Drug-target affinity prediction method and system based on gene ontology guidance and multi-modal attention

Abstract

The invention relates to a drug-target affinity prediction method and system based on gene ontology guidance and multi-modal attention. The method comprises the following steps: complementing gene ontology GO function annotation of target protein; gO feature representation is extracted, sequence features are extracted through an ESM-2 model, and fusion is carried out through a gating mechanism to obtain protein features; using a Molformer model to obtain drug sequence features, converting an SMILES sequence of a drug molecule into a graph structure, using Transformer to extract topological structure features, and fusing to obtain drug features; a double-branch multi-head cross attention mechanism is used to realize bidirectional interaction of protein features and drug features, a Mama layer is used to perform deep feature extraction, and then prediction is performed. By extracting protein features and drug features, complex interaction between a drug and a target spot can be effectively captured; and a head cross attention mechanism is combined with a Mama layer to carry out deep feature extraction, so that the feature expression capability and the prediction performance are enhanced.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a method and system for predicting drug-target affinity based on gene ontology guidance and multimodal attention. Background Technology

[0002] Drug development is characterized by long cycles, high costs, and extremely high failure rates. Traditional experimental methods such as X-ray crystallography, nuclear magnetic resonance, and high-throughput screening, while highly accurate, are generally time-consuming, expensive, and difficult to scale up. To reduce costs and improve development efficiency, computer-aided drug discovery technology has developed rapidly in recent years. Among these methods, drug-target affinity prediction (DTA), as a key step in virtual drug screening, can effectively improve the efficiency and accuracy of drug screening and has become a research hotspot in the field of drug development.

[0003] Existing DTA prediction methods can be broadly classified into three categories: The first category is sequence-based prediction methods, which utilize the drug's smils string or molecular fingerprint and the target's amino acid sequence to directly extract features from the original sequence using a one-dimensional convolutional neural network or Transformer model; the second category is graph-based prediction methods, which represent drug molecules as molecular graphs and use graph neural networks to extract chemical topological features. Some methods also utilize protein structure diagrams or residue contact diagrams to describe the spatial features of proteins; the third category is prediction methods that combine three-dimensional structural information. By introducing the three-dimensional structure of protein binding pockets or complexes, structural information is used to model intermolecular interaction mechanisms in order to improve prediction accuracy and interpretability.

[0004] However, traditional DTA prediction methods rely solely on sequence or structural information to characterize targets, lacking explicit modeling of protein biological functions. This results in limited performance of the models when faced with proteins of unknown structure, complex function, or insufficient structural information, making it difficult to accurately locate function-related binding regions. Therefore, traditional DTA prediction methods, due to their over-reliance on sequence or structural feature extraction, suffer from deficiencies in generalization and robustness, often resulting in low prediction accuracy. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, a drug-target affinity prediction method and system based on gene ontology guidance and multimodal attention is provided. This method can overcome the limitations of traditional methods in feature extraction and prediction performance, and achieve efficient and accurate prediction of drug-target affinity.

[0006] A drug-target affinity prediction method based on gene ontology guidance and multimodal attention, the method comprising:

[0007] Collect affinity datasets containing drug molecular structures and target protein sequences, and perform completion operations on target proteins in the dataset that lack gene ontology GO functional annotations.

[0008] GO functional annotations of target proteins in the dataset are extracted, and GO feature representations are obtained by encoding the GO functional annotations and then extracting them using CNN. Sequence features are extracted from the target protein sequences using the ESM-2 model. The GO feature representations are fused with the sequence features through a gating mechanism to obtain protein features.

[0009] Based on the drug molecule structures in the dataset, the SMILES sequences of the drug molecules are encoded using a Molformer model and then extracted using a CNN to obtain drug sequence features; the SMILES sequences of the drug molecules are converted into a graph structure, and topological features in the graph structure are extracted using a Transformer-based graph convolutional network; the drug sequence features and the topological features are fused to obtain the drug features.

[0010] The protein features and the drug features are respectively input into a two-branch multi-head cross-attention mechanism to achieve bidirectional interaction between the protein features and the drug features. A Mamba layer is used for deep feature extraction, and multimodal fusion features are obtained through splicing, pooling, and weighted fusion.

[0011] The regression module based on the multilayer perceptron performs regression prediction on the multimodal fusion features to obtain the binding affinity between the drug and the target protein.

[0012] In one embodiment, the completion operation for target proteins lacking GO functional annotations in the dataset includes:

[0013] Locate the target proteins in the dataset that lack GO functional annotations of the gene ontology;

[0014] Retrieve gene ontology GO functional annotations for each target protein from the UniProt database;

[0015] Use the NetGO tool to complete the GO functional annotations of genes with missing target proteins.

[0016] In one embodiment, the GO function annotation is encoded and then the GO feature representation is extracted using a CNN, including:

[0017] The GO feature annotations are encoded using the pre-trained model BlueBERT to obtain GO feature representations.

[0018] The GO feature representation is input into a CNN model, which captures the local and global dependencies between GO function annotations. After convolution processing, the GO feature representation is obtained.

[0019] In one embodiment, the ESM-2 model is used to extract sequence features from the target protein sequence; the GO feature representation is fused with the sequence features through a gating mechanism to obtain protein features, including:

[0020] The ESM-2 model was used to extract sequence features from the target protein sequence and then convolutional processing was performed to obtain the final sequence features.

[0021] The GO feature representation is average pooled along the annotation dimension to obtain a compact vector, and a gate control vector is calculated based on the compact vector through linear transformation and Sigmoid activation;

[0022] The gating vector is applied to the final sequence features for fusion to obtain protein features.

