A drug virtual screening system for newly discovered targets based on zero-shot learning
Through a virtual drug screening system based on zero-sample learning, the graph convolution network and attention mechanism are used to solve the problem of poor predictive performance of newly discovered targets, and accurate drug screening and development guidance for newly discovered targets is achieved.
Patent Information
- Application Number
- CN202210608235.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-31
- Publication Date
- 2024-12-10
- Estimated Expiration
- 2042-05-31
AI Technical Summary
Existing data-driven drug target interaction prediction methods are difficult to work effectively when newly discovered targets are found because the lack of homologous proteins and interaction data leads to poor prediction performance.
Using a virtual drug screening system based on zero-sample learning, the new prediction is predicted by constructing protein isomerographic and drug molecular maps, using graph convolution networks and attention mechanisms, combining prior protein target characterization and pharmacophore extraction module parameters. Discover the probability of interaction between targets and drug molecules.
Accurate drug screening and development guidance for newly discovered targets has been achieved, the problem of lack of homologous proteins has been solved, and the efficiency of drug discovery has been improved.
Smart Images
Figure CN114974409B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of medical data processing, and in particular, relates to a drug virtual screening system for newly discovered targets based on zero-sample learning. Background Art
[0002] Traditional drug target interaction (DTI) is a drug discovery process that requires a large number of wet experiments on drug molecules and target proteins, which is time-consuming and labor-intensive. Therefore, computer-integrated DTI prediction methods play an increasingly important role in pharmaceutical research as the first step to quickly screen candidate molecules. Rule-based in silico DTI methods, such as docking and pharmacophore mapping, often simulate the docking process through chemical structure or energy analysis. At their core, they use chemical rules and similarities between samples to identify key substructures of proteins and ligands (i.e., targets and pharmacophores), respectively. However, since these rules are defined by humans with prior knowledge, these methods are often considered inflexible and computationally expensive. In contrast, data-driven learning methods (such as machine learning, deep learning, etc.) can flexibly and autonomously learn interaction rules, and their acceleration effect on drug discovery is more significant.
[0003] Existing data-driven methods usually model DTI prediction as a binary classification problem for a given drug-target pair, and research mainly focuses on how to effectively learn representations of molecules and proteins, and then input these representations into advanced classification models or ranking models. Representation learning methods can start from a one-dimensional (1D) perspective, such as using molecular fingerprints, simplified molecular input line input system (SMILES), FASTA sequences of target proteins as input features, or use a 3D perspective, using GNN and CNN models to embed the 3D structure of target proteins and molecules. In recent years, how to model 3D structures based on geometric deep learning has attracted a lot of research attention. However, the prediction performance of these deep learning methods essentially depends on the similarity between the target protein and its homologous proteins tested in the training set, either implicitly or explicitly. When a newly discovered target protein is given, it may be difficult to find homologous proteins with similar structures, or there are few known molecular interactions. This makes it difficult for these methods to work well because there is little validated knowledge for these data-driven methods. In this case, the prediction performance of these methods is poor and unsatisfactory.
[0004] The introduction of graph neural networks fits well with the application of drug discovery in the field of deep learning. By representing atoms in chemical molecules as nodes in the graph, atomic bonds in chemistry represent edges in the graph, and introducing features such as bond angles, rotations, and chemical energy into the features of edges and nodes, the original information of chemical molecules is well preserved. Existing work such as MPNN, SchNet, DimNet and other backbone chemical molecular graph neural networks have also performed well on chemical datasets.
[0005] There are two main types of few-shot learning (a general type of meta-learning) methods: (1) gradient-based methods and (2) metric-based methods. The former uses a meta-learner to guide the update of the base learner in order to learn the median initialization of the parameters between all tasks. For gradient-based methods, Finn et al. proposed the MAML algorithm, which adds up the losses of all seen tasks to balance the initialized parameters. Andrychowicz et al. compared the gradient descent function with the LSTM (Long-Short-Term Memory) model and used the LSTM module to guide the base learner to update. For metric-based methods, Snell et al. proposed a prototype network that learns a prototype representation of each class in a metric feature space; in testing, the network calculates the distance between the test sample and all seen classes to perform classification.
[0006] Zero-shot learning is a special case of meta-learning in which a model directly predicts unseen classes with additional information during testing. This approach usually exploits task-level information to quickly adapt to the target task. However, considering that newly discovered proteins may have few homologous proteins and drug target interactions, applying zero-shot learning to the DTI prediction problem is still a less explored direction. There is also a similar concept called the master-slave regularization model; but its purpose is to use the model to directly predict the parameters of the logistic regression model for company revenue prediction. Summary of the invention
[0007] In view of the above-mentioned technical deficiencies, the purpose of the present invention is to provide a drug virtual screening system for newly discovered targets based on zero-sample learning to predict the interaction of newly discovered protein targets with drug molecules.
[0008] To achieve the above-mentioned purpose of the invention, an embodiment provides a drug virtual screening system for newly discovered targets based on zero-sample learning, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, wherein the computer memory has a drug virtual screening prediction model with optimized parameters, and the drug virtual screening prediction model comprises a protein analysis module, a drug molecule virtual screening module, and a parameter storage module, wherein the protein analysis module comprises a protein characterization module, a protein target extraction module, and a target similarity attention module, the drug molecule virtual screening module comprises a drug molecule characterization module, a pharmacophore extraction module, and a classification prediction module, and the parameter storage module stores a priori protein target characterizations corresponding to known proteins, and also stores a priori parameters of the pharmacophore extraction module and a priori parameters of the classification prediction module corresponding to the priori protein targets;
[0009] When the computer processor executes the computer program, the following steps are implemented:
[0010] Construct a protein isomerism graph based on the target protein structure data and initialize node attributes;
[0011] Construct a drug molecule graph based on the molecular structure data of candidate drugs and initialize node attributes;
[0012] The target protein is analyzed by using a protein analysis module, including: the input protein isomerization graph is characterized by a protein characterization module to obtain a protein characterization, and the atomic characterization extracted from the protein characterization is input to a protein target extraction module, and the protein target characterization is output to a target similarity attention module after protein target extraction, and the similarity between the protein target characterization and the prior protein target characterization corresponding to all known proteins is calculated as an attention weight, and then the weighted sum of the attention weight and the prior parameters of all pharmacophore extraction modules and the prior parameters of all classification prediction modules is calculated as the extraction prediction parameter of the pharmacophore extraction module and the classification prediction parameter of the classification prediction module corresponding to the target protein;
[0013] Loading the extraction prediction parameters and the classification prediction parameters into the pharmacophore extraction module and the classification prediction module respectively;
[0014] The candidate drug molecules are screened by using the drug molecule virtual screening module, including: the input drug molecule graph is characterized by the drug molecule characterization module, the drug molecule characterization is output to the pharmacophore extraction module whose parameters are extraction prediction parameters, the pharmacophore characterization is output to the classification prediction module whose parameters are classification prediction parameters after extraction, and the prediction probability of the interaction between the target protein and the candidate drug molecule is output after classification calculation;
[0015] Among them, the predicted probability is used to guide drug screening and development for newly discovered protein targets.
