of adjacent atoms (regarded as the j-th atom) are aggregated, and the feature vector of the i-th atom is updated according to the attention score . The calculation process of this step is specifically:

[0144]

[0145] wherein N i represents a set of adjacent atoms of the i-th atom in the SMILES sequence, is the linear transformation matrix of the adjacent atom feature vector of the i-th atom of the first transformer layer, ω ij is The attention score after activation by the Softmax function.

[0146] For the message passing of chemical bonds, directly add to the intermediate chemical bond feature vector That is,

[0147] (5-9) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the intermediate atom feature vector corresponding to the i-th atom and the intermediate chemical bond feature vector corresponding to the i-th atom and the j-th atom are obtained in step (5-8) are input into the first transformer layer of the encoder feature extraction module to splice the intermediate atom feature vectors and the intermediate chemical bond feature vectors obtained by all attention heads in the first transformer layer, respectively, and add the spliced results to the original atom feature vectors and the chemical bond feature vectors input into the transformer layer to obtain the processed intermediate atom feature vector corresponding to the i-th atom and the intermediate chemical bond feature vector

[0148] (5-10) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the intermediate atom feature vector corresponding to the i-th atom and the intermediate chemical bond feature vector obtained in step (5-9) are respectively subjected to residual sum processing to obtain the processed atom feature vector corresponding to the i-th atom and the chemical bond feature vector corresponding to the i-th atom and the j-th atom

[0149] The calculation process of this step is specifically:

[0150]

[0151] where Norm represents normalization, and represent learnable parameters, and ReLU represents an activation function.

[0152] (5-11) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the atom feature vector corresponding to the i-th atom and the bond feature vector corresponding to the i-th atom and the j-th atom The second transformer layer of the input encoder feature extraction module is processed in the same way as steps (5-7) to (5-10) to obtain the feature vector corresponding to the i-th atom in the final output of the second transformer layer and the bond feature vector corresponding to the i-th atom and the j-th atom Similarly, after the third transformer layer is processed, the atom feature vector corresponding to the i-th atom in the final output of the third transformer layer is obtained and the bond feature vector corresponding to the i-th atom and the j-th atom

[0153] (5-12) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the atom feature vector corresponding to the i-th atom obtained in step (5-11) is The virtual node extraction layer of the input encoder feature extraction module is processed (by atom number index, the feature vector of the virtual node represented by the last atom is found, which is regarded as the feature representation of the entire molecular structure graph) to obtain the 128-dimensional drug carrier feature vector and the 128-dimensional linker feature vector corresponding to the i-th atom.

[0154] The advantages of the above sub-steps (5-5) to (5-12) are that the Graph Transformer is used to extract the features of the drug carrier and the linker, not only considering the neighbor features, but also providing a wider global receptive field, and more effectively extracting the features of the molecular structure graph.

[0155] (5-13) The feature vectors corresponding to each antibody heavy chain amino acid sequence, each antibody light chain amino acid sequence, and each antigen amino acid sequence obtained in step (5-2) and each drug carrier feature vector and each linker feature vector obtained in step (5-12) are input into the graph construction layer of the structure feature extraction module to obtain a virtual ADC graph.

[0156] (5-14) The virtual ADC graph obtained in step (5-13) is input into the linear preprocessing layer of the structure feature extraction module for dimension mapping processing to obtain 128-dimensional node feature vectors and 128-dimensional edge feature vectors.

[0157] (5-15) The 128-dimensional node feature vectors and 128-dimensional edge feature vectors obtained in step (5-14) are input into the three transformer layers of the virtual graph structure feature extraction module based on the graph transformer for feature updating processing to obtain updated 128-dimensional node feature vectors and 128-dimensional edge feature vectors.

[0158] (5-16) The updated node feature vector and edge feature vector obtained in step (5-15) are input into the virtual node extraction layer for lookup operation (by atomic number index, find the feature vector of the virtual node represented by the last atom, and regard it as the feature representation of the entire molecular structure graph), to obtain a 128-dimensional feature vector of the virtual ADC graph.

