A drug-target affinity prediction method based on multimodal synergy

Abstract

The present invention relates to a drug-target affinity prediction method based on multimodal collaborative generation, comprising: preprocessing drugs and proteins respectively to obtain preset characterization data of the drugs and proteins; extracting features from the preset characterization data of the drugs and proteins respectively to obtain feature extraction results of the preset characterization data of the drugs and proteins; fusing the feature extraction results of the drugs and the feature extraction results of the proteins respectively to obtain drug fusion features and protein fusion features; using a joint attention mechanism to interactively process the drug fusion features and protein fusion features to obtain interactive features; and predicting the binding affinity between the drugs and proteins based on the interactive features.

Description

Technical Field

[0001] The present invention relates to the technical field of DTA prediction, and in particular to a drug target affinity prediction method based on multimodal synergistic generation. Background Art

[0002] Drug treatment targeting a specific target often achieves significant therapeutic effects. Therefore, the prediction of drug-target affinity (DTA) plays a key role in drug discovery. However, traditional wet lab methods are not only time-consuming and costly, but can only analyze a limited number of drug combinations. With the accumulation of a large amount of biomedical data, deep learning provides an efficient DTA prediction method. By using technologies such as convolutional neural networks, recurrent neural networks, graph neural networks, and Transformer, researchers can automatically extract key features of drugs and targets for DTA prediction. These methods not only speed up drug screening, but also shorten the development cycle and reduce costs, which has a significant impact on modern drug research and development.

[0003] Although deep learning technology has demonstrated good performance in DTA prediction, there is still room for improvement in terms of representation form, fusion methods, interaction methods, interpretability, and generalization capabilities. First, most models only represent the target structure as an adjacency matrix, such as a contact graph or distance graph, which does not fully utilize the three-dimensional structural features of the protein, limiting the model's utilization of the target's structural complexity and biological interpretability. Second, most models fail to comprehensively consider the sequence and structural features of the drug and target, failing to fully utilize the complementary advantages of these features to enhance the effect of fusion learning. In addition, most models still lack convincing evidence for automatically extracting key features of drugs and targets. Finally, many DTA predictors simulate drug-target interactions through simple concatenation, which cannot accurately reflect the interaction process between drugs and targets. Summary of the Invention

[0004] The purpose of the present invention is to propose a drug-target affinity prediction method based on multimodal synergistic generation to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A drug-target affinity prediction method based on multimodal synergistic generation, comprising:

[0007] Preprocessing the drug and protein separately to obtain preset characterization data of the drug and protein; wherein the preset characterization data of the drug includes: drug SMILES sequence, drug two-dimensional contact map, and pharmacophore; the preset characterization data of the protein includes: protein FASTA sequence, protein two-dimensional distance map, and protein three-dimensional structure map;

[0008] Performing feature extraction on the preset characterization data of the drug and the protein respectively to obtain feature extraction results of the preset characterization data of the drug and the protein;

[0009] The feature extraction results of drugs and protein are fused respectively to obtain drug fusion features and protein fusion features;

[0010] Using a joint attention mechanism, the drug fusion features and the protein fusion features are interactively processed to obtain interactive features;

[0011] Based on the interaction characteristics, the binding affinity between the drug and the protein is predicted.

[0012] Optionally, the drug SMILES sequence is based on different SMILES character compositions;

[0013] The drug SMILES sequence is:

[0014]

[0015] Where D represents the drug SMILES sequence, d i Indicates the i-th SMILES character, N D Represents a vocabulary consisting of 62 SMILES characters, m d Indicates the length of the SMILES sequence;

[0016] The protein FASTA sequences are based on different amino acid compositions;

[0017] The protein FASTA sequence is:

[0018]

[0019] Where P represents the protein FASTA sequence, p i Indicates the i-th amino acid character in the sequence, N P Represents a vocabulary set containing preset amino acids, n p Indicates the sequence length of the protein;

[0020] The two-dimensional drug contact map is constructed based on the two-dimensional topology of the drug map;

[0021] The drug diagram is as follows:

[0022] G D =(V D , E D )

[0023] Among them, V D represents the drug molecule node set, ED represents the drug molecule edge set, G D Represents a drug graph;

[0024] The two-dimensional contact map of the drug:

[0025]

[0026] in, represents the two-dimensional contact graph of drugs, that is, the node adjacency matrix of the drug graph, Represents the number of nodes. If there is an edge between node i and node j, then is equal to 1, otherwise 0; Represents the node feature matrix of the drug graph, where the nodes are represented as dimensional vector;

[0027] The protein two-dimensional distance map is constructed based on the distances between amino acid residues and the two-dimensional topology of the protein;

[0028] The protein map is:

[0029] G P =(V P , E P )

[0030] Among them, V P represents the set of protein residues, E P represents the protein edge set, G P represents a protein map;

[0031] The protein two-dimensional distance map is:

[0032]

[0033] in, represents the two-dimensional distance graph of the protein, i.e., the residue adjacency matrix of the protein graph, Represents the number of residues; if the distance between residue i and residue j is less than Elements is equal to 1, otherwise 0; Represents the node feature matrix of the protein graph, where the nodes are represented as dimensional vector;

[0034] The pharmacophore is constructed by extracting substructure features from the SMILES sequence of each drug;

[0035] The pharmacophore is:

[0036] F C ={f1, f2, ..., f i,...,f 27}, f i ∈N F

[0037] Among them, f i Indicates the frequency of occurrence of the i-th specific pharmacophore feature identifier, N F represents a collection of 27 common pharmacophores, F C represents the pharmacophore;

[0038] The protein three-dimensional structure diagram is based on the node feature X of the protein Tnode Create and add side information features; wherein the side information features include: distance threshold, connectivity between atoms, distance between α carbon atoms and distance between residue centers.