[0016] In one embodiment, constructing a protein isomerogram based on target protein structure data includes: extracting chemical atoms and amino acid residues in the target protein structure data, using chemical atoms and amino acid residues as nodes, chemical bonds between chemical atoms as edges, peptide bonds between amino acid residues as edges, and constructing hyperedges between amino acid residues and their constituent chemical atoms as edges, so as to construct a protein isomerogram.
[0017] In one embodiment, node attributes of a protein isomerization graph are initialized, including: using the electron beam and chirality of chemical atoms as one-hot vectors, using an embedding layer to initialize the one-hot vectors to obtain node attributes corresponding to the chemical atoms; and using a pre-trained natural language model to initialize the amino acid sequence to obtain node attributes corresponding to the amino acid residues.
[0018] In one embodiment, constructing a drug molecule graph based on candidate drug molecular structure data includes: extracting chemical atoms in the candidate drug molecular structure data, using chemical atoms as nodes, and using chemical bonds between chemical atoms as edges to construct a drug molecule graph.
[0019] In one embodiment, the node attributes of the drug molecule graph are initialized, including: using the electron beam and chirality of the chemical atom as the one-hot vector, using the embedding layer to initialize the one-hot vector, and obtaining the node attributes corresponding to the chemical atom.
[0020] In one embodiment, the protein characterization module adopts a graph convolutional network, and uses the graph convolutional network to transfer information to the protein heterogeneous graph with initialized node attributes to update the node attributes and obtain a protein characterization composed of the latest node attributes.
[0021] In one embodiment, the step of extracting the atomic representation from the protein representation to the protein target extraction module includes: extracting the latest node attributes corresponding to the chemical atoms from the protein representation to form the atomic representation, and inputting the atomic representation to the protein target extraction module.
[0022] In one embodiment, the drug molecule characterization module adopts a graph convolutional network, and uses the graph convolutional network to transfer information to the drug molecule graph with initialized node attributes to update the node attributes and obtain the drug molecule characterization composed of the latest node attributes. Preferably, the graph convolutional network adopted by the drug molecule characterization module is SchNet.
[0023] In one embodiment, the protein characterization module uses a graph convolutional network, which includes SchNet, GAT, and TransformerConv. The process of using these three graph convolutional networks to construct protein characterization includes:
[0024] Method 1: Using SchNet to transfer information through chemical bonds between chemical atoms in the protein heterogeneous graph with initialized node attributes, so as to update the node attributes of the nodes corresponding to the chemical atoms;
[0025] Method 2: Using GAT to transfer information through the hyperedges between amino acid residues and their constituent chemical atoms on the protein heterogeneous graph with initialized node attributes, so as to update the node attributes of the nodes corresponding to the chemical atoms and the node attributes of the nodes corresponding to the amino acid residues;
[0026] Method 3: Using TransformerConv to transfer information through peptide bonds between amino acid residues on the protein heterogeneous graph with initialized node attributes, so as to update the node attributes of the nodes corresponding to the amino acid residues;
[0027] The latest node attributes updated by the above three methods are summed or averaged to obtain the protein representation.
[0028] In one embodiment, the protein target extraction module includes a first graph assignment layer and a first average pooling layer; the first graph assignment layer is used to judge whether the node attribute of the input atomic representation is a critical attribute, and the node corresponding to the critical attribute is used as the target; the first average pooling layer is used to average pool the target, and the average pooling result is obtained as the protein target representation; preferably, the first graph assignment layer uses a multilayer perceptron.
[0029] In one embodiment, the pharmacophore extraction module includes a second graph assignment layer and a second average pooling layer; the second graph assignment layer is used to judge whether the node attributes of the input drug molecule representation are critical attributes, and the nodes corresponding to the critical attributes are used as pharmacophoric key points; preferably, the second graph assignment layer adopts a multi-layer perceptron; the second average pooling layer is used to average pool the pharmacophoric key points, and the average pooling result is obtained as the pharmacophore representation.
[0030] In one embodiment, the classification prediction module uses a multi-layer perceptron to perform classification calculations on the input pharmacophore representation to output the predicted probability of the interaction between the target site of the target protein and the pharmacophore of the candidate drug molecule.
[0031] In one embodiment, the drug virtual screening prediction model uses a zero-sample learning-based method to perform parameter optimization, including:
[0032] Setting tasks, including: predicting the interaction of each protein target with an active drug molecule or an inactive drug molecule as a task;
[0033] Constructing samples, including: for each task, obtaining protein structure data of a specific target and corresponding drug molecular structure data and constructing a protein isomerism graph and a drug molecular graph to form a single sample, so as to construct a sample set for each task, and dividing the drug molecular structure data in the sample of each task into a support set and a query set; wherein the drug molecular structure data includes active drug molecular structure data and inactive drug molecular structure data;
[0034] Training a drug molecule virtual screening module comprises: (a1) inputting a drug molecule graph contained in samples in a support set of each task into the drug molecule virtual screening module, and optimizing the parameters of the drug molecule virtual screening module corresponding to each task under the supervision of screening loss learning; (b1) inputting a drug molecule graph contained in samples in a query set of each task into the drug molecule virtual screening module, calculating the screening loss corresponding to each task based on the parameters updated by using the support set, and summing up the screening losses of all tasks, and then updating the parameters of the drug molecule virtual screening module corresponding to each task based on the original parameters by using the total screening loss; (c1) iteratively repeating steps (a1) and (b1) until the accuracy of the first prediction probability average value of the drug molecule virtual screening module for all tasks is stable, and then repeating step (a1), and extracting the parameters of the pharmacophore extraction module and the classification prediction module after the support set of each task is updated and storing them in the parameter storage module as the prior parameters for each protein target;
[0035] The parameters of the drug molecule characterization module in the drug molecule virtual screening module are fixed and the protein analysis module is trained, including: (a2) adding an amino acid residue alignment module in the protein analysis module to align the amino acid residue characterizations extracted from the protein characterizations of the same batch; (b2) inputting the protein isomerization graph contained in the samples of each task into the protein analysis module, and at the same time inputting the drug molecule graph contained in the same sample into the drug molecule virtual screening module to output a second prediction probability, and optimizing the parameters of the protein analysis module corresponding to each task under the supervised learning of the analysis loss until the accuracy of the average value of the second prediction probability for all tasks is stable and unchanged, the parameter optimization is completed, and the protein target characterization corresponding to each task is extracted and stored in the parameter storage module as a priori protein target characterization.