[0159] (5-17) The DAR value obtained in step (5-1), the feature vector corresponding to the amino acid sequence of each antibody heavy chain, the amino acid sequence of each antibody light chain, and the amino acid sequence of each antigen obtained in step (5-2), and the Morgan feature vector corresponding to the SMILES sequence of each drug carrier and linker obtained in step (5-4) are spliced to obtain a 5889-dimensional feature vector, which is then input into the PCA dimension reduction module for dimension reduction processing to obtain a 128-dimensional feature vector.

[0160] The advantage of the above sub-step (5-17) is that PCA dimension reduction can quickly and effectively extract the most important information from high-dimensional data, preventing too much useful feature information from being missed during dimension reduction.

[0161] (5-18) The feature vector of the virtual ADC graph obtained in step (5-16) and the 128-dimensional feature obtained in step (5-17) are input into the splicing layer of the multi-layer perception module for vector splicing processing to obtain a 256-dimensional fusion feature vector x, which is then input into the multi-layer perception layer for training processing to obtain an activity prediction value p t .

[0162] The specific steps of this calculation process are as follows:

[0163] p t = Linear(ReLU(Linear(ReLU(Linear(x))))

[0164] In the multi-layer perception layer containing three linear layers, the input dimension of the first linear layer is 256, the output dimension is 1024, the input dimension of the second linear layer is 1024, the output dimension is 128, the input dimension of the third linear layer is 128, the output dimension is 1, and x represents the 256-dimensional fusion feature vector.

[0165] (5-19) According to the label value in the antibody-drug conjugate data pair obtained in step (5-1) and the activity prediction value p The loss function is calculated, and the drug target binding affinity prediction model is trained using the loss function until the antibody drug conjugate activity prediction model converges, thereby obtaining a preliminarily trained antibody drug conjugate activity prediction model.

[0166] Specifically, the loss function calculation formula of the step is as follows:

[0167]

[0168] Wherein, α t is a balance factor for adjusting the weight of positive and negative samples, which is set to 0.4 in the method, and γ is a focusing parameter for adjusting the influence of samples difficult to classify, which is set to 3 in the method.

[0169] (5-20) The test set obtained in step (5-1) is used to test the drug target binding affinity prediction model preliminarily trained in step (5-19) until the value of the loss function reaches the optimal value, thereby obtaining a trained drug target binding affinity prediction model.

[0170] Test results

[0171] The test environment of the application: in the Ubuntu 22.04 operating system, the CPU is Intel(R) Xeon(R) Gold 6133, the GPU is 1 block of NVIDIA RTX3090 24GB, and PyTorch 3.7 is used to program the algorithm of the application.

[0172] In order to ensure the reliability of our experimental results, we use five-fold cross-validation in this study. All the data are randomly divided into five parts: one part is used as the test set, and the remaining four parts are used for training. In this way, each experiment will produce five different results. We report the average score as the final performance indicator. This setting ensures that our results are stable and reliable. The learning rate has a significant impact on the performance of the model. In order to improve the training efficiency, we initialize the learning rate to 0.0005, the batch size of the data set is uniformly set to 64, and the training epoch is set to 200.

[0173] According to the test results on the test set in Table 1, it can be seen that the antibody drug conjugate activity prediction method based on graph virtual node proposed in the application is superior to the existing method in terms of Matthews Correlation Coefficient (MCC), Accuracy (ACC), Area Under the ROC Curve (AUC) and F1 score, which are four classification prediction indicators.

[0174] Table 1 Comparison of experimental results

[0175]

[0176] It is to be understood that the above description is intended to be illustrative and not restrictive. Many other embodiments will be apparent to those of skill in the art upon reading the above description. The scope of the application should therefore, be determined not with reference to the above description, but should instead be determined with reference to the appended claims, along with their full scope of equivalents. The disclosure of all articles and references referred to herein are incorporated by reference in their entirety.