[0039] Optionally, obtain the node feature X of the protein Tnode include:

[0040] Calculate interatomic distances;

[0041] Calculate dihedral information in protein residues, expressed as interatomic rotation angles;

[0042] Based on the interatomic distances and interatomic rotation angles, the three-dimensional structural features of the protein are encoded. Td ;

[0043] Transfer the embedding vectors of the pre-trained ESM-2 model As a supplement to the three-dimensional structure of proteins; p represents the length of the protein pocket, d p represents the embedding dimension of the protein;

[0044] By splicing X Td and X Te , obtain the node features of the protein.

[0045] Optionally, performing feature extraction on the preset characterization data of the drug and protein respectively includes:

[0046] Feature extraction of drug SMILES sequences:

[0047] Embed the drug SMILES sequence to obtain the high-dimensional features of the drug;

[0048] Position encoding of drug SMILES sequences to obtain drug position features;

[0049] Add the high-dimensional features of the drug and the position features of the drug to obtain the one-dimensional sequence feature X of the drug SMILES sequence feature extraction result. SD ;

[0050] Feature extraction of protein FASTA sequences:

[0051] Embed the protein FASTA sequence to obtain the high-dimensional features of the protein;

[0052] Perform position encoding on the protein FASTA sequence to obtain the positional characteristics of the protein;

[0053] Add the high-dimensional features of the protein and the positional features of the protein to obtain the feature extraction result of the protein FASTA sequence, and obtain the one-dimensional sequence feature X SP ;

[0054] Perform feature extraction on the two-dimensional contact map of the drug:

[0055] The node adjacency matrix of the drug two-dimensional contact graph and node feature matrix Input to the multi-head graph attention network to obtain output in, represents the number of nodes in the two-dimensional contact graph of the drug, Represents the characteristic dimension of each node in the drug molecule, Represent the feature vector space of nodes in the drug graph;

[0056] The output of the multi-head graph attention network After two GCNs, two linear layers and one adaptive pooling layer, the feature extraction result of the two-dimensional contact map of the drug is obtained. The two-dimensional contact map feature X GC ;

[0057] Extracting features from the protein two-dimensional distance map includes:

[0058] The node adjacency matrix of the protein two-dimensional distance graph and node feature matrix Input to the multi-head graph attention network to obtain output in, represents the characteristic dimension of each node in the protein, represents the number of nodes in the two-dimensional distance graph of proteins, A feature vector space representing nodes in a protein graph;

[0059] Then After two GCNs, two linear layers and one adaptive pooling layer, the feature extraction result of the two-dimensional distance map of the protein is obtained, and the two-dimensional contact map feature X is obtained. GP ;

[0060] Feature extraction of the pharmacophore:

[0061] The pharmacophore is projected to a high dimension through a linear layer, and each layer is normalized and subjected to a nonlinear activation function ReLU, and then projected to a low dimension through a linear layer to obtain the feature extraction result of the pharmacophore: pharmacophore feature X DC ;

[0062] Perform feature extraction on the three-dimensional structure diagram of the protein:

[0063] Extract the node features g in the protein three-dimensional structure graph n and edge feature g e , then g n Perform embedding coding, adding position coding and Dropout operations to g e Perform embedding coding operation, and then convert the protein three-dimensional structure diagram X T 、g n and g e As input, the updated g is obtained through multiple layers of Graph Transformer Layer n ; for g n Use dgl.split to reorganize into computable tensors as K and V, and use a trainable weight matrix as Q, input into the attention mechanism, and add the output results and take the average value to obtain the final protein three-dimensional structure feature extraction result three-dimensional structure feature X TP .

[0064] Optionally, obtaining drug fusion features and protein fusion features includes:

[0065] X GC With X DC Add them as K and V, and X SC As Q, an intra-molecular multi-head attention mechanism is used for modality fusion;

[0066] X GP With X TP Weighted fusion as K and V, X SP As Q, an intra-molecular multi-head attention mechanism is used for modality fusion;

[0067] After inputting all Q, K, and V into the scaled dot product attention, the attention calculation coefficient of each head is obtained. After splicing the calculation results of each head, two feedforward operations are performed to obtain the final drug fusion feature X. C and protein fusion feature X P .

[0068] Optionally, interactively processing the drug fusion feature and the protein fusion feature to obtain the interactive feature includes:

[0069] Fusion of drug features X Cand protein fusion feature X P Divide into two parts, namely X C1 With X P1 、X C2 With X P2 ;

[0070] X C1 With X P1 They are activated by ReLU respectively, and then the Einstein summation operation and the second nonlinear activation by Sigmoid are performed to obtain X cp , then X Cp Perform Einstein summation and take the reciprocal, and compare it with X cp1 Perform Einstein summation again to get the interaction result X of the first channel cp1 ,

[0071] X C2 With X P2 The interaction result X of the first channel is obtained by directly performing Einstein summation and nonlinear activation through Tanh cp2 ;

[0072] The interaction result X of the two channels cP1 With X cp2 Once again, the Einstein summation operation is performed to obtain the final interaction feature X final .

