Drug target affinity prediction system based on cross-modal feature fusion
Patent Information
- Application Number
- CN202510634027.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-16
- Publication Date
- 2026-11-03
- Estimated Expiration
- 2045-05-16
AI Technical Summary
Existing methods for predicting drug target affinity suffer from low prediction accuracy, mainly due to the lack of drug and protein characteristic information, which makes it impossible to accurately express the characteristics of drugs and proteins.
A drug target affinity prediction system employing cross-modal feature fusion is proposed. The system obtains drug SMILES strings and protein amino acid sequences through a data preprocessing module, extracts graph modal and sequence modal features of drugs and proteins using GCN and BiLSTM, fuses these features using the GMHSA algorithm, and finally uses a cross-attention network to predict drug target affinity.
It improves the accuracy of drug target affinity prediction and the generalization ability of the model, alleviates redundancy in different modal features, makes full use of information gain, and improves prediction accuracy.
Smart Images
Figure CN120544664B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bioinformatics, and in particular to a drug target affinity prediction system based on cross-modal feature fusion. Background Technology
[0002] Accurate prediction of drug target affinity (DTA) can help advance new drug development. Traditional biological experimental methods for DTA prediction are time-consuming and costly, while deep learning-based DTA prediction has the advantages of low cost and high efficiency. Therefore, deep learning-based DTA prediction methods currently occupy a mainstream position.
[0003] In recent years, researchers have proposed many DTA prediction methods. For example, GraphDTA uses graph neural networks to extract graph modal features of drugs and convolutional neural networks to extract sequence modal features of proteins. Then, a feedforward neural network (FNN) is used to predict drug target affinity using the joint features of drug graph modal features and protein sequence modal features. DGraphDTA uses graph neural networks to extract graph modal features of drugs and proteins, and then uses FNN to predict drug target affinity using the joint features of drug graph modal features and protein graph modal features. AttentionMGT fuses graph and sequence modal features using Transformer and cross-attention to obtain higher-quality features for predicting drug target affinity. However, the drug and protein features extracted by these methods are missing information, thus failing to accurately represent drugs and proteins, resulting in low DTA prediction accuracy. Summary of the Invention
[0004] The purpose of this invention is to address the problem of low prediction accuracy in existing DTA prediction methods, and to propose a drug target affinity prediction system based on cross-modal feature fusion.
[0005] The drug target affinity prediction system based on cross-modal feature fusion includes: a data preprocessing module, a sequence feature extraction module, a sequence modality feature extraction module, a cross-modal feature fusion module, and a drug target affinity prediction module;
[0006] The data preprocessing module is used to obtain drug SMILES strings and protein amino acid sequences, and to preprocess the drug SMILES strings and protein amino acid sequences to obtain drug molecular graphs, drug SMILES sequence embedding features, protein contact graphs, and protein amino acid sequence embedding features.
[0007] The sequence feature extraction module is used to obtain drug map modal features based on drug molecule maps and protein map modal features based on protein contact maps.
[0008] The sequence modality feature extraction module obtains drug sequence modality features using drug SMILES sequence embedding features and obtains protein sequence modality features using protein amino acid sequence embedding features.
[0009] The cross-modal feature fusion module is used to fuse drug graph modal features and drug sequence modal features to obtain drug fusion features, and to fuse protein graph modal features and protein sequence modal features to obtain protein fusion features;
[0010] The drug target affinity prediction module uses drug fusion features and protein fusion features to obtain drug target affinity values.
[0011] Furthermore, the data preprocessing module is used to obtain the drug SMILES string and the protein amino acid sequence, and to preprocess the drug SMILES string and the protein amino acid sequence to obtain the drug molecular map, drug SMILES sequence embedding features, protein contact map, and protein amino acid sequence embedding features, specifically:
[0012] a1. Obtain the PubChem ID of the drug, and based on the PubChem ID, obtain the SMILES string of the drug. Use the RDKit software package to parse the SMILES string of the drug to obtain the drug molecular diagram G. d ={V d ,X d E d};
[0013] Among them, V d It is the set of nodes in the drug molecule graph, X d It is the set of feature matrices of drug molecule graph nodes, E d It is the edge set of the drug molecule graph;
[0014] a2. Obtain SMILES sequence embedding features using the drug's SMILES string:
[0015] Emb d =RetNet(OneHot(Str d ))
[0016] Among them, Str d It is the SMILES string for the drug, OneHot() is the OneHot encoding, and Emb... d These are SMILES sequence embedding features, and RetNet() is the RetNet network;
[0017] a3. Obtain the PDB file corresponding to the amino acid sequence of the protein. Extract the spatial distance between any two amino acids in the protein recorded in the PDB file. Perform sparsification on the distances between amino acids using a preset threshold to obtain the amino acid contact matrix. Use the amino acid contact matrix as the adjacency matrix to obtain the protein contact map G. p ={V p ,X p E p};
[0018] Among them, V p It is the node set of the protein contact graph, X p It is the set of node feature matrices of the protein contact graph, E p It is the edge set of the protein contact graph;
[0019] a4. Using the amino acid sequence of a protein, obtain the amino acid sequence embedding features:
[0020] Emb p =RetNet(OneHot(Str p ))
[0021] Among them, Str p It is the amino acid sequence of a protein, Emb p It is an amino acid sequence embedding feature.