[0036] In one embodiment, the screening loss includes a self-supervised loss constructed based on drug molecule characterization, a connectivity loss and a first mutual information loss constructed based on pharmacophore characterization, and a first classification loss constructed based on a first prediction probability;
[0037] Among them, constructing a self-supervised loss based on drug molecule representation includes: multiplying the drug molecule representations corresponding to any two tasks of the same batch to obtain the bond reconstruction self-supervised loss of atomic bond reconstruction, inputting the drug molecule representation into a multi-layer perceptron to regress and predict the number of atomic electrons, constructing the electron number regression self-supervised loss based on the regression prediction result and the true value of the electron number, and taking the weighted sum of the bond reconstruction self-supervised loss and the electron number regression self-supervised loss as the self-supervised loss;
[0038] Constructing connectivity loss and first mutual information loss based on pharmacophore representation, including: extracting the encoding vector of the node index in the pharmacophore representation, and calculating the transpose of the encoding vector of the node index, the adjacency matrix of the drug molecule graph, and the encoding vector of the node index, and taking the norm of the difference between the product of the three and the unit matrix as the connectivity loss; calculating the mutual information between the pharmacophore representation of the current drug molecule and the drug molecule representation, calculating the average mutual information between the pharmacophore representation of the current drug molecule and the drug molecule representation of other drug molecules in the same batch, and taking the difference between the mutual information and the average mutual information as the first mutual information loss;
[0039] Constructing a first classification loss based on the first prediction probability, including: taking the cross entropy loss of the first prediction probability and the true value label of the interaction of the drug molecule as the first classification loss;
[0040] The weighted sum of self-supervision loss, connectivity loss, first mutual information loss and first classification loss is used as the screening loss.
[0041] In one embodiment, the analysis loss includes an alignment loss constructed based on amino acid residue representation, a connection loss and a second mutual information loss constructed based on protein target representation, a parameter loss constructed based on generation parameters, and a second classification loss constructed based on prediction probability;
[0042] Among them, constructing the alignment loss based on the amino acid residue representation includes: aligning the amino acid residue representation corresponding to a single task with the amino acid residue representation corresponding to other tasks belonging to the same batch, and constructing the alignment loss based on the predicted alignment result;
[0043] Constructing a connection loss and a second mutual information loss based on the protein target representation, including: extracting the encoding vector of the node index in the protein target representation, and calculating the transpose of the encoding vector of the node index, the adjacency matrix of the protein heterogeneous graph, and the encoding vector of the node index, and taking the norm of the difference between the product of the three and the unit matrix as the connection loss; calculating the mutual information between the protein target representation of the current protein and the protein representation, calculating the average mutual information between the protein target representation of the current protein and the protein representation of other proteins in the same batch, and taking the difference between the mutual information and the average mutual information as the second mutual information loss;
[0044] The parameter loss constructed based on the generated parameters includes: calculating the similarity between the protein target representation corresponding to a single task and the protein target representation corresponding to other tasks in the same batch as the attention weight, and weighting and summing the prior parameters corresponding to other tasks according to the attention weight to obtain the prediction parameters for the single task, wherein the prediction parameters include the extraction prediction parameters of the pharmacophore extraction module and the classification prediction parameters of the classification prediction module, and taking the difference between the prediction parameters and the stored prior parameters as the parameter loss;
[0045] Constructing a first classification loss based on the second predicted probability, including: taking the cross entropy loss of the second predicted probability and the true value label of the interaction of the drug molecule as the second classification loss;
[0046] The weighted sum of alignment, connectivity loss, second mutual information loss, and second classification loss is taken as the analysis loss.
[0047] Compared with the prior art, the present invention has the following beneficial effects:
[0048] Through zero-shot learning, a drug virtual screening prediction model is constructed, and the prior protein target representation for the prior protein target, as well as the prior parameters of the corresponding pharmacophore extraction module and the prior parameters of the classification prediction module are stored; in this way, for newly discovered targets, the attention weights of the new target and the prior target are calculated by using the prior protein target representation, and the prediction parameters (extraction prediction parameters and classification prediction parameters) for the newly discovered target are constructed based on the attention weights and prior parameters. In this way, the drug molecule virtual screening module using the prediction parameters can realize the prediction probability of the interaction between the target and the drug molecule, and the activity of the newly discovered target on the drug molecule can be obtained according to the prediction probability. This system can solve the problem of lack of homologous proteins of newly discovered target proteins in practical applications and difficulty in screening active drug molecules, thereby guiding the development of drugs for the newly discovered target proteins. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0050] Figure 1 is a schematic diagram of the structure of a drug virtual screening prediction model provided in an embodiment;
[0051] Figure 2 It is a flowchart of using a drug virtual screening prediction model to perform drug virtual screening of newly discovered targets provided in an embodiment. DETAILED DESCRIPTION
[0052] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.
[0053] In order to solve the problem of lack of homologous proteins of newly discovered target proteins in practical applications and difficulty in screening active drug molecules, the embodiment provides a drug virtual screening prediction model based on zero-sample learning. Figure 1 As shown, the drug virtual screening prediction model provided by the embodiment includes a protein analysis module, a drug molecule virtual screening module and a parameter storage module, wherein the protein analysis module includes a protein characterization module, a protein target extraction module, and a target similarity attention module; the drug molecule virtual screening module includes a drug molecule characterization module, a pharmacophore extraction module, and a classification prediction module; the parameter storage module stores the prior protein target characterization corresponding to the known protein, and also stores the prior parameters of the pharmacophore extraction module and the prior parameters of the classification prediction module corresponding to the prior protein target.