[0073] Optionally, performing a binding affinity prediction between a drug and a protein includes:

[0074] The interaction features are linearly adjusted to a first preset dimension, then normalized to adjust the data size, Dropout to reduce overfitting, ReLU to perform nonlinear activation, linearly adjusted to a second preset dimension, then normalized to adjust the data size again, Dropout to reduce overfitting, ReLU to perform nonlinear activation, and finally linearly adjusted to the final dimension to obtain the prediction result.

[0075] The beneficial effects of the present invention are:

[0076] The present invention first processes drugs and proteins into one-dimensional sequence representation, two-dimensional contact (or distance) graph representation, pharmacophore representation and protein three-dimensional structure graph representation. These different representation forms can fully capture the characteristics of drugs and proteins, ensuring the richness and comprehensiveness of information. Then, the multi-channel module extracts the multiple perspective features of drugs and proteins respectively, realizing the synchronous capture of drug and protein sequence and structural features. Then, the intramolecular multi-head attention mechanism is used to fuse the multiple modal features of drugs and proteins, making full use of their sequence and structural information, thereby improving the comprehensiveness and accuracy of feature fusion. Subsequently, the interactive fusion of drug and protein features is performed through the joint attention module to more accurately simulate the interaction between them and enhance the prediction ability of drug-target binding. Finally, the DTA prediction results are obtained through the multi-layer perceptron prediction module composed of multiple layers of fully connected layers, providing strong support for drug design and target research. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. 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 paying any creative work.

[0078] Figure 1 Schematic diagram of a DTA prediction model structure based on multimodal collaborative generation according to an embodiment of the present invention;

[0079] Figure 2 This is a schematic diagram of the embedded module structure according to an embodiment of the present invention;

[0080] Figure 3 This is a schematic diagram of a graph stream aggregation network structure according to an embodiment of the present invention;

[0081] Figure 4 This is a schematic diagram of the structure of a Graph Transfromer module according to an embodiment of the present invention;

[0082] Figure 5 This is a schematic diagram of the structure of the Graph Transfromer Layer in an embodiment of the present invention;

[0083] Figure 6 Schematic diagram of the structure of the joint attention mechanism of an embodiment of the present invention. DETAILED DESCRIPTION

[0084] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0085] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0086] This example proposes a drug-target affinity prediction method based on multimodal synergy, including:

[0087] Preprocessing drugs and proteins separately to obtain preset characterization data of drugs and proteins; wherein the preset characterization data of drugs include: drug SMILES sequence, drug two-dimensional contact map, and pharmacophore; the preset characterization data of proteins include: protein FASTA sequence, protein two-dimensional distance map, and protein three-dimensional structure map;

[0088] Specifically, in this embodiment, the SMILES sequence of the drug is composed of different SMILES characters, which are defined as follows:

[0089]

[0090] Where D represents the drug SMILES sequence, d i Indicates the i-th SMILES character, N D Represents a vocabulary consisting of 62 SMILES characters, m d Indicates the length of the SMILES sequence. Each character in the SMILES sequence is mapped to an integer using integer encoding, and each sequence is unified to a length of l by truncation or trailing zeros.

[0091] Similar to the definition of drug SMILES sequences, protein FASTA sequences are composed of different amino acids and are defined as follows:

[0092]

[0093] Where P represents the protein FASTA sequence, p i Indicates the i-th amino acid character in the sequence, N P Represents a vocabulary set containing 25 common amino acids, n p Indicates the length of the protein sequence. Use integer encoding to map the characters in each FASTA sequence to integers, and then unify the length of each sequence to h by truncating or padding with zeros.

[0094] The drug 2D contact graph is constructed based on the 2D topology and physicochemical properties of the drug graph. Specifically, the drug graph G D It can be expressed as:

[0095] G D =(V D , E D ) (3)

[0096] Among them, V D represents the drug molecule node set, E D represents the drug molecule edge set, represents node i, represents the edge between node i and node j. represents the node adjacency matrix of GD, where Represents the number of nodes. If there is an edge between node i and node j, then is equal to 1, otherwise 0. represents the node feature matrix, where the node is represented as dimensional vector. The two-dimensional contact map is the processed two-dimensional map data, which refers to That is, the node adjacency matrix. In addition, the two-dimensional input also includes the node feature matrix extracted from the adjacency matrix.

[0097] Similar to the 2D contact map of drugs, a 2D distance map of proteins is constructed based on the distances between amino acid residues and the 2D topology of proteins. Specifically, the protein map G P Defined as:

[0098] G P =(V P , E P ) (4)

[0099] Among them, V P represents the set of protein residues, E P represents a protein edge set. is residue i, is the edge between residue i and residue j. Represents G P The residue adjacency matrix of Represents the number of residues. If the distance between residue i and residue j is less than Elements is equal to 1, otherwise 0. represents the node feature matrix, where the node is represented as dimensional vector.

[0100] The pharmacophore is constructed by extracting the substructure features in the SMILES sequence of each drug. A d F dimensional vector To represent the pharmacophore characteristics of each drug SMILES sequence based on RDKit, F C The presence and frequency of different common pharmacophore features (such as donor, acceptor, aromaticity, etc.) are indicated in Pharmacophore F C The definition is as follows:

[0101] F C ={f1, f2, ..., f i ,...,f 27}, f i ∈N F (5)

[0102] Among them, f i Indicates the frequency of occurrence of the i-th specific pharmacophore feature identifier, N F Represents a collection of 27 common pharmacophores.