[0022] Furthermore, the sequence feature extraction module is used to obtain drug map modal features based on the drug molecule map and protein map modal features based on the protein contact map, specifically as follows:
[0023] b1. Aggregate neighbor features from the drug molecule graph and protein contact graph respectively to obtain node features with local dependencies in the protein contact graph and node features with local dependencies in the drug molecule graph, specifically:
[0024]
[0025] Where k is the layer label of the graph neural network, ReLU is the nonlinear activation function, MLP is the multilayer perceptron, GCN is the graph neural network, and i,j are the node labels in the drug molecule diagram. It is the feature representation of the i-th node in the drug molecule graph after the (k+1)-th update. This is the feature representation of node i in the drug molecule graph after the k-th update. This is the feature representation of node j in the drug molecule graph after the k-th update. w is the set of neighboring nodes of node i in the drug molecule graph. 1 w 2 w 3 w4 These are learnable parameters. This is the feature representation of node j' in the protein contact graph after the k-th update. This is the feature representation of node i' after the k-th update in the protein contact graph, where i', j' are the labels of the nodes in the protein contact graph. It is the feature representation of node i' after the (k+1)th update in the protein contact graph. MLP() is a multilayer perceptron network.
[0026] b2. Utilize node features with local dependencies in the protein contact graph and node features with local dependencies in the drug molecule graph to obtain node features with global dependencies in the drug molecule graph and node features with global dependencies in the protein contact graph:
[0027]
[0028] Among them, F d；i F is a drug node feature with global dependencies at node i in the drug molecule graph. p；i' is the protein node feature with global dependencies of node i' in the protein contact graph, and FC is a fully connected layer;
[0029] b3. Obtain drug graph modal features and protein graph modal features by utilizing node features with global dependencies in the drug molecule graph and protein contact graph.
[0030] Furthermore, in step b3, the drug graph modal features and protein graph modal features obtained by utilizing node features with global dependencies in the drug molecule graph and node features with global dependencies in the protein contact graph are specifically as follows:
[0031] GM d =GlobalMaxPool(F d )
[0032] GM p =GlobalMeanPool(F p )
[0033] Among them, F d F is a concatenation matrix obtained by concatenating the column features of nodes with global dependencies in the drug molecule diagram. p GM is a concatenated matrix obtained by concatenating the column features of nodes with global dependencies in the protein contact graph. p These are the graph modal features of the protein contact graph. GlobalMaxPool is the graph max pooling function, GlobalMeanPool is the graph average pooling function, and GM... d It is a graphical modal feature of the drug molecule diagram.
[0034] Furthermore, the sequence modality feature extraction module obtains drug sequence modality features using drug SMILES sequence embedding features and obtains protein sequence modality features using protein amino acid sequence embedding features, specifically as follows:
[0035] c1. Context-dependent drug SMILES sequence features and context-dependent protein amino acid sequence features are obtained using drug SMILES sequence embedding features and protein amino acid sequence embedding features, specifically:
[0036] S d =BiLSTM(Emb) d )
[0037] S p =BiLSTM(Emb) p )
[0038] Among them, S d It is a context-dependent drug SMILES sequence feature, S p It is a protein amino acid sequence feature that is context-dependent; BiLSTM() is a BiLSTM network.
[0039] c2. Context-dependent drug SMILES sequence features and context-dependent protein amino acid sequence features are used to obtain drug sequence modal features and protein sequence modal features.
[0040] Furthermore, in c2, drug sequence modality features and protein sequence modality features are obtained by utilizing context-dependent drug SMILES sequence features and context-dependent protein amino acid sequence features, specifically as follows:
[0041] c21. Utilize the learnable weight matrix and drug SMILES sequence features to obtain the query embedding q, key embedding k, and value embedding v. Then, use q, k, and v to obtain the drug sequence modal features:
[0042]
[0043] Where, q i* This represents the representation of the i-th word in the query embedding q, where k i* It is the representation of the i*th word in the key embedding k, d k It is the dimension of the key embedding k, a i* It is the attention coefficient of the i*th word in the drug SMILES sequence, e i* It is the attentional importance of the i*th word in the drug SMILES sequence, e j*It represents the attention importance of the j*th word in the SMILES sequence, v i* It is the representation of the i*th word in the value embedding v, SM d These are drug sequence modal features, where i*, j* are word labels in the drug SMILES sequence, and n* is the total number of words in the drug SMILES sequence.
[0044] c22. Using the learnable weight matrix and amino acid sequence features, obtain the query embedding q', bond embedding k', and value embedding v', respectively. Then, use q', k', and v' to obtain the protein sequence modal features:
[0045]
[0046] Where, q' i'* This represents the representation of the i'*th word in the query embedding q', where k' i'* v' represents the i'*th word in the key embedding k'. i'* For the representation of the i'*th word in the value embedding v', pe i'* For the attentional importance of the i'*th word in the protein amino acid sequence, pa i'* SM represents the attention coefficient of the i'*th word in the protein amino acid sequence. p These are protein sequence modal features, where i'*, j'* are word labels in the protein amino acid sequence, n'* is the total number of words in the protein amino acid sequence, and d'* is the total number of words in the protein amino acid sequence. k It is the dimension of the key embedding k'.