[0054] In an embodiment, a protein characterization module is used to perform feature characterization on an input protein isomerogram to obtain a protein characterization, wherein the protein isomerogram is constructed based on protein structure data, specifically including: extracting chemical atoms and amino acid residues in the protein structure data, using chemical atoms and amino acid residues as nodes, chemical bonds between chemical atoms as edges, peptide bonds between amino acid residues as edges, and constructing a hyperedge as an edge between amino acid residues and their constituent chemical atoms (understood as constituent atoms of the amino acid residues of the chemical origin), thereby constructing a protein isomerogram.
[0055] In the embodiment, before the protein isomerization graph is input into the protein characterization module, the node attributes need to be initialized, including: using the electron beam and chirality of the chemical atom as the one-hot vector, using the embedding layer to initialize the one-hot vector to obtain the node attributes corresponding to the chemical atom; using a pre-trained natural language model (such as the Transformer model ProtTrans) to initialize the amino acid sequence to obtain the node attributes corresponding to the amino acid residues.
[0056] In one possible implementation, the protein characterization module may use a graph convolutional network to transfer information to a protein heterogeneous graph with initialized node attributes, so as to update the node attributes and obtain a protein characterization composed of the latest node attributes.
[0057] In another possible implementation, the graph convolutional networks used in the protein characterization module include SchNet, GAT, and TransformerConv; the process of using these three graph convolutional networks for information transmission to construct protein characterization includes:
[0058] Method 1: SchNet is used to transfer information through chemical bonds between chemical atoms on the protein heterogeneous graph with initialized node attributes, so as to update the node attributes of the nodes corresponding to the chemical atoms; Method 2: GAT is used to transfer information through hyperedges between amino acid residues and their constituent chemical atoms on the protein heterogeneous graph with initialized node attributes, so as to update the node attributes of the nodes corresponding to the chemical atoms and the node attributes of the nodes corresponding to the amino acid residues; Method 3: TransformerConv is used to transfer information through peptide bonds between amino acid residues on the protein heterogeneous graph with initialized node attributes, so as to update the node attributes of the nodes corresponding to the amino acid residues; the latest node attributes updated by the above three methods are summed or averaged to obtain protein representation.
[0059] In an embodiment, a protein target extraction module is used to extract protein targets. Specifically, the latest node attributes corresponding to chemical atoms are extracted from the protein representation to form an atomic representation, and the atomic representation is input into the protein target extraction module. After the protein target extraction calculation, the protein target to target similarity attention module is output.
[0060] In the embodiment, the protein target extraction module includes a first graph assignment layer and a first average pooling layer; the first graph assignment layer is used to judge whether the node attribute of the input atomic representation is a critical attribute, and the node corresponding to the critical attribute is used as the target; then, the first average pooling layer is used to average pool the target, and the average pooling result is obtained as the protein target representation; in the embodiment, the first graph assignment layer preferably can use a multilayer perceptron, that is, the multilayer perceptron is used to judge whether the node attribute of the atomic representation is a critical attribute to output a judgment probability value.
[0061] In an embodiment, a similarity attention module is used to calculate the attention weight according to the protein target and then calculate the corresponding prediction parameters of the protein target, wherein the prediction parameters include the extraction prediction parameters of the pharmacophore extraction module and the classification prediction parameters of the classification prediction module. The specific process includes: the protein target characterization is input into the target similarity attention module, and the similarity between the protein target characterization and the prior protein target characterization corresponding to all known proteins is calculated as the attention weight; then the weighted sum of the attention weight and the prior parameters of all pharmacophore extraction modules is calculated as the extraction prediction parameter of the pharmacophore extraction module corresponding to the protein target, and then the weighted sum of the attention weight and the prior parameters of all classification prediction modules corresponding to the protein target is calculated as the classification prediction parameter of the classification prediction module corresponding to the protein target. Among them, the prior protein target characterization and the prior parameters are obtained through training process.
[0062] In the embodiment, the drug molecule characterization module is used to characterize the input drug molecule graph to obtain drug molecule features. The drug molecule graph is constructed based on the drug molecule structure data, specifically including: extracting chemical atoms in the candidate drug molecule structure data, using chemical atoms as nodes, and using chemical bonds between chemical atoms as edges to construct the drug molecule graph.
[0063] In the embodiment, before the drug molecule graph is input into the drug molecule characterization module, the node attributes of the drug molecule graph are initialized, including: using the electron beam and chirality of the chemical atom as the one-hot vector, using the embedding layer to initialize the one-hot vector, and obtaining the node attributes corresponding to the chemical atom.
[0064] In an embodiment, the drug molecule characterization module can use a 3D backbone molecular graph convolutional network, and use the graph convolutional network to transfer information to the drug molecule graph with initialized node attributes to update the node attributes and obtain a drug molecule characterization composed of the latest node attributes. In a possible implementation, the 3D backbone molecular graph convolutional network used by the drug molecule characterization module is preferably SchNet, and SchNet is used to transfer information to the drug molecule graph with initialized node attributes through chemical bonds between chemical atoms to update the node attributes of the nodes corresponding to the chemical atoms.
[0065] In the embodiment, the pharmacophore extraction module is used to extract the pharmacophore representation according to the drug molecule representation. When extracting, the extraction prediction parameters of the pharmacophore extraction module generated by the protein analysis module are loaded, and the drug molecule representation is calculated under the extraction prediction parameters to extract the pharmacophore representation that plays a major pharmacodynamic role.
[0066] In the embodiment, the pharmacophore extraction module includes a second graph assignment layer and a second average pooling layer; the second graph assignment layer is used to judge whether the node attribute of the input drug molecule representation is a critical attribute, and the node corresponding to the critical attribute is judged as the pharmacophoric key point; then, the second average pooling layer is used to average pool the pharmacophoric key points, and the average pooling result is obtained as the pharmacophoric representation. In the embodiment, the second graph assignment layer preferably adopts a multilayer perceptron, that is, the multilayer perceptron is used to judge whether the node attribute of the drug molecule representation is a pharmacophoric key point, so as to output a judgment probability value.