[0103] According to the UniProtID of the FASTA sequence, the corresponding protein three-dimensional data T is downloaded from the AlphaFold protein database to construct the protein three-dimensional structure diagram X T First, calculate the interatomic distance Ad:

[0104] Ad=(Ad max +Ad min +Ad x，y )×α (6)

[0105] Among them, Ad represents the calculation of a specific set of distance metrics in protein residues, Ad max Represents the maximum distance between all atomic pairs, Ad min Represents the minimum distance between all atomic pairs, Ad x，y represents the last distance between atoms x and y, x∈{CA, O, N}, y∈{O, N, C}, α represents the scaling factor. Next, the four main dihedral angles in the protein residues are calculated, expressed as the interatomic rotation angle Ar:

[0106] Ar=(Ar φ +Ar ψ +Ar ω +Ar χ )×β (7)

[0107] Among them, Ar φ represents the rotation angle between the pre-spinal nitrogen atom and the α carbon atom, Ar ψ The rotation angle between the α carbon atom of the spine and the carbonyl group, Ar ω represents the rotation angle between peptide bonds (representing an important feature of protein secondary structure), Ar χThe rotation angle of the side chain atoms relative to the spine (representing an important feature of the protein tertiary structure) and β represent the scaling ratio. Td :

[0108] X Td =onehot(T)+Ad(T)+Ar(T) (8)

[0109] Among them, onehot represents integer mapping encoding. Td , migrated the embedding vector of the pre-trained ESM-2 model (with 6M parameters) As a supplement to the three-dimensional structure of proteins, n p represents the length of the protein pocket, d p Represents the embedding dimension of the protein. By splicing X Td and X Te Get the final node features of the protein:

[0110] X Tnode =X Td ||X Te (9)

[0111] Finally, according to the node feature X of the protein Tnode Create a 3D protein structure diagramX T , and add edge information features (determined by distance threshold, inter-atom connectivity, distance between alpha carbon atoms, and distance between residue centers). Nodes are filtered based on the three-dimensional position information of the convex hull pockets and a subgraph of residues within the protein is created. Laplacian position encoding is applied to the subgraph to enhance the structural information of the graph.

[0112] Performing feature extraction on the preset characterization data of the drug and the protein respectively to obtain feature extraction results of the preset characterization data of the drug and the protein;

[0113] Furthermore, this embodiment performs feature extraction on the preset characterization data of drugs and proteins respectively based on the feature extraction module;

[0114] The feature extraction module includes: embedding submodule, GFAN submodule, linear projection submodule and graph transformer submodule;

[0115] Embedding submodule, used to extract features from drug SMILES sequences and protein FASTA sequences respectively;

[0116] Feature extraction of drug SMILES sequences includes:

[0117] Embed the drug SMILES sequence to obtain the high-dimensional features of the drug;

[0118] Position encoding of drug SMILES sequences to obtain drug position features;

[0119] Add the high-dimensional features of the drug and the position features of the drug to obtain the feature extraction results of the drug SMILES sequence;

[0120] Feature extraction of protein FASTA sequences includes:

[0121] Embed the protein FASTA sequence to obtain the high-dimensional features of the protein;

[0122] Perform position encoding on the protein FASTA sequence to obtain the positional characteristics of the protein;

[0123] Add the high-dimensional features of the protein and the positional features of the protein to obtain the feature extraction results of the protein FASTA sequence;

[0124] The GFAN submodule is used to extract features from the drug 2D contact map and protein 2D distance map respectively;

[0125] Feature extraction of drug two-dimensional contact map includes:

[0126] The drug two-dimensional contact graph node adjacency matrix and node feature matrix Input to the multi-head graph attention network to obtain output in, represents the number of nodes in the two-dimensional contact graph of the drug, Represents the characteristic dimension of each node in the drug molecule, Represent the feature vector space of nodes in the drug graph;

[0127] The output of the multi-head graph attention network After two GCNs, two linear layers, and one adaptive pooling layer, the feature extraction results of the drug two-dimensional contact map are obtained;

[0128] Feature extraction of protein two-dimensional distance map includes:

[0129] The node adjacency matrix of the protein two-dimensional distance graph and node feature matrix Input to the multi-head graph attention network to obtain output in, represents the characteristic dimension of each node in the protein, represents the number of nodes in the two-dimensional distance graph of proteins, A feature vector space representing nodes in a protein graph;

[0130] Then After two GCNs, two linear layers, and one adaptive pooling layer, the feature extraction result of the protein two-dimensional distance map is obtained;

[0131] Linear projection submodule, used for feature extraction of pharmacophores:

[0132] The pharmacophore is projected to a high dimension through a linear layer, each layer is normalized and the nonlinear activation function ReLU is performed, and then projected to a low dimension through a linear layer to obtain the feature extraction result of the pharmacophore;

[0133] The Transformer submodule is used to extract features from protein 3D structure graphs:

[0134] Extract node features g in protein three-dimensional structure graph n and edge feature g e , then g n Perform embedding coding, adding position coding and Dropout operations to g e Perform embedded coding operation, and then convert the protein three-dimensional structure diagram X T 、g n and g e As input, the updated g is obtained through multiple layers of Graph Transformer Layer n ; for g n Use dgl.split to reorganize into computable tensors as K and V, and use a trainable weight matrix as Q, which is input into the attention mechanism. The output results are added and averaged to obtain the final feature extraction result of the protein three-dimensional structure.