[0047] Furthermore, the cross-modal feature fusion module is used to fuse drug graph modal features and drug sequence modal features to obtain drug fusion features, and to fuse protein graph modal features and protein sequence modal features to obtain protein fusion features, specifically as follows:
[0048] d1. The drug fusion feature is obtained by fusing drug graph modal features and drug sequence modal features using the GMHSA algorithm, specifically as follows:
[0049] d11. The drug graph modal features and drug sequence modal features are first stacked and then normalized to obtain the normalized drug features F'. d Specifically:
[0050] F' d =Normal(stack(GM) d SM d )
[0051] Among them, stack() is for stacking, and Normal() is for normalization.
[0052] d12. Use matrix partitioning to divide F' dThe attention heads are divided evenly along the second dimension, and then F' is used. d The query vector Q corresponding to the l-th attention head is obtained from the l-th feature part. l Key vector K l Value vector V l Then, using the Q of the l-th attention head l and K l Obtain the attention importance coefficient matrix Specifically:
[0053] First, using the segment F' d Obtain the query vector Q corresponding to the l-th feature part and the l-th attention head. l Key vector K l Value vector V l :
[0054]
[0055] Among them, d' k It is K l Dimension size, It is F' d The l-th feature part after segmentation, W Q W K W V It is a weight matrix;
[0056] Then, using the Q of the l-th attention head l and K l Obtain the attention importance coefficient matrix Specifically:
[0057]
[0058] Among them, d' k It is K l Dimension size;
[0059] d13. Construct the sparse mask matrix SP, specifically as follows:
[0060]
[0061] Where w is the size of the perception window, n_global is the global perception point, and t and s are the normalized drug features F'. d Column labels;
[0062] d14. Using the sparse mask matrix SP and the attention importance coefficient matrix Obtain drug fusion characteristics;
[0063] d2. Protein fusion features are obtained by fusing protein map modal features and protein sequence modal features using the GMHSA algorithm.
[0064] Furthermore, the sparse mask matrix SP and the attention importance coefficient matrix are utilized in d14. Obtaining drug fusion characteristics, specifically:
[0065]
[0066] head l =mean2(T) l V l )
[0067]
[0068] Among them, T l Let be the attention weight matrix corresponding to the l-th attention head, and mean2(·) represents the mean of the input along the second dimension. l This represents the output value corresponding to the l-th attention head, DF is the drug fusion feature, h represents the number of attention heads, and w o It is head l weights, SP l It is the sparse mask matrix corresponding to the l-th attention head.
[0069] Furthermore, the drug target affinity prediction module obtains drug target affinity values using drug fusion features and protein fusion features, specifically:
[0070] e1. Using a combination of layer normalization and L2 normalization, the distributions of drug fusion features and protein fusion features are aligned to obtain the aligned drug fusion feature DF' and the aligned protein fusion feature PF':
[0071] DF' = Normal(LayerNormal(DF))
[0072] PF' = Normal(LayerNormal(PF))
[0073] Among them, LayerNormal() is an L2 normalization process;
[0074] e2. Using the aligned distribution of drug fusion features and aligned distribution of protein fusion features, the drug feature F of the fused protein features is obtained. p->d Protein features that integrate drug characteristics F d->p Specifically:
[0075] F d->p =CrossAttend (DF′,PF′)
[0076] F p->d =CrossAtten p (PF', DF')
[0077]
[0078] Among them, W q W k W v It is a weight vector;
[0079] e3. The drug feature and the fused drug feature are spliced together to obtain the drug-protein joint feature, and then the drug-protein joint feature is used to obtain the drug target affinity value.
[0080] Furthermore, the drug feature obtained by splicing and fusing the protein feature and the protein feature obtained by fusing the drug feature in e3 is used to obtain the drug-protein joint feature. Then, the drug target affinity value is obtained using the drug-protein joint feature, specifically:
[0081] affinity=MLP(BN(ReLU(H)))
[0082] H = concat(F) d->p ,F p->d )
[0083] Here, concat(·) means concatenating rows, BN means batch normalization, and H is the drug-protein co-feature.
[0084] The beneficial effects of this invention are as follows:
[0085] This invention uses a combination of GCN, FC, and pooling functions to extract drug graph modal features and protein graph modal features from drug molecule graphs and protein contact graphs, respectively. It employs a combination of RetNet, BiLTSM, and attention pooling to extract drug sequence modal features and protein sequence modal features. Furthermore, it proposes a multi-head sparse attention network (GMHSA) with a gating mechanism, and utilizes GMHSA to fuse the graph and sequence modal features of drugs and proteins to obtain drug fusion features and protein fusion features. This invention significantly reduces redundant features in drug and protein features across different modalities, resulting in drug and protein features with higher information purity. It fully utilizes the information gain from different modalities, enhancing the generalization ability of the prediction model while improving the prediction accuracy of DTA. This invention fully utilizes known drug and protein data, improving the accuracy of drug target affinity prediction. Attached Figure Description
[0086] Figure 1 This is a flowchart of the present invention;
[0087] Figure 2(a) is a visualization of the MSE index in the comparative experiment on the Davis dataset;
[0088] Figure 2(b) is a visualization of the Person index in the comparative experiment on the Davis dataset;
[0089] Figure 2(c) is a visualization of the CI index in the comparative experiment on the Davis dataset;
[0090] Figure 3(a) is a visualization of the MSE index in the comparative experiment on the KIBA dataset;
[0091] Figure 3(b) is a visualization of the Person index in the comparative experiment on the KIBA dataset;
[0092] Figure 3(c) is a visualization of the CI index in the comparative experiment on the KIBA dataset;
[0093] Figure 4(a) is a scatter plot visualizing the predicted and true values in the Davis dataset;
[0094] Figure 4(b) is a scatter plot visualizing the predicted and true values in the Davis dataset. Detailed Implementation
[0095] Specific implementation method one: as follows Figure 1 As shown, the drug target affinity prediction system based on cross-modal feature fusion in this embodiment includes: a data preprocessing module, a sequence feature extraction module, a sequence modal feature extraction module, a cross-modal feature fusion module, and a drug target affinity prediction module;
[0096] The data preprocessing module is used to obtain drug SMILES strings and protein amino acid sequences, and to preprocess the drug SMILES strings and protein amino acid sequences to obtain drug molecular graphs, drug SMILES sequence embedding features, protein contact graphs, and protein amino acid sequence embedding features.