[0067] In the embodiment, the classification prediction module is used to predict whether it is an active drug molecule according to the pharmacophore characterization. When it is predicted to be an active drug molecule, it indicates that the protein target interacts with the drug molecule. When it is predicted to be an inactive drug molecule, it indicates that the protein target does not interact with the drug molecule. When making a specific prediction, the classification prediction parameters of the classification prediction module generated by the protein analysis module are loaded, and the pharmacophore characterization is calculated under the classification prediction parameters to output the predicted probability of the interaction between the protein target and the drug molecule.
[0068] In an embodiment, the classification prediction module may use a multi-layer perceptron to perform classification calculations on the input pharmacophore representation to output the predicted probability of the pharmacophore interaction between the protein target and the drug molecule.
[0069] The above drug virtual screening prediction model needs to be parameter optimized before being applied. The embodiment uses zero-sample learning to optimize the model parameters to update the model parameters and generate prior parameters of the pharmacophore extraction module corresponding to the prior protein target and the classification prediction module. The specific parameter optimization process includes:
[0070] 1. Set tasks, including: predicting the interaction of each protein target with active drug molecules or inactive drug molecules as a task. During the learning process, N proteins will be input in the same batch, and there will be N tasks corresponding to N protein targets.
[0071] 2. Constructing samples, including: for each task, obtaining protein structure data of specific targets and corresponding drug molecular structure data and constructing protein isomerization graphs and drug molecular graphs to form a single sample, so as to construct a sample set for each task, and dividing the drug molecular structure data in the sample of each task into a support set and a query set; wherein the drug molecular structure data includes active drug molecular structure data and inactive drug molecular structure data.
[0072] In the embodiment, the protein structure data includes the number of electrons, atomic chirality (UNSPECIFIED, CHI_TETRAHEDRAL_CW, CHI_TETRAHEDRAL_CCW, CHI_OTHER), bond type (single bond, double bond, triple bond, benzene ring) and bond direction (NONE, ENDUPRIGHT, ENDDOWNRIGHT). Based on these protein structure data, a protein isomerization graph is constructed in the above manner and node attributes are initialized.
[0073] In the embodiment, for the drug molecular structure data, 5 active molecules and 5 inactive molecules can be randomly selected as the support set for zero-shot learning, and then 125 active molecules and 125 inactive molecules can be randomly selected as the query set for zero-shot learning, and whether the drug molecule reacts with the protein of the task is used as the final label, and the label is 0, 1, indicating that the drug molecule is an inactive molecule and an active molecule for the protein, respectively;
[0074] 3. Training of drug molecule virtual screening module, including:
[0075] (a1) The drug molecule graph contained in the samples in the support set of each task is input into the drug molecule virtual screening module, that is, the drug molecule representation is obtained by using the drug molecule characterization module, and the drug molecule representations corresponding to any two tasks in the same batch are multiplied to obtain the bond reconstruction self-supervised loss of atomic bond reconstruction, and the drug molecule representation is input into the multi-layer perceptron to regress and predict the number of atomic electrons, and the electron number regression self-supervised loss is constructed based on the regression prediction result and the true value of the electron number, and the weighted sum of the bond reconstruction self-supervised loss and the electron number regression self-supervised loss is used as the self-supervised loss; the drug molecule representation is input into the pharmacophore extraction module, and after extracting the pharmacophore representation, the encoding vector of the node index in the pharmacophore representation is extracted, and the transpose of the encoding vector of the node index, the adjacency matrix of the drug molecule graph, and the encoding vector of the node index are calculated, and the range of the difference between the three products and the unit matrix is calculated. number as connectivity loss; calculate the mutual information between the pharmacophore representation of the current drug molecule and the drug molecule representation, calculate the average mutual information between the pharmacophore representation of the current drug molecule and the drug molecule representation of other drug molecules in the same batch, and take the difference between the mutual information and the average mutual information as the first mutual information loss; average pool the pharmacophore key points corresponding to the critical attributes to obtain the pharmacophore representation and input it into the classification prediction module, output the first prediction probability after calculation, take the cross entropy loss of the first prediction probability and the true value label of the interaction between the drug molecule (that is, whether the drug molecule is an active drug molecule or an inactive drug molecule) as the first classification loss, and take the weighted sum of the self-supervised loss, connectivity loss, first mutual information loss and first classification loss as the screening loss; optimize the parameters of the drug molecule virtual screening module corresponding to each task under the supervised learning of the screening loss;
[0076] (b1) The drug molecule graph contained in the samples in the query set of each task is input into the drug molecule virtual screening module, and the screening loss corresponding to each task is calculated based on the parameters updated by the support set. The screening loss is calculated in the same way as in step (a1). After summing the screening losses of all tasks, the total screening loss is used to update the parameters of the drug molecule virtual screening module corresponding to each task based on the original parameters. It should be noted that the original parameters refer to the initial parameters of the drug molecule virtual screening module.
[0077] (c1) Iteratively repeat steps (a1) and (b1) until the accuracy of the first prediction probability average value of the drug molecule virtual screening module for all tasks remains stable, that is, until the first prediction probability average value on all tasks no longer increases, and then repeat step (a1), that is, use the drug molecule graph in the support set to continue updating the parameters updated by the query set, and extract the updated parameters of the pharmacophore extraction module and the classification prediction module of each task in the support set and store them in the parameter storage module as the prior parameters of the pharmacophore extraction module and the classification prediction module for each protein target. It should be noted that there is a set of prior parameters for each task, that is, for each protein target. Figure 1 Three drug molecule virtual screening modules with different parameters corresponding to three protein data Ta, Tb, and Tc are given as examples.