Claims

1. A method for predicting antibody-drug conjugate binding based on virtual graphs and multiscale features, characterized in that, The method comprises the following steps: (1) obtaining an antibody-drug conjugate pair to be predicted, preprocessing the antibody-drug conjugate pair to be predicted to obtain a preprocessed antibody-drug conjugate pair, which comprises a simplified molecular linear input canonical (SMILES) sequence of a drug carrier and a linker, an amino acid sequence of an antibody heavy chain, an amino acid sequence of an antibody light chain, an amino acid sequence of an antigen, a drug-antibody ratio (DAR) value, and a label of the antibody-drug conjugate; (2) inputting the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, and the amino acid sequence of the antigen into a protein large model ESM-2 to obtain a feature vector corresponding to each of the amino acid sequences; (3) converting the SMILES sequence into a molecular structure graph, and adding a virtual node and a feature code to the molecular structure graph to obtain a molecular structure graph corresponding to the SMILES sequence; (4) converting the SMILES sequence into a molecular fingerprint feature using an open source chemical library RDKit, as a Morgan feature vector corresponding to the SMILES sequence; (5) inputting the feature vectors corresponding to the amino acid sequences of all antibody heavy chains, the amino acid sequences of all antibody light chains, and the amino acid sequences of all antigens obtained in step (2), the molecular structure graphs corresponding to the SMILES sequences of all drug carriers and linkers obtained in step (3), and the Morgan feature vectors corresponding to the SMILES sequences of all drug carriers and linkers obtained in step (4) into a pre-trained antibody-drug conjugate activity prediction model to obtain an activity prediction result of the antibody-drug conjugate pair; The antibody-drug conjugate activity prediction model comprises a virtual graph structure feature extraction module based on a graph transformer, and the virtual graph structure feature extraction module based on the graph transformer comprises a virtual graph construction layer. The input of the virtual graph construction layer is the 1280-dimensional feature vector corresponding to the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, and the amino acid sequence of the antigen, and the drug carrier feature vector and the linker feature vector output by the graph transformer-based encoder feature extraction module. First, the open source library DGL is used for virtual graph construction operation to obtain a virtual ADC graph G'=(V', E'), where V' represents a set of all nodes in the virtual ADC graph, and E' represents a set of edges connecting the nodes in the virtual ADC graph. Then, a virtual node vn is added to the obtained virtual ADC graph, so that the virtual node is connected to all nodes in the virtual ADC graph and forms a virtual edge. Subsequently, feature coding operation is performed on the virtual ADC graph after adding the virtual node, and the 1280-dimensional feature vectors corresponding to the amino acid sequences of the antibody heavy chain, the antibody light chain, and the antigen, and the 128-dimensional drug feature vectors of the drug carrier and the linker are inserted into the corresponding node positions in the virtual ADC graph to obtain the node feature vectors of the virtual ADC graph. The edges of different types are 0-1 coded, and the DAR value is added to the edge feature between the linker and the antibody heavy chain and light chain to obtain the 7-dimensional edge feature vector of the virtual ADC graph. Finally, the output is the virtual ADC graph corresponding to the antibody conjugated drug and its 7-dimensional edge feature vector. 2.The method of claim 1, wherein the method is characterized by, The pre-processing of the antibody conjugated drug data pair to be predicted in step (1) is specifically as follows: First, the antibody conjugated drug data pair to be predicted is screened to delete data with data missing problems to obtain the SMILES sequence of the drug carrier and the linker, the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, the amino acid sequence of the antigen, the antibody drug ratio, and the biological activity test data including IC50, EC50 and GI50; Then, the label of the antibody conjugated drug is obtained according to the following rules, which are specifically as follows: the biological activity test data including IC50, EC50 and GI50 are uniformly converted to nM units at the cell level, if an antibody conjugated drug has multiple biological activity values, the invalid value is discarded, and the remaining minimum biological activity value is retained as the only biological activity value of the antibody conjugated drug; for the antibody conjugated drug that has been approved or is in clinical I, II and III stages, it is labeled as "positive sample"; for the antibody conjugated drug in the experimental stage, if its IC50, EC50 or GI50 is less than or equal to 100 nM, it is labeled as "positive sample", otherwise, it is labeled as "negative sample"; Finally, the SMILES sequence of the drug carrier and the linker, the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, the amino acid sequence of the antigen, the antibody drug ratio, and the label of the antibody drug are combined to obtain the pre-processed antibody conjugated drug data pair.