[0097] The sequence feature extraction module is used to obtain drug map modal features based on drug molecule maps and protein map modal features based on protein contact maps.
[0098] The sequence modality feature extraction module obtains drug sequence modality features using drug SMILES sequence embedding features and obtains protein sequence modality features using protein amino acid sequence embedding features.
[0099] The cross-modal feature fusion module uses the GMHSA algorithm to fuse drug graph modal features and drug sequence modal features to obtain drug fusion features, and fuses protein graph modal features and protein sequence modal features to obtain protein fusion features;
[0100] The drug target affinity prediction module uses drug fusion features and protein fusion features to obtain drug target affinity values.
[0101] Specific implementation method two: The data preprocessing module performs the following specific processes:
[0102] a1. Obtain the PubChem ID of the drug, and based on the PubChem ID, obtain the SMILES string of the drug. Use the RDKit software package to parse the SMILES string of the drug to obtain the drug molecular diagram G. d ={V d ,X d E d};
[0103] Among them, V d It is the set of nodes in the drug molecule graph, X d；i ∈X d X d；i It is the i-th node V in the drug molecule diagram d；i The node feature matrix, X d It is a set of node feature matrices, E d It is an edge set, E d element E d；ij This indicates that there is a chemical bond between the i-th node and the j-th node, where i,j∈[1,n], and n is the number of nodes in the drug molecule diagram;
[0104] a2. Obtain a binary matrix of SMILES sequences using OneHot encoded drug SMILES strings, and then extract SMILES sequence embedding features from the binary SMILES matrix using a RetNet network:
[0105] Emb d =RetNet(OneHot(Str d ))
[0106] Among them, Str d It is the SMILES string for the drug, OneHot() is the OneHot encoding, and Emb... d It is a SMILES sequence embedding feature;
[0107] a3. Obtain the corresponding PDB file from the PDB website using the protein's amino acid sequence. Using the spatial distance between any two amino acids in the protein recorded in the PDB file and a preset threshold, perform sparsification on the distances between amino acids to obtain an amino acid contact matrix. Use the amino acid contact matrix as the adjacency matrix to obtain the protein contact graph G. p ={V p ,X p E p};
[0108] Among them, V p It is the set of nodes in the protein contact graph, X p It is a set of node feature matrices, E p It is an edge set;
[0109] In this step, the sparsity processing of the distance between amino acids is specifically performed as follows: obtain the spatial distance between any two amino acids in the protein recorded in the PDB file, and then retain the amino acids corresponding to the spatial distance between amino acids that is less than a preset threshold, thereby completing the sparsity processing.
[0110] a4. Using the amino acid sequence of the protein encoded by OneHot, obtain the binary matrix of the amino acid sequence, and then use the RetNet network to extract the amino acid sequence embedding features from the binary matrix of the amino acid sequence:
[0111] Emb p =RetNet(OneHot(Str p ))
[0112] Among them, Str p It is the amino acid sequence of a protein, Emb p It is an amino acid sequence embedding feature.
[0113] Specific Implementation Method 3: The sequence feature extraction module is used to obtain drug map modal features based on the drug molecule map and protein map modal features based on the protein contact map, specifically as follows:
[0114] b1. Using a Graph Convolutional Network (GCN), neighbor features are aggregated from the drug molecule graph and the protein contact graph to obtain node features with local dependencies in the protein contact graph and the drug molecule graph, respectively. Specifically:
[0115]
[0116] Where k is the layer label of the graph neural network, ReLU is the nonlinear activation function, MLP is the multilayer perceptron, and i,j are the node labels in the drug molecule graph. It is the feature representation of the i-th node in the drug molecule graph after the (k+1)-th update. This is the feature representation of node i in the drug molecule graph after the k-th update. This is the feature representation of node j in the drug molecule graph after the k-th update. w is the set of neighboring nodes of node i in the drug molecule graph. 1 w 2 w 3 w 4 These are learnable parameters. This is the feature representation of node j' in the protein contact graph after the k-th update. This is the feature representation of node i' after the k-th update in the protein contact graph, where i', j' are the labels of the nodes in the protein contact graph. It is the feature representation of node i' after the (k+1)th update in the protein contact graph. MLP() is a multilayer perceptron network.
[0117] In this step, the output of each layer of the graph neural network consists of node features with local dependencies. The set of node features constituting the local dependencies of node i in the drug molecule graph is analogous to the set of node features with local dependencies in the protein contact graph. When k = 0, It is the node feature matrix of the i-th node in the drug molecule graph. It is the node feature matrix of the i'th node in the protein screenshot.