[0078] 4. After completing the training of the drug molecule virtual screening module, the parameters of the drug molecule characterization module in the drug molecule virtual screening module are fixed and the protein analysis module is trained, including:
[0079] (a2) Figure 1 As shown, an amino acid residue alignment module is added to the protein analysis module to align the amino acid residue representations extracted from the protein representations of the same batch;
[0080] (b2) Inputting the protein isomerization graph contained in the samples of each task into the protein analysis module, obtaining the protein representation after feature characterization by the protein characterization module, and extracting the atomic representation and amino acid residue representation from the protein representation. Using the multi-sequence alignment technology, the amino acid residue representation corresponding to a single task is aligned with the amino acid residue representation corresponding to other tasks in the same batch, that is, the amino acid residue representation corresponding to a single task is multiplied with the amino acid residue representation corresponding to other tasks in the same batch to obtain the predicted alignment result, and constructing the alignment loss based on the predicted alignment result for self-supervision and constraining the training results; inputting the atomic representation into the protein target extraction module to extract the protein target After characterization, extract the encoding vector of the node index in the protein target representation, and calculate the transpose of the encoding vector of the node index, the adjacency matrix of the protein heterogeneous graph, and the encoding vector of the node index. Then, take the norm of the difference between the three products and the unit matrix as the connection loss; calculate the mutual information between the protein target representation of the current protein and the protein representation, calculate the average mutual information between the protein target representation of the current protein and the protein representation of other proteins in the same batch, and take the difference between the mutual information and the average mutual information as the second mutual information loss; perform average pooling on the targets judged to correspond to the key attributes, obtain the protein target representation and input it into the target similarity attention module, and calculate the loss of a single task. The similarity between the protein target representation corresponding to the task and the protein target representation corresponding to other tasks in the same batch is used as the attention weight, and the prior parameters corresponding to other tasks stored are weighted and summed according to the attention weight to obtain the prediction parameters for the current single task, wherein the prediction parameters include the extraction prediction parameters of the pharmacophore extraction module and the classification prediction parameters of the classification prediction module, and the difference between the prediction parameters and the stored prior parameters is used as the parameter loss; the prediction parameters are loaded into the pharmacophore extraction module and the classification prediction module; the drug molecule graph contained in the same sample is input into the drug molecule virtual screening module, and successively passes through the drug molecule characterization module, the pharmacophore extraction module with the parameters as the extraction prediction parameters, and the classification prediction module. Take the module, the parameters of the classification prediction module are the calculation of the classification prediction parameters to output the second prediction probability, and the cross entropy loss of the second prediction probability and the true value label of the drug molecule interaction is taken as the second classification loss; the weighted sum of the alignment, connectivity loss, second mutual information loss and second classification loss is taken as the analysis loss; under the supervised learning of the analysis loss, the parameters of the protein analysis module corresponding to each task are optimized until the accuracy of the second prediction probability average value for all tasks is stable, that is, the second prediction probability average value for all tasks reaches the optimal value, the parameter optimization is completed, and the protein target representation corresponding to each task is extracted and stored in the parameter storage module as the prior protein target representation.
[0081] Compared with the existing technology, for newly discovered protein targets, there is no need for related homologous protein structure analysis and knowledge transfer, nor is there a need for a large number of related active and inactive molecules to be trained. It can directly generate a drug virtual screening prediction model with a high accuracy. This model can realize the prediction of drug activity for this newly discovered protein target and guide drug development.
[0082] Based on the parameter-optimized drug virtual screening prediction model, the embodiment also provides a drug virtual screening system for newly discovered targets based on zero-sample learning, which is a hardware device with the function of drug virtual screening for newly discovered targets. Specifically, the system includes a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, and also stores a drug virtual screening prediction model, wherein the drug virtual screening prediction model includes a parameter storage module that stores a priori protein target representations corresponding to known proteins, and also stores a priori parameters of a pharmacophore extraction module corresponding to a priori protein targets and a priori parameters of a classification prediction module.
[0083] When the computer processor executes the computer program, a virtual drug screening process for newly discovered targets is implemented, such as Figure 2 As shown, the following steps are included:
[0084] Step 1: construct a protein isomorphic graph based on the target protein structure data and initialize node attributes.
[0085] In the embodiment, the target protein includes a newly discovered protein target. The target protein structure data also includes the number of atomic electrons, atomic chirality, and bond direction, etc. The protein isomerization graph of the target protein structure data is constructed in the above manner and the node attributes are initialized, which will not be repeated here.
[0086] Step 2: construct a drug molecule graph based on the candidate drug molecular structure data and initialize node attributes.
[0087] In the embodiment, the candidate drug molecular structure data may be active drug molecular data for newly discovered protein targets, or may be inactive drug molecular data. The above method is used to construct a drug molecular graph of the candidate drug molecular structure data and initialize node attributes, which will not be repeated here.
[0088] Step 3: Analyze the target protein using the protein analysis module to determine the extraction prediction parameters and classification prediction parameters for the target protein target.
[0089] In an embodiment, a target protein is analyzed using a protein analysis module, including: an input protein isomorphic graph is subjected to feature characterization by a protein characterization module to obtain a protein characterization, and an atomic characterization extracted from the protein characterization is input to a protein target extraction module, and the protein target characterization is output to a target similarity attention module after protein target extraction, and the similarity between the protein target characterization and a priori protein target characterizations corresponding to all known proteins is calculated as an attention weight, and then the attention weight is calculated and the weighted sum of the stored prior parameters of all pharmacophore extraction modules and the prior parameters of all classification prediction modules is used as the extraction prediction parameter of the pharmacophore extraction module and the classification prediction parameter of the classification prediction module corresponding to the target protein.
[0090] Step 4: Load the extraction prediction parameters and classification prediction parameters into the pharmacophore extraction module and the classification prediction module respectively.
[0091] Step 5: Screen candidate drug molecules using the drug molecule virtual screening module.
[0092] In an embodiment, a drug molecule virtual screening module is used to screen candidate drug molecules, including: inputting a drug molecule graph through feature characterization by a drug molecule characterization module with fixed parameters, outputting the drug molecule characterization to a pharmacophore extraction module with extraction prediction parameters, outputting the pharmacophore characterization through extraction to a classification prediction module with classification prediction parameters, and outputting a predicted probability of interaction between a target protein and a candidate drug molecule through classification calculation.
[0093] In the embodiment, the predicted probability of each candidate drug molecule is used as the activity probability for detecting drug activity, which is used to guide the rapid screening and development of drugs for target proteins. At the same time, according to the protein target extraction module, a protein target characterization map can be obtained, which can be used as a structural reference for the protein target. At the same time, the pharmacophore characterization map extracted by the pharmacophore extraction module is a substructure of the commonality of active drugs, which can be used as a structural reference for the corresponding pharmacophore and can be used to guide the interpretation and exploration of subsequent drug development work.