3. The antibody-drug conjugate binding prediction method based on virtual graph and multi-scale features according to claim 1 or 2, characterized in that, Step (2) is specifically, first, converting the amino acid sequence of each antibody heavy chain, the amino acid sequence of each antibody light chain, and the amino acid sequence of each antigen into a token vector through the alphabet tool of the esm library, then calling the pretrained.esm2_t33_650M_UR50D model of the esm library as an ESM-2 model, and inputting the token vector corresponding to the amino acid sequence of each antibody heavy chain, the amino acid sequence of each antibody light chain, and the amino acid sequence of each antigen into the ESM-2 model respectively to obtain the feature vector corresponding to the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, and the amino acid sequence of the antigen, respectively, which has a length of 1280 dimensions; Step (3) is specifically, first, using the open source chemical library RDKit to convert the SMILES sequence into a molecular structure graph wherein V represents a set of all atoms in the SMILES sequence, E represents a set of chemical bonds between atoms in the SMILES sequence; then, adding a virtual node vn to the molecular structure graph G, so that the virtual node is connected with all atoms in the molecular structure graph G and forms a virtual edge; subsequently, performing feature coding on the molecular structure graph after adding the virtual node, to obtain an M-dimensional atomic feature vector and an N-dimensional chemical bond feature vector, and the two vectors constitute the molecular structure graph corresponding to the SMILES sequence; wherein M is equal to 44 and N is equal to 10.

4. The antibody-drug conjugate binding prediction method based on virtual graph and multi-scale features according to claim 3, characterized in that, The antibody-drug conjugate activity prediction model further comprises an encoder feature extraction module based on a graph transformer, a principal component analysis (PCA) dimension reduction module, and a multi-layer perception module. The encoder feature extraction module based on the graph transformer comprises a linear preprocessing layer, three transformer layers, and a virtual node extraction layer, and the specific structure is as follows: The input of the preprocessing layer is the M-dimensional atom feature vector and the N-dimensional bond feature vector in the molecular structure graph, which is first subjected to linear regression operation by a linear regression function to obtain an intermediate atom feature vector of 128 dimensions and a bond feature vector of 128 dimensions, respectively; then, the connection relationship between the nodes and edges in the molecular structure graph is used to obtain the unit matrix, the degree matrix, and the adjacency matrix corresponding to the molecular structure graph; then, the symmetric normalized Laplacian matrix is obtained according to the unit matrix, the degree matrix, and the adjacency matrix corresponding to the molecular structure graph, and the Laplacian position feature vector of the symmetric normalized Laplacian matrix is obtained; then, the intermediate atom feature vector obtained after the linear regression operation and the Laplacian position feature vector are added to obtain the processed 128-dimensional atom feature vector, and the final output is the 128-dimensional atom feature vector and the 128-dimensional bond feature vector; The input of the first transformer layer is the 128-dimensional atom feature vector and the 128-dimensional bond feature vector output by the preprocessing layer, which is first subjected to attention mechanism calculation operation by four independent attention heads inside the transformer to obtain the attention score between two atoms; Then, the attention score is used for message passing processing to aggregate the current features of adjacent atoms to obtain a 16-dimensional intermediate atomic feature vector and a 16-dimensional intermediate chemical bond feature vector; then, the intermediate atomic feature vector and the intermediate chemical bond feature vector after the message passing processing are subjected to residual sum processing to obtain a processed 128-dimensional atomic feature vector and a 128-dimensional chemical bond feature vector, and the final output is the 128-dimensional atomic feature vector and the 128-dimensional chemical bond feature vector; The input of the second transformer layer is the 128-dimensional atomic feature vector and the 128-dimensional chemical bond feature vector output by the first transformer layer, which are subjected to attention calculation, message passing and residual sum processing by the attention head in sequence to output a 128-dimensional atomic feature vector and a 128-dimensional chemical bond feature vector, respectively; The input of the third transformer layer is the 128-dimensional atomic feature vector and the 128-dimensional chemical bond feature vector output by the second transformer layer, which are subjected to attention calculation, message passing and residual sum processing by the attention head in sequence to output a 128-dimensional atomic feature vector and a 128-dimensional chemical bond feature vector; The input of the virtual node extraction layer is the 128-dimensional atomic feature vector output by the third transformer layer, which obtains the feature vector of the last atom in the atomic feature vector according to the atomic serial number index, and takes it as the feature vector of the drug carrier and the feature vector of the linker, and outputs a 128-dimensional drug carrier feature vector and a 128-dimensional linker feature vector.