[0118] b2. Using Fully Connected Network (FC), extract the global dependencies of node features in the drug molecule graph and protein contact graph respectively, to obtain node features with global dependencies in the drug molecule graph and the protein contact graph:
[0119]
[0120] Among them, F d；i F is a drug node feature with global dependencies at node i in the drug molecule graph. p；i' is the protein node feature with global dependencies of node i' in the protein contact graph, and FC is a fully connected layer;
[0121] b3. For node features with local and global dependencies, the drug graph modal features and protein graph modal features are obtained by compression using graph max pooling and graph average pooling functions, respectively, as shown in the following equation:
[0122] GM d =GlobalMaxPool(F d )
[0123] GM p =GlobalMeanPool(F p )
[0124] Among them, F d It is a feature matrix of nodes in the drug molecule graph representing global dependencies (a feature matrix obtained by concatenating the features of nodes in the drug molecule graph after FC update, column-wise). p It is a feature matrix of nodes with global dependencies in the protein contact graph (the feature matrix obtained by concatenating the features of nodes in the protein contact graph after FC update by column), GM d It is a graph modal feature of drug molecule diagrams, GM p These are the graph modal features of the protein contact graph. GlobalMaxPool is the graph max pooling function, and GlobalMeanPool is the graph average pooling function.
[0125] Specific Implementation Method Four: The sequence modality feature extraction module obtains drug sequence modality features using drug SMILES sequence embedding features and obtains protein sequence modality features using protein amino acid sequence embedding features, specifically as follows:
[0126] c1. Using BiLSTM, bidirectional cyclic dependencies are mined from drug SMILES sequence embedding features and protein amino acid sequence embedding features, respectively, to obtain context-dependent drug SMILES sequence features and context-dependent protein amino acid sequence features, specifically:
[0127] S d =BiLSTM(Emb) d )
[0128] S p =BiLSTM(Emb) p )
[0129] Among them, S d It is a context-dependent drug SMILES sequence feature, S p It is a protein amino acid sequence feature that is context-dependent; BiLSTM() is a BiLSTM network.
[0130] c2. The attention pooling function is used to compress the context-dependent drug SMILES sequence features and the context-dependent protein amino acid sequence features, respectively, to obtain the drug sequence modality features and protein sequence modality features, specifically:
[0131] c21. Utilize the learnable weight matrix and drug SMILES sequence features to obtain the query embedding q, key embedding k, and value embedding v. Then, use q, k, and v to obtain the drug sequence modal features:
[0132]
[0133] Where, q i* This represents the representation of the i*th word in the query embedding, k i* It is the representation of the i*th word in the key embedding, d k It is the key embedding dimension, a i* It is the attention coefficient of the i*th word obtained after softmax calculation, e i* It is the attentional importance of the i*th word in the drug SMILES sequence, e j* It represents the attention importance of the j*th word in the SMILES sequence, v i* The representation of the i*th word in the value embedding, SM d These are the drug sequence modal features obtained through attention pooling, where i*,j* are the word labels in the drug SMILES sequence, and n* is the total number of words in the drug SMILES sequence.
[0134] This step uses the scaled dot product to calculate the attention importance coefficient, the softmax function to calculate the attention weight, and finally calculates the drug sequence modal features based on the attention weight.
[0135] c22. Using three learnable weight matrices and amino acid sequence features, obtain the query embedding q', bond embedding k', and value embedding v' respectively. Then, use q', k', and v' to obtain the protein sequence modal features:
[0136]
[0137] Where, q' i'* This represents the representation of the i'*th word in the query embedding q', where k' i'* Let v' represent the i'*th word in the key embedding. i'* For the representation of the i'*th word in the value embedding v', pe i'* For the attentional importance of the i'*th word in the protein amino acid sequence, pa i'* SM represents the attention coefficient of the i'*th word in the protein amino acid sequence obtained after softmax calculation. p These are protein sequence modal features obtained through attention pooling, where i'*,j'* are word labels in the protein amino acid sequence, n'* is the total number of words in the protein amino acid sequence, and d'* is the total number of words. k It is the dimension of the key embedding k'.
[0138] Specific Implementation Method 5: The cross-modal feature fusion module uses the GMHSA algorithm to fuse drug graph modal features and drug sequence modal features to obtain drug fusion features, and fuses protein graph modal features and protein sequence modal features to obtain protein fusion features, specifically as follows:
[0139] d1. A multi-head sparse attention algorithm with gating mechanism, GMHSA, is proposed. The GMHSA algorithm is used to fuse drug graph modal features and drug sequence modal features to obtain drug fusion features, specifically:
[0140] d11. Stack and normalize the drug graph modal features and drug sequence modal features, align the feature distributions of the two modalities, and obtain the normalized drug features F'. d Specifically:
[0141] F' d =Normal(stack(GM) d SM d )
[0142] Among them, stack() is for stacking, and Normal() is for normalization.