[0135] Specifically, in this embodiment, the feature extraction module mainly includes an embedding module, a GFAN module, a linear projection module and a graph transformer module.

[0136] (1) Embedded module

[0137] like Figure 2 As shown in Figure 3, the embedding module mainly consists of embedding encoding and position embedding.

[0138] The SMILES sequence after drug preprocessing is used as input. First, the sequence is embedded and encoded to obtain the high-dimensional feature E of the drug. D ∈ R l×d , where d represents the word embedding dimension of the drug. Then, the sequence is positionally encoded to obtain the drug’s position feature P D ∈R l×d , the formula is defined as

[0139]

[0140]

[0141] Among them, i is the position of the molecule in the sequence, and j is the index of the word embedding dimension. Finally, the embedding feature E D With position feature P D Add together to get the final result X of drug sequence feature extraction SD :

[0142] X SD =E D +P D (12)

[0143] The FASTA sequence feature extraction method for proteins is the same as that for drugs. First, the sequence is embedded and encoded to obtain the high-dimensional feature E of the protein. P ∈R h×d Then, the sequence is positionally encoded to obtain the protein position feature P P ∈R h×d . Preferably, embedding feature E P With position feature P P Add together to get the final result X of protein sequence feature extraction SP :

[0144] X SP =E P +P P (13)

[0145] (2) GFAN module

[0146] like Figure 3 As shown in the figure, the Graph Flow Aggregation Network (GFAN) module is mainly composed of a multi-head graph attention network (Multi-Head GAT) and a graph convolutional network (GCN).

[0147] Take the two-dimensional contact graph of the drug as input. First, the node adjacency matrix of the two-dimensional contact graph of the drug is and node feature matrix Input to the multi-head graph attention network to obtain output The graph attention mechanism is used to enhance the model's ability to mine features across the entire graph.

[0148] The two-dimensional contact graph is the adjacency matrix of the drug. If there is an edge between a pair of nodes, it is 1, otherwise it is 0, so its size is The node feature matrix is a drug molecule A set of nodes, where each node is dimensional word embedding vector. So the size of the node feature matrix is It is extracted from the 2D contact map.

[0149] Multi-Head GAT consists of K graph attention layers (GATs). Each GAT first receives the node feature matrix X of the two-dimensional contact graph D As input, through a trainable weight matrix X D Convert to a feature matrix applicable to all nodes

[0150] h=X D W1 (14)

[0151] Then construct the source node feature of each edge according to the index of the non-zero elements of h and target node features Through the trainable weight matrix h i and h j Conversion, by splicing the converted h i and h j As input, to calculate the attention coefficient

[0152]

[0153]

[0154] σ2=max(0,x) (17)

[0155] Where σ1(·) represents the ReLU activation function and dropout(·) represents the random inactivation operation. ij and target node feature h j Update source node characteristics Get the output of the graph attention layer k=1,2,...,K, the calculation formula is as follows:

[0156] h i =a ij W3h j (18)

[0157] in Represents a trainable weight matrix. Finally, the results of the K graph attention layers are added and averaged and the output of Multi-Head GAT is obtained by the activation function ELU. The formula is defined as follows:

[0158]

[0159]

[0160] Where σ2(·) represents the ELU activation function. Then After two GCNs, two linear layers, and one adaptive pooling layer, the final feature extraction result of the two-dimensional contact map is obtained. The graph convolution mechanism is used to enhance the model's ability to mine local range features. The formula of the graph convolution network is defined as follows:

[0161]

[0162]

[0163] in Represented by the adjacency matrix of the two-dimensional contact graph of the drug The degree matrix calculated, I represents the degree matrix obtained by The self-loop matrix of each node, W l-1 Represents the weight matrix of the l-1th layer GCN. After two GCNs, two linear layers and one adaptive pooling layer, the final feature extraction result of the two-dimensional contact map is obtained.

[0164] Similar to drugs, the two-dimensional distance graph of proteins also passes through the GFAN module to obtain the final feature extraction results of the two-dimensional distance graph of proteins. First, the node feature matrix of the two-dimensional distance graph of proteins is Input to the multi-head graph attention network to obtain output Then After two GCNs, two linear layers and one adaptive pooling layer, the final feature extraction result of the two-dimensional distance map is obtained.

[0165] (3) Linear projection layer

[0166] The linear projection layer mainly consists of a linear layer, a layer normalization layer, and a ReLU activation function. First, the pharmacophore feature F C After projecting to high dimensions through a linear layer, normalizing and nonlinear activation function ReLU are performed on each layer, and then projecting to low dimensions through a linear layer, the final feature extraction result of the pharmacophore is obtained.

[0167] (4)Graph Tranformer module

[0168] like Figure 4 As shown in the figure, the Graph Transfromer module applies Transformer specifically to process three-dimensional structure graph data to maintain the geometric properties of the three-dimensional graph, such as rotation and translation invariance.