5. The antibody-drug conjugate binding prediction method based on virtual graph and multi-scale features according to claim 4, characterized in that, The virtual graph structure feature extraction module based on the graph transformer further comprises a preprocessing layer, three transformer layers and a virtual node extraction layer; The input of the preprocessing layer is the 1280-dimensional amino acid sequence of the antibody heavy chain, the 1280-dimensional amino acid sequence of the antibody light chain, the 1280-dimensional amino acid sequence of the antigen corresponding to the feature vector, and the virtual ADC graph and its edge feature vector output by the virtual graph construction layer, which first performs linear regression operation on the 1280-dimensional feature vector and the 7-dimensional feature vector by a linear regression function to obtain a 128-dimensional intermediate node feature vector and a 128-dimensional edge feature vector, respectively; then, the connection relationship between the nodes and edges in the virtual ADC graph is used to obtain the unit matrix, the degree matrix and the adjacency matrix corresponding to the virtual ADC graph; then, the symmetric normalized Laplacian matrix is obtained according to the unit matrix, the degree matrix and the adjacency matrix corresponding to the virtual ADC graph, and the Laplacian position feature vector of the symmetric normalized Laplacian matrix is obtained; then, the intermediate node feature vector obtained after the linear regression operation and the Laplacian position feature vector are added to obtain a processed 128-dimensional node feature vector, and the final output is the 128-dimensional node feature vector and the 128-dimensional edge feature vector; The input of the first transformer layer is the 128-dimensional node feature vector and the 128-dimensional edge feature vector output by the preprocessing layer, which first uses four independent attention heads inside the transformer to perform attention mechanism calculation operations on the node feature vector and the edge feature vector in turn to obtain the attention scores between two atoms; Then, the message passing process is performed using the attention scores to aggregate the current features of adjacent atoms to obtain a 16-dimensional intermediate node feature vector and a 16-dimensional intermediate edge feature vector, respectively; then, the intermediate node feature vector and the intermediate edge feature vector after the message passing process are subjected to residual sum processing to obtain the processed 128-dimensional node feature vector and 128-dimensional edge feature vector, and the final output is the 128-dimensional node feature vector and the 128-dimensional edge feature vector; The input of the second transformer layer is the 128-dimensional node feature vector and the 128-dimensional edge feature vector output by the first transformer layer, which uses attention heads to perform attention calculation, message passing and residual sum processing on the node feature vector and the edge feature vector, respectively, and outputs a 128-dimensional node feature vector and a 128-dimensional edge feature vector; The input of the third transformer layer is the 128-dimensional node feature vector and the 128-dimensional edge feature vector output by the second transformer layer, which uses attention heads to perform attention calculation, message passing and residual sum processing on the node feature vector and the edge feature vector, respectively, and outputs a 128-dimensional node feature vector and a 128-dimensional edge feature vector; The input of the virtual node extraction layer is the 128-dimensional node feature vector output by the third transformer layer, which obtains the feature vector of the last node in the node feature vector according to the node sequence index and uses it as the feature vector of the virtual ADC graph, and outputs a 128-dimensional feature vector of the virtual ADC graph.