[0143] d12. Use matrix segmentation to divide the input features F' d Based on the selected number of attention heads, the data is divided evenly along the second dimension. Then, for each part, three linear transformations are applied to obtain the query vector Q corresponding to the l-th attention head. l Key vector K l Value vector V l Then, the Q of the l-th attention head is calculated using the scaled dot product. l and K l Obtain the attention importance coefficient matrix Specifically:
[0144]
[0145] Among them, d' k It is K l Dimension size, It is F' d The l-th feature part after average segmentation;
[0146] d13. Construct the sparse mask matrix SP, specifically as follows:
[0147]
[0148] Where w is the size of the perception window, n_global is the global perception point, and t and s are the normalized drug features F'. d Column labels;
[0149] In this step, a sparse mask matrix for each attention head is generated based on the size of the third dimension of the input features according to the sparse mask generation algorithm. A perceptual window w is set; if the difference between the positions of the two parties calculating the importance coefficient in the feature matrix is less than w, then within the attention importance range, it is assigned a value of 1. A global perceptual point n_global is set; if the party calculating the attention importance coefficient is within [0, n_global], then within the attention importance range, it is assigned a value of 1; otherwise, the attention coefficient is not calculated.
[0150] d14. Using matrix multiplication, the sparse attention weight matrix of the i-th attention head is obtained based on the attention importance matrix and the sparse mask matrix. Then, matrix multiplication is used to process the sparse attention weight matrix and the value vector to obtain the output value corresponding to the i-th attention head. Finally, a gating mechanism is used to calculate the weight of each attention head in the final output value, and multiplied by the corresponding attention head output to obtain the final output value of the multi-head gated sparse attention, which is the drug fusion feature. Specifically:
[0151]
[0152] head l =mean2(T) l V l )
[0153]
[0154] Among them, T l Let be the attention weight matrix corresponding to the l-th attention head, and mean2(·) represents the mean of the input along the second dimension. l This represents the output value corresponding to the l-th attention head, DF is the drug fusion feature, h represents the number of attention heads, and w o It is head l weights, SP l It is the sparse mask matrix corresponding to the l-th attention head.
[0155] d2. Protein fusion features are obtained by fusing protein map modal features and protein sequence modal features using the GMHSA algorithm;
[0156] The steps of d2 are similar to those of d1, with the input being GM. p SM p To obtain the protein fusion feature PF;
[0157] Specific Implementation Method Six: The drug target affinity prediction module obtains drug target affinity values using drug fusion features and protein fusion features, specifically as follows:
[0158] e1. Using a combination of layer normalization and L2 normalization, the distributions of drug fusion features and protein fusion features are aligned to obtain the aligned drug fusion feature DF' and the aligned protein fusion feature PF':
[0159] DF' = Normal(LayerNormal(DF))
[0160] PF' = Normal(LayerNormal(PF))
[0161] e2. Using a standard cross-attention network to process the aligned drug fusion features and aligned protein fusion features, the drug feature F of the fused protein features is obtained. p->d Protein features that integrate drug characteristics F d->p Specifically:
[0162] F d->p =CrossAtten d (DF′,PF′)
[0163] F p->d =CrossAtten p (PF', DF')
[0164]
[0165] Among them, W q W k W v It is a weight vector;
[0166] e3. The drug features obtained by fusion of protein features and drug features calculated through cross-attention are concatenated to obtain drug-protein joint features. Then, the ReLU nonlinear activation function and batch normalization are used to enhance the expressive power of the drug-protein joint features. Finally, the enhanced joint features are input into a Multi-Layer Perceptron (MLP) model to predict the drug target affinity value. Specifically:
[0167] H = concat(F) d->p ,F p->d )
[0168] affinity=MLP(BN(ReLU(H)))
[0169] Here, concat(·) means concatenating two features row by row, BN means batch normalization operation, and H is the drug-protein joint feature.
[0170] Example: To verify the beneficial effects of the present invention, the following experiments were conducted:
[0171] Taking a specific drug SMILES sequence and protein amino acid sequence pair as an example, the SMILES sequence is processed into a drug molecule graph and SMILES sequence embeddings, while the protein amino acid sequence is processed into a protein contact graph and amino acid sequence embeddings. For the drug graph modal features, the original node features of the drug molecule graph are extracted using GCN to obtain node features. Then The updated node features are input into the FC (Full Functional) system. Drug map modal features were obtained after max pooling. For protein graph modal features, the original node features of the protein contact graph are extracted using GCN to obtain node features. Then The updated node features are input into the FC (Full Functional) system. Drug map modal features were obtained after max pooling. For drug sequence modal features, SMILES sequence embeddings are processed by BiLSTM to obtain SMILES sequence features. Then, the SMILES sequence features are input into an attention pooling network to obtain drug sequence modality features. For protein sequence modality features, amino acid sequence embeddings are obtained by processing with BiLSTM to obtain amino acid sequence features. Then, the amino acid sequence features are input into an attention pooling network to obtain protein sequence modality features. Drug fusion features are obtained by fusing drug graph modal features and drug sequence modal features using the GMHSA algorithm. Protein fusion features are obtained by fusing protein map modal features and protein sequence modal features using the GMHSA algorithm. By aligning the distributions of drug fusion features and protein fusion features using layer normalization and L2 normalization, the feature-aligned drug fusion features are obtained. Drug fusion features aligned with features Cross-attention computation is used to transfer drug information to protein fusion features, thereby obtaining protein features. Cross-attention computation is used to transfer protein information to drug fusion features, thereby obtaining drug features. Finally, the protein and drug characteristics are concatenated to obtain the drug-protein combined characteristic. The input is fed into a three-layer fully connected network to output the predicted drug target affinity value. The prediction performance is evaluated, and the DTA prediction results using different methods on different datasets are shown in Tables 1-3. In the tables, CMFDTA refers to the algorithm of this invention, and the prediction performance results are as follows: Figure 2(a)-Figure 4(b) As shown.