[0094] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A drug virtual screening system for newly discovered targets based on zero-sample learning, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, characterized in that: The computer memory has a drug virtual screening prediction model with optimized parameters, and the drug virtual screening prediction model includes a protein analysis module, a drug molecule virtual screening module and a parameter storage module, wherein the protein analysis module includes a protein characterization module, a protein target extraction module, and a target similarity attention module, the drug molecule virtual screening module includes a drug molecule characterization module, a pharmacophore extraction module, and a classification prediction module, and the parameter storage module stores a priori protein target characterization corresponding to a known protein, and also stores a priori parameters of the pharmacophore extraction module and a priori parameters of the classification prediction module corresponding to the priori protein target; When the computer processor executes the computer program, the following steps are implemented: Construct a protein isomerism graph based on the target protein structure data and initialize node attributes; Construct a drug molecule graph based on the molecular structure data of candidate drugs and initialize node attributes; The target protein is analyzed by using a protein analysis module, including: the input protein isomerization graph is characterized by a protein characterization module to obtain a protein characterization, and the atomic characterization extracted from the protein characterization is input to a protein target extraction module, and the protein target characterization is output to a target similarity attention module after protein target extraction, and the similarity between the protein target characterization and the prior protein target characterization corresponding to all known proteins is calculated as an attention weight, and then the weighted sum of the attention weight and the prior parameters of all pharmacophore extraction modules and the prior parameters of all classification prediction modules is calculated as the extraction prediction parameter of the pharmacophore extraction module and the classification prediction parameter of the classification prediction module corresponding to the target protein; Loading the extraction prediction parameters and the classification prediction parameters into the pharmacophore extraction module and the classification prediction module respectively; The candidate drug molecules are screened by using the drug molecule virtual screening module, including: the input drug molecule graph is characterized by the drug molecule characterization module, the drug molecule characterization is output to the pharmacophore extraction module whose parameters are extraction prediction parameters, the pharmacophore characterization is output to the classification prediction module whose parameters are classification prediction parameters after extraction, and the prediction probability of the interaction between the target protein and the candidate drug molecule is output after classification calculation; Among them, the predicted probability is used to guide drug screening and development for newly discovered protein targets.
2. The drug virtual screening system for newly discovered targets based on zero-sample learning according to claim 1, characterized in that: The method of constructing a protein isomerism graph based on the target protein structure data includes: extracting chemical atoms and amino acid residues in the target protein structure data, using chemical atoms and amino acid residues as nodes, using chemical bonds between chemical atoms as edges, using peptide bonds between amino acid residues as edges, and constructing a hyperedge between the amino acid residues and their constituent chemical atoms as an edge, thereby constructing a protein isomerism graph; Initialize the node attributes of the protein heterogeneous graph, including: using the electron beam and chirality of the chemical atom as the one-hot vector, using the embedding layer to initialize the one-hot vector to obtain the node attributes corresponding to the chemical atom; using the pre-trained natural language model to initialize the amino acid sequence to obtain the node attributes corresponding to the amino acid residue; The method of constructing a drug molecule graph based on the candidate drug molecular structure data includes: extracting chemical atoms in the candidate drug molecular structure data, using the chemical atoms as nodes, and using chemical bonds between the chemical atoms as edges to construct a drug molecule graph; Initializing the node attributes of the drug molecule graph includes: using the electron beam and chirality of the chemical atom as the one-hot vector, using the embedding layer to initialize the one-hot vector, and obtaining the node attributes corresponding to the chemical atom.
3. The drug virtual screening system for newly discovered targets based on zero-sample learning according to claim 2, characterized in that: The protein characterization module adopts a graph convolutional network, and uses the graph convolutional network to transmit information to the protein heterogeneous graph with initialized node attributes to update the node attributes and obtain a protein characterization composed of the latest node attributes; The atomic representation extracted from the protein representation to the protein target extraction module includes: extracting the latest node attributes corresponding to the chemical atoms from the protein representation to form the atomic representation, and inputting the atomic representation to the protein target extraction module; The drug molecule characterization module adopts a graph convolutional network, and uses the graph convolutional network to transfer information to the drug molecule graph with initialized node attributes to update the node attributes and obtain the drug molecule characterization composed of the latest node attributes.
4. The drug virtual screening system for newly discovered targets based on zero-sample learning according to claim 3, characterized in that: The graph convolutional network used in the drug molecule characterization module is SchNet.
5. The drug virtual screening system for newly discovered targets based on zero-sample learning according to claim 2, characterized in that: The protein characterization module adopts a graph convolutional network, which includes SchNet, GAT, and TransformerConv. The process of constructing protein characterization using these three graph convolutional networks includes: Method 1: Use SchNet to transfer information through chemical bonds between chemical atoms in the protein heterogeneous graph with initialized node attributes, so as to update the node attributes of the nodes corresponding to the chemical atoms; Method 2: Using GAT to transfer information through the hyperedges between amino acid residues and their constituent chemical atoms on the protein heterogeneous graph with initialized node attributes, so as to update the node attributes of the nodes corresponding to the chemical atoms and the node attributes of the nodes corresponding to the amino acid residues; Method 3: Using TransformerConv to transfer information through peptide bonds between amino acid residues on the protein heterogeneous graph with initialized node attributes, so as to update the node attributes of the nodes corresponding to the amino acid residues; The latest node attributes updated by the above three methods are summed or averaged to obtain protein representation; The atomic representation extracted from the protein representation is transferred to the protein target extraction module, including: extracting the latest node attributes corresponding to the chemical atoms from the protein representation to form the atomic representation, and inputting the atomic representation to the protein target extraction module.
6. The drug virtual screening system for newly discovered targets based on zero-sample learning according to claim 1, characterized in that: The protein target extraction module includes a first graph assignment layer and a first average pooling layer; the first graph assignment layer is used to judge whether the node attribute of the input atomic representation is a critical attribute, and the node judged to correspond to the critical attribute is used as a target; The first average pooling layer is used to average pool the target, and the average pooling result is used as the protein target representation; The pharmacophore extraction module includes a second graph assignment layer and a second average pooling layer; the second graph assignment layer is used to judge whether the node attribute of the input drug molecule representation is a key attribute, and the node corresponding to the key attribute is judged as the key point of drug efficacy; The second average pooling layer is used to average pool the pharmacophore key points, and the average pooling result is used as the pharmacophore representation.
7. The drug virtual screening system for newly discovered targets based on zero-sample learning according to claim 6, characterized in that: The first graph assignment layer uses a multi-layer perceptron; the second graph assignment layer uses a multi-layer perceptron.