6. The antibody-drug conjugate binding prediction method based on virtual graph and multi-scale features according to claim 5, characterized in that, The input of the PCA dimension reduction module is the 1280-dimensional amino acid sequence of the antibody heavy chain, the 1280-dimensional amino acid sequence of the antibody light chain, the 1024-dimensional Morgan feature vector corresponding to the feature vector of the antigen amino acid sequence, the 1024-dimensional Morgan feature vector of the drug carrier and the linker, and the 1-dimensional DAR value, which first splices these features into a 5889-dimensional vector, and then performs dimension reduction and linear regression operations on the 5889-dimensional vector to obtain a 128-dimensional feature vector and output; The multi-layer perception module includes a splicing layer and a multi-layer perception layer, and the specific structure is as follows: The input of the splicing layer is the 128-dimensional vector obtained by the virtual graph structure feature extraction module based on the graph transformer and the 128-dimensional vector obtained by the PCA dimension reduction module, which splices the two to output a 256-dimensional feature vector; The multi-layer perception layer includes three linear layers, and the specific structure is as follows: The input of the first linear layer is the 256-dimensional feature vector obtained by the splicing layer, which processes the feature vector through an activation function, and outputs a 1024-dimensional feature vector; The input of the second linear layer is the 256-dimensional feature vector obtained by the first linear layer, which is processed by an activation function, and the output is a 128-dimensional feature vector; The input of the third linear layer is the 128-dimensional feature vector obtained by the second linear layer, which is processed by an activation function, and the output is a 1-dimensional ADC activity prediction value.