[0169] First, extract the protein three-dimensional structure X T Node feature g in n and edge feature g e , then g n Perform embedding coding, adding position coding and Dropout operations to g e Perform embedded coding operation, then X T 、g n and g e As input, the updated g is obtained through multiple layers of Graph Transformer Layer n . n Use dgl.split to reorganize into computable tensors as K and V, and input a trainable weight matrix as Q into the attention mechanism, and add the output results and take the average value to obtain the final protein three-dimensional structure feature extraction result X TP , the attention calculation formula is as follows:

[0170]

[0171] The structure of Graph Transfromer Layer is as follows Figure 5 As shown:

[0172] The process of Graph Transfromer Layer is as follows: First, extract the protein three-dimensional structure graph X T The node features in the attention model are input as Q and K, and the edge features as V. The results are scaled and biased to separate the node and edge features. The node and edge features calculated after attention are then reshaped, dropout, and linearly transformed. Together with the input node and edge features, they are passed through two feed-forward layers to obtain the final node and edge features.

[0173] The feature extraction results of drugs and protein are fused respectively to obtain drug fusion features and protein fusion features;

[0174] Specifically, in this embodiment, after feature extraction, the one-dimensional sequence features X of the drugs are obtained. SC , two-dimensional contact map feature X GC , pharmacophore characteristics X DC and the one-dimensional sequence feature X of the protein SP , two-dimensional contact map feature X GP , three-dimensional structural features X TP , then Figure 1 As shown in the figure, the features of different modalities of drugs and proteins are modally fused to obtain the output result X after modal fusion.C With X P . GC With X DC Add them as K and V, and X SC As Q, the intramolecular multi-head attention module is used for modality fusion. GP With X TP Weighted fusion as K and V, X SP As Q, an intra-molecular multi-head attention module is adopted for modality fusion.

[0175] After Q, K, and V are input into the scaled dot product attention, the attention calculation coefficient of each head is obtained. The calculation results of each head are concatenated and passed through two feedforward layers to obtain the final drug and protein modality fusion feature X. C With X P The feed-forward layer includes residual connection, linear transformation and normalization.

[0176] A joint attention mechanism is used to interactively process drug fusion features and protein fusion features to obtain interactive features;

[0177] Specifically, in this embodiment, after modality fusion, the joint attention mechanism is used to extract drug features X C with protein feature X P Interaction result X finaL The structure of the joint attention mechanism is as follows Figure 6 shown.

[0178] First, X C With X P Divide into two parts, namely X C1 With X P1 、X C2 With X P2 , where X C1 With X P1 They are activated by ReLU respectively, and then the Einstein summation operation and the second nonlinear activation by Sigmoid are performed to obtain X cp , Einstein's summation formula is as follows:

[0179]

[0180] The index of the output matrix is "ij", the index of the input matrix A is "ik", and the index of the input matrix B is "kj". The convention is "ik, kj→ij". Here, the convention is "bij, bkj→bik". Then, for X cp Perform Einstein summation and take the reciprocal, agreeing on "bij→b", and X cp1 Perform Einstein summation again to get the interaction result X of the first channel cp1, agreed to "bij, b→bij". X C2 With X P2 The interaction result X of the first channel is obtained by directly performing Einstein summation and nonlinear activation through Tanh cp2 , agreed to "bij, bkj→bikj". The interaction result of the two channels X cp1 With X cp2 Once again, the Einstein summation operation is performed to obtain the final interaction feature X final , the convention is "bijk, bij→bk".

[0181] Based on the interaction characteristics, the binding affinity between drugs and proteins is predicted.

[0182] Specifically, in this embodiment, using the prediction module to predict the binding affinity between the drug and the protein includes:

[0183] The prediction module is a multi-layer perceptron consisting of a series of linear layers, normalization layers, Dropout operations, and ReLU activation functions. First, X final After adjusting to 2048 dimensions through the linear layer, the data size is adjusted through the normalization layer, overfitting is reduced through Dropout, nonlinear activation is performed through a layer of ReLU, and then adjusted to 512 dimensions through the linear layer. Then the data size is adjusted through the normalization layer, overfitting is reduced through Dropout, nonlinear activation is performed through a layer of ReLU, and finally adjusted to 1 dimension through the linear layer to obtain the prediction result.

[0184] This embodiment proposes a DTA prediction model based on multimodal collaborative generation. First, drugs and proteins are processed into one-dimensional sequence representation, two-dimensional contact (or distance) graph representation, pharmacophore representation and protein three-dimensional structure graph representation. Then, the multi-channel module is used to extract multiple perspective features of drugs and proteins respectively, so as to achieve the synchronous capture of drug and protein sequence and structural features. Then, the intramolecular multi-head attention mechanism is used to fuse the multiple modal features of drugs and proteins, making full use of their sequence and structural information. Subsequently, the interactive fusion of drug and protein features is performed through the joint attention module to more accurately simulate the interaction between them. Finally, the DTA prediction result is obtained through a multi-layer perceptron prediction module composed of multiple layers of fully connected layers.

[0185] The prediction model proposed in this embodiment consists of four parts: feature extraction module, modality fusion module, interaction module and prediction module. Figure 1 shown.

[0186] (1) Feature extraction module

[0187] The drug's SMILES sequence, two-dimensional contact map, and pharmacophore are respectively processed through the embedding module, GFAN module, and linear projection layer to extract one-dimensional and two-dimensional features.

[0188] The protein's FASTA sequence, two-dimensional distance graph, and three-dimensional structure graph are respectively extracted through the embedding module, GFAN module, and graph Transformer module to extract one-dimensional, two-dimensional, and three-dimensional features.

