AD drug affinity prediction method based on data enhancement and graph neural network

Abstract

The invention provides an AD drug affinity prediction method based on data enhancement and a graph neural network, and the method comprises the steps: obtaining a drug sample database by looking up a target protein of an AD target, hybridizing the drug sample database with target protein sequence information to form input data, and then arranging the input data into an input data table; the method comprises the following steps: preprocessing data, performing data enhancement on a drug sample, extracting expression of a drug and a target sequence, constructing a drug molecular map, calculating a protein contact matrix, constructing a protein residue sub-map, and splicing and enhancing the protein residue sub-map; constructing a joint graph, enhancing the connectivity of the joint graph, and constructing full connectivity and root node self-loop by using node index offset; carrying out characterization coding and node initialization; and inputting the joint graph feature vector and the node feature into a model for reasoning and regression prediction, and finally outputting affinity and writing the affinity into a collection table. According to the method, the prediction process is optimized, and the robustness and accuracy of the AD target drug affinity prediction process are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to an improved affinity prediction technology for Alzheimer's disease (AD) drug development, which optimizes the prediction performance of glutamatergic system targets (such as NMDA receptor, metabotropic glutamate receptor) by fusing graph neural network and reinforcement learning, and uses small sample data to enhance the prediction of drug and target affinity. It is mainly used for protein representation, drug screening, discovery and design. BACKGROUND

[0002] The role of glutamate in the pathophysiology of Alzheimer's disease (AD) is an important research area. Glutamate is an excitatory neurotransmitter, and its abnormal concentration and function in the brain are closely related to various nervous system diseases, including AD. We will integrate AD-related transcriptomic data and drug databases to extract key features and predict the relationship between drug and gene expression.

[0003] Drug-target affinity prediction applications include drug discovery and development and disease treatment. In practical applications, neural networks are often used to predict the affinity between drugs and targets. Common deep learning-based methods, such as convolutional neural networks, are widely used in drug-target affinity prediction. These models usually convert three-dimensional compounds into one-dimensional sequences to extract basic feature information of binding affinity. However, one-dimensional sequences lack sufficient structural information, limiting prediction accuracy. At the same time, existing models have some problems, such as data scarcity: AD-related target available complex structures only account for 2.3% of PDB; evaluation defects (traditional indicators cannot capture neuroprotective effects) and so on.

[0004] Some recent methods use the structure of a graph to represent compounds and proteins. In compounds, drug atoms are treated as nodes, while chemical bonds are edges. In proteins, residues are treated as nodes, while contact maps define topology, which is a matrix with distances between residues. Compared with one-dimensional sequence input, graph-based models can capture more information from drugs and proteins, including atom types, bonds between atoms, and topological structures, thereby improving drug-target affinity prediction accuracy. These methods include representing compounds as graph structures, using atoms as nodes, and chemical bonds as edges; representing proteins as graph structures, using residues as nodes, and contact maps defining topology. These methods can capture more structural information, thereby improving drug-target affinity prediction accuracy.

[0005] The structure of the graph can retain more structural information and increase the diversity of drug and protein features, thereby facilitating the development of various graph neural network-based methods to capture these features. However, these methods, although making progress, still cannot fully represent the physical and chemical properties of proteins. After changing the structure of the graph, they also lack appropriate ways of information transmission. Therefore, accurately representing the topology close to the binding pocket of the drug and target is crucial for improving DTA (Drug Target Affinity, drug and target affinity) prediction. SUMMARY

[0006] Drug-target affinity prediction is an important part of drug discovery, and traditional methods rely on experimental determination and are costly. In recent years, deep learning-based architectures such as sequence models and graph neural networks have made progress in protein-ligand interaction prediction. However, existing methods still have problems such as poor robustness in handling heterogeneous data fusion, ensuring graph integrity, and sequence-graph interaction, and insufficient cross-target generalization ability. The present invention makes some improvements based on the unified DMFF architecture (see Figure 1 ), combining GINConv and bidirectional LSTM attention modules to achieve efficient and stable affinity prediction.