7. The method of claim 6, wherein the method is based on a virtual graph and multiscale features. The antibody conjugate drug activity prediction model is obtained by the following steps: (5-1) Obtain a plurality of antibody conjugate drug data pairs to be predicted, preprocess all the antibody conjugate drug data pairs to be predicted to obtain a plurality of preprocessed antibody conjugate drug data pairs, including the SMILES sequence of the drug carrier and the linker, the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, the amino acid sequence of the antigen, the DAR value, and the label value of the antibody conjugate drug, and divide all the preprocessed antibody conjugate drug data pairs into a training set and a test set according to a ratio of 4:1; (5-2) For each antibody heavy chain amino acid sequence, antibody light chain amino acid sequence, and antigen amino acid sequence in the training set obtained in step (1), input these amino acid sequences into the protein large model ESM-2 to obtain the feature vectors corresponding to each antibody heavy chain amino acid sequence, each antibody light chain amino acid sequence, and each antigen amino acid sequence; (5-3) For each SMILES sequence of the drug carrier and the linker in the training set obtained in step (1), convert the SMILES sequence into a molecular structure graph, and add virtual nodes and feature encodings to the molecular structure graph to obtain a molecular structure graph corresponding to the SMILES sequence, which includes an M-dimensional atomic feature vector and an N-dimensional chemical bond feature vector, where M equals 44 and N equals 10; (5-4) For each SMILES sequence of the drug carrier and the linker in the training set obtained in step (5-1), use the open-source chemical library RDKit to convert the SMILES sequence into a molecular fingerprint feature representation to obtain a Morgan feature vector corresponding to the SMILES sequence of the drug carrier and the linker; (5-5) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), input the M-dimensional atomic feature vector and the N-dimensional chemical bond feature vector in the molecular structure graph corresponding to the SMILES sequence obtained in step (5-3) into the mapping layer of the encoder feature extraction module for mapping processing to obtain a 128-dimensional intermediate atomic feature vector of each atom , and a 128-dimensional chemical bond feature vector between the ith atom and the jth atom , where i ∈ [1, the total number of atoms in the molecular structure graph corresponding to the SMILES sequence], j ∈ [1, the total number of atoms in the molecular structure graph corresponding to the SMILES sequence], and i ≠ j; (5-6) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the intermediate atom feature vector of the i th atom obtained in step (5-5) The pre-processing layer of the input encoder feature extraction module is subjected to Laplacian feature position embedding processing to obtain a processed 128-dimensional atom feature vector corresponding to the i th atom ; (5-7) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), input the atomic feature vector corresponding to the i-th atom in the atomic feature vector set obtained in step (5-6) and the chemical bond feature vector between the i-th atom and the j-th atom obtained in step (5-5) into the first transformer layer of the encoder feature extraction module to calculate the attention score between the i-th atom and the j-th atom ; (5-8) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the attention score corresponding to the i th atom and the j th atom in the pair obtained in step (5-7) The first transformer layer of the input encoder feature extraction module is subjected to message passing processing to obtain an updated intermediate atom feature vector corresponding to the i th atom , and an intermediate chemical bond feature vector corresponding to the i th atom and the j th atom ; (5-9) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the intermediate atom feature vector corresponding to the i-th atom and the intermediate bond feature vector corresponding to the i-th atom and the j-th atom obtained in step (5-8) are input into the first transformer layer of the encoder feature extraction module to splice the intermediate atom feature vectors and the intermediate bond feature vectors obtained by all attention heads in the first transformer layer respectively, and add the spliced results to the original atom feature vectors and the bond feature vectors input into the transformer layer respectively to obtain the processed intermediate atom feature vector corresponding to the i-th atom and the intermediate bond feature vector corresponding to the i-th atom and the j-th atom. ;​​​​​ (5-10) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the intermediate atom feature vector corresponding to the i th atom in the intermediate atom feature vector and the intermediate chemical bond feature vector respectively are subjected to residual sum processing to obtain the processed atom feature vector corresponding to the i th atom and the chemical bond feature vector corresponding to the i th atom and the j th atom respectively. (5-11) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the atomic feature vector corresponding to the i-th atom in the atomic feature vector set obtained in step (5-10) and the bond feature vector corresponding to the i-th atom and the j-th atom is input into the second transformer layer of the encoder feature extraction module for the same processing as steps (5-7) to (5-10) above to obtain the feature vector corresponding to the i-th atom in the final output of the second transformer layer and the bond feature vector corresponding to the i-th atom and the j-th atom , and so on, after the processing of the third transformer layer, the atomic feature vector corresponding to the i-th atom in the final output of the third transformer layer is obtained and the bond feature vector corresponding to the i-th atom and the j-th atom ; (5-12) For each drug carrier and linker SMILES sequence in each antibody-drug conjugate data pair in the training set obtained in step (5-1), the atomic feature vector corresponding to the i th atom in the atomic feature vector set obtained in step (5-11) The virtual node extraction layer of the input encoder feature extraction module performs a lookup operation to obtain a 128-dimensional drug carrier feature vector and a 128-dimensional linker feature vector corresponding to the i th atom; (5-13) Input the feature vectors corresponding to each antibody heavy chain amino acid sequence, each antibody light chain amino acid sequence, and each antigen amino acid sequence obtained in step (5-2) and each drug carrier feature vector and each linker feature vector obtained in step (5-12) into the graph construction layer of the structure feature extraction module to obtain a virtual ADC graph; (5-14) Input the virtual ADC graph obtained in step (5-13) into the linear preprocessing layer of the structure feature extraction module for dimension mapping processing to obtain a 128-dimensional node feature vector and a 128-dimensional edge feature vector; (5-15) Input the 128-dimensional node feature vector and the 128-dimensional edge feature vector obtained in step (5-14) into the three transformer layers of the virtual graph structure feature extraction module based on the graph transformer for feature updating processing to obtain updated 128-dimensional node feature vectors and 128-dimensional edge feature vectors; (5-16) The updated node feature vector and edge feature vector obtained in step (5-15) are input into the virtual node extraction layer for lookup operation to obtain a 128-dimensional feature vector of the virtual ADC graph; (5-17) The DAR value obtained in step (5-1), the feature vectors corresponding to the amino acid sequences of each antibody heavy chain, the amino acid sequences of each antibody light chain, and the amino acid sequences of each antigen obtained in step (5-2), and the Morgan feature vectors corresponding to the SMILES sequences of each drug carrier and linker obtained in step (5-4) are spliced to obtain a 5889-dimensional feature vector, and the vector is input into a PCA dimension reduction module for dimension reduction processing to obtain a 128-dimensional feature vector; (5-18) Concatenate the feature vector of the virtual ADC map obtained in step (5-16) and the 128-dimensional feature vector obtained in step (5-17) to obtain a 256-dimensional fusion feature vector and then to the input multi-layer perception module for training processing to obtain an activity prediction value ; (5-19) the label value of the data pair of the antibody conjugated drug obtained in step (5-1) and the activity prediction value obtained in step (5-18) The loss function is calculated, and the drug target binding affinity prediction model is trained using the loss function until the antibody conjugated drug activity prediction model converges, thereby obtaining a preliminarily trained antibody conjugated drug activity prediction model; (5-20) The test set obtained in step (5-1) is tested on the drug target binding affinity prediction model preliminarily trained in step (5-19) until the value of the loss function reaches the optimal value, thereby obtaining the trained drug target binding affinity prediction model.