[0189] (2) Modal fusion module

[0190] Drug feature fusion: The pharmacophore features are added to the two-dimensional features of the drug, and then modally fused with the one-dimensional features of the drug through the intramolecular multi-head attention module to obtain the final features of the drug.

[0191] Protein feature fusion: The three-dimensional features of the protein are weightedly fused with the two-dimensional features, and then modally fused with the one-dimensional features of the protein through the intramolecular multi-head attention module to obtain the final features of the protein.

[0192] (3) Interaction module

[0193] Through the intermolecular joint attention module, the interaction relationship between the final features of drugs and proteins is further mined and learned.

[0194] (4) Prediction module

[0195] Apply the predictor module to process the learned features and obtain the binding affinity prediction results between drugs and proteins.

[0196] Furthermore, this example also uses two common benchmark datasets, KIBA and Davis, to evaluate the model of this example. The Davis dataset contains 442 kinase proteins and 68 inhibitors (drugs), forming 30,056 drug-target pairs, which are characterized by the dissociation constant (K d ) value measurement. The KIBA dataset contains 229 proteins and 2111 drugs, forming 118,254 drug-target pairs. The data is divided into training set and test set data according to 5 / 1, and the data is randomly divided 5 times. Based on the actual value of the affinity of the test set and the predicted value of the model output, the DTA prediction evaluation indicators - mean square error (MSE), consistency index (CI), regression mean And the standard deviation (std) of these three indicators. After calculation, in the Davis data set, the results of the model in this embodiment are MSE (0.248), CI (0.880), (0.667). In the KIBA data set, the results of the model in this embodiment are MSE (0.184), CI (0.857), (0.726). This embodiment's performance was compared with some existing DTA prediction models, as shown in Tables 1 and 2. The results show that the DTA prediction method of this embodiment outperforms the other methods in the table, and the lower std value indicates that the model of this embodiment has high stability.

[0197] Table 1 Comparison of results on the Davis dataset

[0198]

[0199] Table 2 Comparison of results on the KIBA dataset

[0200]

[0201] The above embodiments are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A drug-target affinity prediction method based on multimodal synergistic generation, characterized in that: include: Preprocessing the drug and protein separately to obtain preset characterization data of the drug and protein; wherein the preset characterization data of the drug includes: drug SMILES sequence, drug two-dimensional contact map, and pharmacophore; the preset characterization data of the protein includes: protein FASTA sequence, protein two-dimensional distance map, and protein three-dimensional structure map; Performing feature extraction on the preset characterization data of the drug and the protein respectively to obtain feature extraction results of the preset characterization data of the drug and the protein; Perform feature extraction on the three-dimensional structure diagram of the protein: Extract the node features g in the protein three-dimensional structure graph n and edge feature g e , then g n Perform embedding coding, adding position coding and Dropout operations to g e Perform embedding coding operation, and then convert the protein three-dimensional structure diagram X T 、g n and g e As input, the updated g is obtained through multiple layers of Graph Transformer Layer n ; for g n Use dgl.split to reorganize into computable tensors as K and V, and use a trainable weight matrix as Q, input into the attention mechanism, and add the output results and take the average value to obtain the final protein three-dimensional structure feature extraction result three-dimensional structure feature X TP ; The feature extraction results of drugs and proteins are fused to obtain the drug fusion feature X C and protein fusion feature X P ; The joint attention mechanism is used to fusion feature X C and protein fusion feature X P Perform interaction processing and obtain interaction features; The drug fusion feature X C and protein fusion feature X P Perform interaction processing and obtain interaction features including: The drug is fused with feature X C and protein fusion feature X P Divide into two parts, namely X C1 With X P1 、X C2 With X P2 ; X C1 With X P1 They are activated by ReLU respectively, and then the Einstein summation operation and the second nonlinear activation by Sigmoid are performed to obtain X cp , then X cp Perform Einstein summation and take the reciprocal, and compare it with X cp1 Perform Einstein summation again to get the interaction result X of the first channel cp1 ; X C2 With X P2 The interaction result X of the first channel is obtained by directly performing Einstein summation and nonlinear activation through Tanh cp2 ; The interaction result X of the two channels cp1 With X cp2 Once again, the Einstein summation operation is performed to obtain the final interaction feature X final ; Based on the interaction characteristics, the binding affinity between the drug and the protein is predicted.