8. The drug virtual screening system for newly discovered targets based on zero-sample learning according to claim 1, characterized in that: The classification prediction module adopts a multi-layer perceptron, and uses the multi-layer perceptron to perform classification calculations on the input pharmacophore representation to output the predicted probability of the interaction between the target point of the target protein and the pharmacophore of the candidate drug molecule.
9. The drug virtual screening system for newly discovered targets based on zero-sample learning according to claim 1, characterized in that: The drug virtual screening prediction model uses a zero-sample learning-based method to perform parameter optimization, including: Setting tasks, including: predicting the interaction of each protein target with an active drug molecule or an inactive drug molecule as a task; Constructing samples, including: for each task, obtaining protein structure data of a specific target and corresponding drug molecular structure data and constructing a protein isomerism graph and a drug molecular graph to form a single sample, so as to construct a sample set for each task, and dividing the drug molecular structure data in the sample of each task into a support set and a query set; wherein the drug molecular structure data includes active drug molecular structure data and inactive drug molecular structure data; Training a drug molecule virtual screening module includes: (a1) inputting a drug molecule graph contained in samples in a support set of each task into the drug molecule virtual screening module, and optimizing the parameters of the drug molecule virtual screening module corresponding to each task under supervised learning of screening loss; (b1) inputting a drug molecule graph contained in samples in a query set of each task into the drug molecule virtual screening module, calculating the screening loss corresponding to each task based on the parameters updated using the support set, and summing up the screening losses of all tasks, and then updating the parameters of the drug molecule virtual screening module corresponding to each task based on the original parameters using the total screening loss; (c1) iteratively repeating steps (a1) and (b1) until the accuracy of the first prediction probability average value of the drug molecule virtual screening module for all tasks becomes stable, and then repeating step (a1), and extracting the parameters of the pharmacophore extraction module and the classification prediction module after the support set of each task is updated and stored in the parameter storage module as the prior parameters for each protein target; The parameters of the drug molecule characterization module in the drug molecule virtual screening module are fixed and the protein analysis module is trained, including: (a2) adding an amino acid residue alignment module to the protein analysis module to align the amino acid residue characterizations extracted from the protein characterizations of the same batch; (b2) inputting the protein isomerization graph contained in the samples of each task into the protein analysis module, and at the same time inputting the drug molecule graph contained in the same sample into the drug molecule virtual screening module to output a second prediction probability, and optimizing the parameters of the protein analysis module corresponding to each task under the supervised learning of the analysis loss until the accuracy of the average value of the second prediction probability for all tasks remains stable and unchanged, the parameter optimization is completed, and the protein target characterization corresponding to each task is extracted and stored in the parameter storage module as a priori protein target characterization.
10. The drug virtual screening system for newly discovered targets based on zero-shot learning according to claim 1, characterized in that: The screening loss includes a self-supervised loss constructed based on drug molecule characterization, a connectivity loss and a first mutual information loss constructed based on pharmacophore characterization, and a first classification loss constructed based on a first prediction probability; Among them, constructing a self-supervised loss based on drug molecule representation includes: multiplying the drug molecule representations corresponding to any two tasks of the same batch to obtain the bond reconstruction self-supervised loss of atomic bond reconstruction, inputting the drug molecule representation into a multi-layer perceptron to regress and predict the number of atomic electrons, constructing the electron number regression self-supervised loss based on the regression prediction result and the true value of the electron number, and taking the weighted sum of the bond reconstruction self-supervised loss and the electron number regression self-supervised loss as the self-supervised loss; Constructing connectivity loss and first mutual information loss based on pharmacophore representation, including: extracting the encoding vector of the node index in the pharmacophore representation, and calculating the transpose of the encoding vector of the node index, the adjacency matrix of the drug molecule graph, and the encoding vector of the node index, and taking the norm of the difference between the product of the three and the unit matrix as the connectivity loss; calculating the mutual information between the pharmacophore representation of the current drug molecule and the drug molecule representation, calculating the average mutual information between the pharmacophore representation of the current drug molecule and the drug molecule representation of other drug molecules in the same batch, and taking the difference between the mutual information and the average mutual information as the first mutual information loss; Constructing a first classification loss based on the first prediction probability, including: taking the cross entropy loss of the first prediction probability and the true value label of the interaction of the drug molecule as the first classification loss; The weighted sum of self-supervision loss, connectivity loss, first mutual information loss and first classification loss is used as the screening loss.
11. The drug virtual screening system for newly discovered targets based on zero-shot learning according to claim 1, characterized in that: The analysis loss includes an alignment loss constructed based on amino acid residue representation, a connection loss and a second mutual information loss constructed based on protein target representation, a parameter loss constructed based on generation parameters, and a second classification loss constructed based on prediction probability; Among them, constructing the alignment loss based on the amino acid residue representation includes: aligning the amino acid residue representation corresponding to a single task with the amino acid residue representation corresponding to other tasks belonging to the same batch, and constructing the alignment loss based on the predicted alignment result; Constructing a connection loss and a second mutual information loss based on the protein target representation, including: extracting the encoding vector of the node index in the protein target representation, and calculating the transpose of the encoding vector of the node index, the adjacency matrix of the protein heterogeneous graph, and the encoding vector of the node index, and taking the norm of the difference between the product of the three and the unit matrix as the connection loss; calculating the mutual information between the protein target representation of the current protein and the protein representation, calculating the average mutual information between the protein target representation of the current protein and the protein representation of other proteins in the same batch, and taking the difference between the mutual information and the average mutual information as the second mutual information loss; The parameter loss constructed based on the generated parameters includes: calculating the similarity between the protein target representation corresponding to a single task and the protein target representation corresponding to other tasks in the same batch as the attention weight, and weighting and summing the prior parameters corresponding to other tasks according to the attention weight to obtain the prediction parameters for the single task, wherein the prediction parameters include the extraction prediction parameters of the pharmacophore extraction module and the classification prediction parameters of the classification prediction module, and taking the difference between the prediction parameters and the stored prior parameters as the parameter loss; Constructing a first classification loss based on the second predicted probability, including: taking the cross entropy loss of the second predicted probability and the true value label of the interaction of the drug molecule as the second classification loss; The weighted sum of alignment, connectivity loss, second mutual information loss, and second classification loss is taken as the analysis loss.