[0172] Table 1 Evaluation Results
[0173]
[0174] Table 2 shows the comparative experimental results on the Davis dataset.
[0175]
[0176] Table 3 shows the comparative experimental results on the KIBA dataset.
[0177]
[0178] This embodiment conducts extensive experiments on the Davis and KIBA datasets, verifying that the present invention outperforms several state-of-the-art DTA prediction methods under different experimental settings. The present invention exhibits good robustness and generalization ability. Furthermore, the present invention fully utilizes known drug and protein data to improve the accuracy of drug target affinity prediction.
Claims
1. A drug target affinity prediction system based on cross-modal feature fusion, characterized in that: The system includes: a data preprocessing module, a sequence feature extraction module, a sequence modality feature extraction module, a cross-modality feature fusion module, and a drug target affinity prediction module; The data preprocessing module is used to obtain drug SMILES strings and protein amino acid sequences, and to preprocess the drug SMILES strings and protein amino acid sequences to obtain drug molecular graphs, drug SMILES sequence embedding features, protein contact graphs, and protein amino acid sequence embedding features. The sequence feature extraction module is used to obtain drug map modal features based on drug molecule maps and protein map modal features based on protein contact maps. The sequence modality feature extraction module obtains drug sequence modality features using drug SMILES sequence embedding features and obtains protein sequence modality features using protein amino acid sequence embedding features. The cross-modal feature fusion module is used to fuse drug graph modal features and drug sequence modal features to obtain drug fusion features, and to fuse protein graph modal features and protein sequence modal features to obtain protein fusion features, specifically: d1. The drug fusion feature is obtained by fusing drug graph modal features and drug sequence modal features using the GMHSA algorithm of a multi-head sparse attention network with gating mechanism, specifically as follows: d11. The drug graph modal features and drug sequence modal features are first stacked and then normalized to obtain the normalized drug features. Specifically: in, It is a stacking process. It is a normalization process. These are the graphical modal features of drug molecule diagrams. These are drug sequence modal characteristics; d12. Use matrix partitioning method to divide The attention heads are divided evenly along the second dimension, and then... The The first feature part is obtained The query vector corresponding to each attention head Key vector Value vector Then use the first A person's attention and Obtain the attention importance coefficient matrix Specifically: First, using segmentation Obtain the The first feature part is obtained The query vector corresponding to each attention head Key vector Value vector : in, yes Dimension size, yes The segmented first Each feature part, , , It is a weight matrix; Then, using the first A person's attention and Obtain the attention importance coefficient matrix Specifically: in, yes Dimension size; d13. Construct a sparse mask matrix Specifically: in, It is a window size sensor. The global perception point is t, and s are the normalized drug characteristics. Column labels; d14. Using sparse masking matrices Attention Importance Coefficient Matrix Obtaining drug fusion characteristics, specifically: in, For the first Attention weight matrix corresponding to each attention head This means that the input will be averaged along the second dimension. Indicates the first The output value corresponding to each attention head As a drug fusion feature, Indicates the number of attention heads. yes The weight, It is the first The sparse mask matrix corresponding to each attention head; d2. Protein fusion features are obtained by fusing protein map modal features and protein sequence modal features using the GMHSA algorithm; The drug target affinity prediction module uses drug fusion features and protein fusion features to obtain drug target affinity values.
2. The drug target affinity prediction system based on cross-modal feature fusion according to claim 1, characterized in that: The data preprocessing module is used to obtain drug SMILES strings and protein amino acid sequences, and to preprocess the drug SMILES strings and protein amino acid sequences to obtain drug molecular graphs, drug SMILES sequence embedding features, protein contact graphs, and protein amino acid sequence embedding features, specifically: a1. Obtain the PubChem ID of the drug, and based on the PubChem ID, obtain the SMILES string of the drug. Use the RDKit software package to parse the SMILES string of the drug to obtain the drug molecular diagram. ; in, It is a set of nodes in the drug molecule graph. It is a set of feature matrices of drug molecule graph nodes. It is the edge set of the drug molecule graph; a2. Obtain SMILES sequence embedding features using the drug's SMILES string: in, It is the SMILES string for the drug. yes coding, It is a SMILES sequence embedding feature. It is a RetNet network; a3. Obtain the PDB file corresponding to the amino acid sequence of the protein. Extract the spatial distance between any two amino acids in the protein recorded in the PDB file. Perform sparsification on the distances between amino acids using a preset threshold to obtain an amino acid contact matrix. Use the amino acid contact matrix as the adjacency matrix to obtain the protein contact map. ; in, It is the node set of the protein contact graph. It is a set of node feature matrices of the protein contact graph. It is the edge set of the protein contact graph; a4. Using the amino acid sequence of a protein, obtain the amino acid sequence embedding features: in, It is the amino acid sequence of a protein. It is an amino acid sequence embedding feature.