2. The drug-target affinity prediction method based on multimodal synergistic generation according to claim 1, characterized in that: The drug SMILES sequence is based on different SMILES character compositions; The drug SMILES sequence is: Where D represents the drug SMILES sequence, d i Indicates the i-th SMILES character, N D Represents a vocabulary consisting of 62 SMILES characters, m d Indicates the length of the SMILES sequence; The protein FASTA sequences are based on different amino acid compositions; The protein FASTA sequence is: Where P represents the protein FASTA sequence, p i Indicates the i-th amino acid character in the sequence, N P Represents a vocabulary set containing preset amino acids, n p Indicates the sequence length of the protein; The two-dimensional drug contact map is constructed based on the two-dimensional topology of the drug map; The drug diagram is as follows: G D =(V D ,E D ) Among them, V D represents the drug molecule node set, E D represents the drug molecule edge set, G D Represents a drug graph; The two-dimensional contact map of the drug: in, represents the two-dimensional contact graph of drugs, that is, the node adjacency matrix of the drug graph, Represents the number of nodes. If there is an edge between node i and node j, then is equal to 1, otherwise 0; Represents the node feature matrix of the drug graph, where the nodes are represented as dimensional vector; The protein two-dimensional distance map is constructed based on the distances between amino acid residues and the two-dimensional topology of the protein; The protein map is: G P =(V P ,E P ) Among them, V P represents the set of protein residues, E P represents the protein edge set, G P represents a protein map; The protein two-dimensional distance map is: in, represents the two-dimensional distance graph of the protein, i.e., the residue adjacency matrix of the protein graph, |V P | represents the number of residues; if the distance between residue i and residue j is less than Elements is equal to 1, otherwise 0; Represents the node feature matrix of the protein graph, where the nodes are represented as dimensional vector; The pharmacophore is constructed by extracting substructure features from the SMILES sequence of each drug; The pharmacophore is: F C ={f1,f2,…,f i ,…,f 27 },f i ∈N F Among them, f i Indicates the frequency of occurrence of the i-th specific pharmacophore feature identifier, N F represents a collection of 27 common pharmacophores, F C represents the pharmacophore; The protein three-dimensional structure diagram is based on the node feature X of the protein Tnode Create and add side information features; wherein the side information features include: distance threshold, connectivity between atoms, distance between α carbon atoms and distance between residue centers.

3. The drug-target affinity prediction method based on multimodal synergistic generation according to claim 2, characterized in that: Get the node feature X of the protein Tnode include: Calculate interatomic distances; Calculate dihedral information in protein residues, expressed as interatomic rotation angles; Based on the interatomic distances and interatomic rotation angles, the three-dimensional structural features of the protein are encoded. Td ; Transfer the embedding vectors of the pre-trained ESM-2 model As a supplement to the three-dimensional structure of proteins; p represents the length of the protein pocket, d p represents the embedding dimension of the protein; By splicing X Td and X Te , obtain the node features of the protein.

4. The drug-target affinity prediction method based on multimodal synergistic generation according to claim 1, characterized in that: Feature extraction for the preset characterization data of drugs and proteins includes: Feature extraction of drug SMILES sequences: Embed the drug SMILES sequence to obtain the high-dimensional features of the drug; Position encoding of drug SMILES sequences to obtain drug position features; Add the high-dimensional features of the drug and the position features of the drug to obtain the one-dimensional sequence feature X of the drug SMILES sequence feature extraction result. SD ; Feature extraction of protein FASTA sequences: Embed the protein FASTA sequence to obtain the high-dimensional features of the protein; Perform position encoding on the protein FASTA sequence to obtain the positional characteristics of the protein; Add the high-dimensional features of the protein and the positional features of the protein to obtain the feature extraction result of the protein FASTA sequence, and obtain the one-dimensional sequence feature X SP ; Perform feature extraction on the two-dimensional contact map of the drug: The node adjacency matrix of the drug two-dimensional contact graph and node feature matrix Input to the multi-head graph attention network to obtain output in, represents the number of nodes in the two-dimensional contact graph of the drug, Represents the characteristic dimension of each node in the drug molecule, Represent the feature vector space of nodes in the drug graph; The output of the multi-head graph attention network After two GCNs, two linear layers and one adaptive pooling layer, the feature extraction result of the two-dimensional contact map of the drug is obtained. The two-dimensional contact map feature X GC ; Extracting features from the two-dimensional protein distance map includes: The node adjacency matrix of the protein two-dimensional distance graph and node feature matrix Input to the multi-head graph attention network to obtain output in, represents the characteristic dimension of each node in the protein, represents the number of nodes in the two-dimensional distance graph of proteins, A feature vector space representing nodes in a protein graph; Then After two GCNs, two linear layers and one adaptive pooling layer, the feature extraction result of the two-dimensional distance map of the protein is obtained, and the two-dimensional contact map feature X is obtained. GP ; Feature extraction of the pharmacophore: The pharmacophore is projected to a high dimension through a linear layer, each layer is normalized and subjected to a nonlinear activation function ReLU, and then projected to a low dimension through a linear layer to obtain the feature extraction result of the pharmacophore, that is, the pharmacophore feature.

5. The drug-target affinity prediction method based on multimodal synergistic generation according to claim 4, characterized in that: Obtaining drug fusion signatures and protein fusion signatures includes: X GC With X DC Add them as K and V, and X SC As Q, an intra-molecular multi-head attention mechanism is used for modality fusion; X GP With X TP Weighted fusion as K and V, X SP As Q, an intra-molecular multi-head attention mechanism is used for modality fusion; After inputting all Q, K, and V into the scaled dot product attention, the attention calculation coefficient of each head is obtained. After splicing the calculation results of each head, two feedforward operations are performed to obtain the final drug fusion feature X. C and protein fusion feature X P .

6. The drug-target affinity prediction method based on multimodal synergistic generation according to claim 1, characterized in that: Predicting the binding affinity between drugs and proteins includes: The interaction features are linearly adjusted to a first preset dimension, then normalized to adjust the data size, Dropout to reduce overfitting, ReLU to perform nonlinear activation, linearly adjusted to a second preset dimension, then normalized to adjust the data size again, Dropout to reduce overfitting, ReLU to perform nonlinear activation, and finally linearly adjusted to the final dimension to obtain the prediction result.

Images