[0007] The technical solution of the present invention is: this AD drug target affinity prediction method based on data augmentation and graph neural networks includes the following steps: S1 consults the target proteins of AD targets, obtains a drug sample database, and hybridizes the target protein sequence information into input data and arranges them into an input data table; S2 pre-processes the data, performs data augmentation on the drug samples, extracts the representation of the drug and target sequences, constructs a drug molecule graph, calculates the protein contact matrix, constructs a protein residue subgraph, and performs splicing and augmentation; S3 construct a joint graph and enhance its connectivity, use node index offset, construct full connectivity and root node self-loop; S4 feature coding and node initialization; S5 input the joint graph feature vector and node features into the model for inference and regression prediction, and finally output the affinity and write it into the collection table.

[0008] By consulting the target protein of the AD target, the drug sample database is obtained, and the input data is arranged into the input data table after hybridization with the target protein; the data is preprocessed, the drug samples are data enhanced, the representation of the drug and target sequence is extracted, the drug molecular graph is constructed, the protein contact matrix is calculated, the protein residue subgraph is constructed and spliced and enhanced; the joint graph is constructed and its connectivity is enhanced, the node index offset is used, the full connectivity and root node self-loop are constructed; feature coding and node initialization are performed; the joint graph feature vector and node features are input into the model for inference and regression prediction, and finally the affinity is output and written into the collection table. BRIEF DESCRIPTION OF DRAWINGS

[0009] The application is further illustrated by the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the application. For ordinary skilled persons in the art, other drawings can be obtained without creative labor on the basis of the following drawings.

[0010] Figure 1 It is the deep neural network model project framework inherited by the application, which is used to realize the specific prediction of AD (Alzheimer's disease) drugs by improving the model framework structure.

[0011] Figure 2 It is the abstract drawing of the application, which generally shows the main technical process and module relationship of the application.

[0012] Figure 3 It is seven AD-related target data determined by literature screening and bioinformatics analysis in the application, including the key information of each target: the starting position (Start) on the protein chain, the ending position (End), the PDB file name (PdbPath), the target name and the target sequence (Sequence), which provides accurate target positioning data for subsequent model construction.

[0013] Figure 4 It is the complete structure of the model input data set in the application, which contains the following key fields: CAS number (biological experimental drug retrieval identifier), compound SMILES sequence (drug molecular structure representation), target protein sequence, target name, target protein chain total length, target starting position, target ending position, PDB file name and data source. DETAILED DESCRIPTION S101, for the selection of AD drug targets, select the relevant literature to screen the target and use Uniport and AlphaFold2 to determine the effective target position, finally screen out the following seven targets and their effective points and obtain their PDB files, see Figure 3 ; S102、For the database of drug samples, we accessed the public drug database DrugBank to obtain three drug databases for AD treatment, including drugs in early research, drugs in clinical trials, and approved drugs; S103、Based on the cross combination of the original database and the seven target points, a total of 100,000+ samples were generated, and the specific sample format is shown in Figure 4 , to generate the input data table input.csv.

[0014] S201、For input data preprocessing, read the samples containing the following fields from the input.csv file: target_sequence, compound_iso_smiles, target_sequence_start, target_sequence_end, pdb_file, parse and verify SMILES and PDB file paths, skip and record NaN in the results when SMILES or path does not exist fails to parse; S202、Data augmentation is performed on drug samples, where for each SMILES string, a sliding window masking strategy is used: randomly select a substructure for masking at each epoch, simulate compound substructure loss, and improve the robustness of the model to incomplete fragments; Add the corresponding segment in the code to introduce restricted Bernoulli noise in the protein residue contact graph (randomly "break" or "increase" a small number of original edges, simulate structure sequencing noise, and maintain the integrity of the overall graph structure), and remove a small number of samples that severely destroy the structure through screening; S203、Extract the representation of drug and target sequence, for each valid sample, if target_sequence is empty, call get_seq_from_pdb to extract the amino acid sequence from the PDB file; based on start and end, extract the binding segment subsequence and verify the legality of the extracted interval; S204、Drug molecule graph construction: call smiles_to_graph, use RDKitMolFromSmiles to convert SMILES to a molecular object, iterate through chemical bonds to generate an edge list, build a directed graph based on NetworkX and add self-loops, and generate edge_index_mol; S205、Calculate the protein contact matrix: call get_contact_map, use Bio.PDB to parse the PDB file structure, calculate the distance matrix between residues , generate a binary contact matrix according to the threshold ​S206, construct protein residue sub-graph and splice and enhance: call target_to_graph to map the intercepted binding segment sequence to the original PDB residue index, extract the corresponding residue pair edge list from the contact matrix, and construct edge_index_pro; merge edge_index according to the node offset of the molecular graph and the residue graph, add full connection edges (any node pair is connected) and virtual root node self-loop, and ensure the connectivity and information flow of the joint graph.