3. The drug target affinity prediction system based on cross-modal feature fusion according to claim 2, characterized in that: The sequence feature extraction module is used to obtain drug map modal features based on the drug molecule map and protein map modal features based on the protein contact map, specifically: b1. Aggregate neighbor features from the drug molecule graph and protein contact graph respectively to obtain node features with local dependencies in the protein contact graph and node features with local dependencies in the drug molecule graph, specifically: in, These are the layer labels of a graph neural network. ReLU is a non-linear activation function, MLP is a multilayer perceptron, and GCN is a graph neural network. These are the labels of nodes in the drug molecule diagram. It is the i-th node in the drug molecule diagram. The updated feature representation It is a node in the drug molecule diagram No. The updated feature representation It is a node in the drug molecule diagram No. The updated feature representation It is a node in the drug molecule diagram The set of neighboring nodes, , , , These are learnable parameters. It is a node in the protein contact diagram. No. The updated feature representation It is a node in the protein contact diagram. No. The updated feature representation These are the labels of the nodes in the protein contact diagram. It is a node in the protein contact diagram. No. The updated feature representation It is a multilayer perceptron network; b2. Utilize node features with local dependencies in the protein contact graph and node features with local dependencies in the drug molecule graph to obtain node features with global dependencies in the drug molecule graph and node features with global dependencies in the protein contact graph: in, It is a drug node feature with global dependencies for node i in the drug molecule graph. It is a node in the protein contact diagram. Protein node features with global dependencies It is a fully connected layer; b3. Obtain drug graph modal features and protein graph modal features by utilizing node features with global dependencies in the drug molecule graph and protein contact graph.
4. The drug target affinity prediction system based on cross-modal feature fusion according to claim 3, characterized in that: The method described in b3, which utilizes node features with global dependencies in the drug molecule graph and node features with global dependencies in the protein contact graph to obtain drug graph modal features and protein graph modal features, specifically involves: in, It is a splicing matrix obtained by concatenating the node features with global dependencies in the drug molecule diagram column by column. It is a concatenation matrix obtained by concatenating the column features of nodes with global dependencies in the protein contact graph. These are graphical modal features of protein contact maps. For graph max pooling functions, The graph average pooling function is used. It is a graphical modal feature of the drug molecule diagram.
5. The drug target affinity prediction system based on cross-modal feature fusion according to claim 4, characterized in that: The sequence modality feature extraction module obtains drug sequence modality features using drug SMILES sequence embedding features and protein sequence modality features using protein amino acid sequence embedding features. Specifically: c1. Context-dependent drug SMILES sequence features and context-dependent protein amino acid sequence features are obtained using drug SMILES sequence embedding features and protein amino acid sequence embedding features, specifically: in, These are context-dependent drug SMILES sequence characteristics. It is a protein amino acid sequence feature that is context-dependent. yes network; c2. Context-dependent drug SMILES sequence features and context-dependent protein amino acid sequence features are used to obtain drug sequence modal features and protein sequence modal features.
6. The drug target affinity prediction system based on cross-modal feature fusion according to claim 5, characterized in that: In step c2, drug sequence modality features and protein sequence modality features are obtained by utilizing context-dependent drug SMILES sequence features and context-dependent protein amino acid sequence features, specifically as follows: c21. Obtaining query embeddings using learnable weight matrices and drug SMILES sequence features. , key embedding Sum embedding Then use , and Obtaining drug sequence modal features: in, Indicates query embedding The Middle The representation of a word, Key embedding The Middle The representation of a word, Key embedding Dimensions It is the first in the drug SMILES sequence. Attention coefficient of each word It is the first in the drug SMILES sequence. The importance of attention for each word It is the SMILES sequence number The importance of attention for each word Value embedding The Middle The representation of a word, It is a drug sequence modal feature. These are word tags in the drug SMILES sequence. It is the total number of words in the drug SMILES sequence; c22. Query embeddings are obtained using the learnable weight matrix and amino acid sequence features respectively. , key embedding Sum embedding Then use , and Obtaining protein sequence modal features: in, Indicates query embedding The Middle The representation of a word, Embedded for key The Middle The representation of a word, Embedding for values The Middle The representation of a word, The first amino acid in the protein sequence The importance of attention for each word The first amino acid in the protein sequence Attention coefficient of each word It is a protein sequence modal feature. These are the word labels in the amino acid sequence of a protein. It is the total number of words in the amino acid sequence of a protein. Key embedding Dimensions.
7. The drug target affinity prediction system based on cross-modal feature fusion according to claim 6, characterized in that: The drug target affinity prediction module obtains drug target affinity values using drug fusion features and protein fusion features, specifically: e1. Align the distributions of drug fusion features and protein fusion features using a combination of layer normalization and L2 normalization to obtain the aligned distribution of drug fusion features. Protein fusion characteristics after alignment and distribution : in, It is an L2 normalization process; e2. Utilize the aligned distribution of drug fusion features and aligned distribution of protein fusion features to obtain drug features of fused protein features. and protein characteristics of fusion drugs Specifically: in, , , It is a weight vector; e3. The drug-protein joint feature is obtained by splicing the drug-protein joint feature and the fused drug-protein joint feature. Then, the drug-protein joint feature is used to obtain the drug target affinity value. .
8. The drug target affinity prediction system based on cross-modal feature fusion according to claim 7, characterized in that: The drug-protein combined feature is obtained by splicing and fusing the drug-protein features in e3, and then the drug-protein combined feature is used to obtain the drug target affinity value. Specifically: in, This indicates concatenation by line. This indicates a batch normalization operation, where H represents the drug-protein co-characteristic.