8. The antibody conjugated drug binding prediction method based on virtual graph and multi-scale features according to claim 7, characterized in that, The specific calculation formula in step (5-7) is: ， wherein , are linear transformation matrices of eigenvectors of the i-th atom and the j-th atom in the first transformer process, respectively, is a linear transformation matrix of eigenvectors of the chemical bond between the i-th atom and the j-th atom in the first transformer layer, is a linear transformation matrix is the dimension of the eigenvectors of the linear transformation matrix Step (5-8) uses the following calculation formula: ， wherein represents a set of adjacent atoms of the i-th atom in the SMILES sequence, is a linear transformation matrix of the adjacent atom feature vector of the i-th atom of the first transformer layer, is is the attention score after activation by the Softmax function; In step (5-8) for the message passing of chemical bonds, directly assign the intermediate chemical bond feature vector i.e. ; Step (5-10) uses the following calculation formula: ， where Norm denotes normalization, and denote learnable parameters, and ReLU denotes an activation function.

9. The antibody conjugated drug binding prediction method based on virtual graph and multi-scale features according to claim 8, characterized in that, Step (5-18) uses the following calculation formula: ， In the three linear layers constituting the multi-layer perceptron module, the input dimension of the first linear layer is 256, the output dimension is 1024, the input dimension of the second linear layer is 1024, the output dimension is 128, and the input dimension of the third linear layer is 128, the output dimension is 1, represents a 256-dimensional fusion feature vector; The loss function calculation formula of step (5-19) is as follows: ， wherein is a balancing factor that regulates the weight of positive and negative samples, is a focusing parameter that regulates the impact of difficult-to-classify samples.

10. An antibody-drug conjugate binding prediction system based on virtual graphs and multi-scale features, which is realized by the antibody-drug conjugate binding prediction method based on virtual graphs and multi-scale features according to claim 1, characterized in that, The antibody conjugated drug binding prediction system comprises: A first module for obtaining an antibody conjugated drug pair to be predicted, and preprocessing the antibody conjugated drug data pair to obtain a preprocessed antibody conjugated drug data pair, which includes the SMILES sequence of the drug carrier and linker, the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, the amino acid sequence of the antigen, the DAR value, and the label of the antibody conjugated drug; A second module for inputting the amino acid sequence of each antibody heavy chain, the amino acid sequence of each antibody light chain, and the amino acid sequence of each antigen in the preprocessed antibody conjugated drug data pair obtained by the first module into the ESM-2 protein model to obtain the feature vectors corresponding to the amino acid sequence of the antibody heavy chain, the amino acid sequence of the antibody light chain, and the amino acid sequence of the antigen, respectively; A third module for converting the SMILES sequence of each drug carrier and linker in the preprocessed antibody conjugated drug data pair obtained by the first module into a molecular structure graph, and adding virtual nodes and feature codes to the molecular structure graph to obtain a molecular structure graph corresponding to the SMILES sequence; The fourth module is configured to, for each SMILES sequence of a drug carrier and a linker in the preprocessed antibody-drug conjugate data pair obtained by the first module, convert the SMILES sequence into a molecular fingerprint feature by using an open-source chemical library RDKit, as a Morgan feature vector corresponding to the SMILES sequence; The fifth module is configured to input the feature vectors corresponding to the amino acid sequences of all antibody heavy chains, the amino acid sequences of all antibody light chains, and the amino acid sequences of all antigens in the antibody-drug conjugate data pair obtained by the second module, the molecular structure graphs corresponding to the SMILES sequences of all drug carriers and linkers obtained by the third module, and the Morgan feature vectors corresponding to the SMILES sequences of all drug carriers and linkers obtained by the fourth module into a pre-trained antibody-drug conjugate activity prediction model, to obtain an activity prediction result of the antibody-drug conjugate data pair.

Images