[0015] S301, use node index offset: shift the node index of the protein residue graph backward by the number of molecular graph nodes to ensure that the merged index does not overlap; S302, then merge Edge, splice the offset protein edge list and the molecular edge list to form the initial joint graph edge_index; S303, build full connection and root node self-loop: add inter-node bidirectional full connection edges to the joint graph, and introduce a virtual root node and its self-loop edge to ensure the overall connectivity of the graph.

[0016] S401, sequence encoding: call WordVocab to encode SMILES and intercepted subsequence into index sequence respectively; S402, embedding and LSTM attention: input the encoded sequence into nn.Embedding and bidirectional LSTM (2 layers, dropout=0.2), and calculate the attention weight through the SpatialGroupEnhance_for_1D module to generate smiles_enc and protein_enc respectively; S403, feature splicing: in reset_feature_single, splice smiles_enc, protein_enc and atomic / residue type one-hot label to generate node feature x, and set num_nodes and batch.

[0017] S5, Model inference and regression prediction: input the edge_index and node features x of the joint graph into GINConvNet (5 layers of GINConv + BatchNorm), aggregate the output graph-level representation gout through global_add_pool; apply multi-head attention fusion to gout and the sequence representation out_cat output by LSTM to obtain the fused representation fused_repr; input fused_repr into two-layer fully connected MLP (including Dropout, ReLU) to output the scalar affinity, and write it into the predicted_affinity column of predictions.csv.

[0018] Repeat this process until the prediction table predictions.csv is output after completion, and the prediction results of the affinity of AD therapeutic drugs to each target are obtained.

Claims

1. A method for predicting AD drug affinity based on data augmentation and graph neural networks, characterized in that: The method includes the following steps: S1: Consult the target proteins of AD targets, obtain the drug sample database, hybridize with the target protein sequence information to form input data, and then organize it into the input data table; S2 preprocesses the data, performs data augmentation on drug samples, extracts representations of drug and target sequences, constructs drug molecular maps, calculates protein contact matrices, constructs protein residue submaps, and performs splicing and augmentation. S3 constructs a joint graph and enhances its connectivity by using node index offsets to build fully connected graphs and root node self-loops; S4 performs feature encoding and node initialization; S5 inputs the joint graph feature vector and node features into the model for inference and regression prediction, and finally outputs the affinity, which is then written into the collection table.

2. The method for predicting AD drug target affinity based on data augmentation and graph neural networks according to claim 1, characterized in that: In step S2, the SMILES string is parsed using RDKit MolFromSmiles and the None result is processed as a null value.

3. The method for predicting AD drug target affinity based on data augmentation and graph neural networks according to claim 1, characterized in that: In step S202, for each SMILES string, a sliding window masking strategy is adopted, and the masking operation is performed with a probability of 0.3 for the window length.

4. The method for predicting AD drug target affinity based on data augmentation and graph neural networks according to claim 1, characterized in that: In step S303, when constructing the joint graph, a bidirectional fully connected edge is added between all node pairs, and a self-loop is added between the root node and itself. The root node index is node_count+1.

5. The method for predicting AD drug target affinity based on data augmentation and graph neural networks according to claim 1, characterized in that: In step S4, the sequence encoder includes: • The Embedding layer maps the one-hot encoded indices to real number vectors; • Bidirectional LSTM (2 layers, dropout=0.2) generates sequence context representations; • The SpatialGroupEnhance_for_1D enhancement module is used for calculating attention weights in the time dimension.

6. The method for predicting AD drug target affinity based on data augmentation and graph neural networks according to claim 1, characterized in that: In step S402, the protein contact matrix is ​​calculated by residue_contact_map, and nodes are considered connected if the distance between them is less than 8.0 Å.

7. The method for predicting AD drug target affinity based on data augmentation and graph neural networks according to claim 1, characterized in that: The graph representation extraction network GINConvNet contains 5 layers of GINConv + BatchNorm, uses global_add_pool for graph pooling, and passes two layers of FC + Dropout before outputting the regression values.

Images