[0023] In one embodiment, the SMILES sequence of the drug molecule is converted into a graph structure, and a Transformer-based graph convolutional network is used to extract the topological features of the graph structure, including:

[0024] The SMILES sequence of the drug molecule was converted into a graph structure using RDKit, and the graph structure was then processed by feature encoding using TransConv in a Transformer-based graph convolutional network.

[0025] Based on the encoded graph structure, attention mechanism is used to aggregate neighbor information, introduce edge features, combine nonlinear activation and batch normalization, and form the input tensor of all node features in the encoded graph structure through zero padding to obtain the topological features in the graph structure.

[0026] In one embodiment, the bidirectional interaction between the protein features and the drug features is achieved through the dual-branch multi-head cross-attention mechanism, and deep feature extraction is performed using a Mamba layer, including:

[0027] A linear transformation is performed on the protein features and the drug features to obtain the corresponding query, key, and value vectors.

[0028] Based on the query, key, and value vectors, cross-attention calculation is performed on the protein features and the drug sequence features to obtain a first fusion result; based on the query, key, and value vectors, cross-attention calculation is performed on the protein features and the topological structure features to obtain a second fusion result;

[0029] The first fusion result and the second fusion result are input into the Mamba layer based on the structured state-space model for deep feature extraction.

[0030] In one embodiment, the method further includes:

[0031] The structured state space model includes a structured state space model module, a fully connected layer, a convolutional layer, and a pooling layer.

[0032] Transformation functions are defined and activation functions are determined based on the structured state space model module, fully connected layer, convolutional layer, and pooling layer.

[0033] Using the transformation function and activation function, deep feature extraction is performed on the data input to the structured state space model to obtain deep protein features and deep drug features.

[0034] In one embodiment, multimodal fusion features are obtained through splicing, pooling, and weighted fusion, including:

[0035] The deep protein features and deep drug features are spliced ​​together using modal features, and max pooling is used to obtain the fused representation of protein and drug sequence features and the fused representation of protein and drug graph features.

[0036] The average values ​​of the protein-drug sequence feature fusion representation and the protein-drug map feature fusion representation are calculated respectively, and the ratio of the average values ​​is used as an adjustment coefficient.

[0037] The aforementioned adjustment coefficients are used to weight the deep features of proteins and drugs, and a fusion layer is introduced for nonlinear interaction to obtain multimodal fusion features.

[0038] In one embodiment, the regression module based on a multilayer perceptron includes multilayer linear transformation and nonlinear activation, and is trained by minimizing the mean squared error loss.

[0039] A drug-target affinity prediction system based on gene ontology guidance and multimodal attention, the system comprising:

[0040] The data collection module is used to collect affinity datasets containing drug molecule structures and target protein sequences, and to complete the target proteins in the dataset that lack gene ontology GO functional annotations.

[0041] The protein feature extraction module is used to extract the GO functional annotations of target proteins in the dataset, encode the GO functional annotations, and extract GO feature representations through CNN; extract sequence features from the target protein sequence using the ESM-2 model; and fuse the GO feature representations with the sequence features through a gating mechanism to obtain protein features.

[0042] The drug feature extraction module is used to encode the SMILES sequences of drug molecules using a Molformer model based on the drug molecule structure in the dataset, and then extract drug sequence features through a CNN; convert the SMILES sequences of drug molecules into a graph structure, and extract the topological structure features in the graph structure using a Transformer-based graph convolutional network; and fuse the drug sequence features with the topological structure features to obtain the drug features.

[0043] The multimodal feature fusion module is used to input the protein features and the drug features into the two-branch multi-head cross-attention mechanism, respectively. The two-branch multi-head cross-attention mechanism realizes the bidirectional interaction between the protein features and the drug features, and uses the Mamba layer to perform deep feature extraction. The multimodal fused features are obtained by splicing, pooling and weighted fusion.

[0044] The regression prediction module is used to perform regression prediction on the multimodal fusion features based on the multilayer perceptron regression module to obtain the binding affinity between the drug and the target protein.

[0045] The aforementioned drug-target affinity prediction method and system based on gene ontology guidance and multimodal attention effectively captures the complex interactions between drugs and targets by extracting protein and drug features through GO functional annotation completion of target proteins. It employs a multi-head cross-attention mechanism combined with Mamba layers for deep feature extraction, enhancing feature expression and prediction performance. Furthermore, the use of a multilayer perceptron regression module significantly improves prediction performance, achieving efficient and accurate prediction of drug-target affinity. Attached Figure Description

[0046] Figure 1 This is an application environment diagram of a drug-target affinity prediction method based on gene ontology guidance and multimodal attention in one embodiment;

[0047] Figure 2 This is a flowchart illustrating a drug-target affinity prediction method based on gene ontology guidance and multimodal attention in one embodiment.

[0048] Figure 3This is a schematic diagram of the GoMA-DTA model structure for implementing a drug-target affinity prediction method based on gene ontology guidance and multimodal attention in one embodiment.

[0049] Figure 4 This is a block diagram of a drug-target affinity prediction system based on gene ontology guidance and multimodal attention in one embodiment;

[0050] Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] The posture data method provided in this application is based on a drug-target affinity prediction method guided by gene ontology and multimodal attention, and can be applied to, for example... Figure 1 The application environment shown. For example... Figure 1As shown, the application environment includes computer device 110. Computer device 110 can collect affinity datasets containing drug molecule structures and target protein sequences, and perform completion operations on target proteins lacking GO functional annotations in the dataset; computer device 110 can extract GO functional annotations of target proteins from the dataset, encode the GO functional annotations, and then extract GO feature representations using a CNN; it uses an ESM-2 model to extract sequence features from the target protein sequences; it fuses the GO feature representations with the sequence features through a gating mechanism to obtain protein features; computer device 110 can, based on the drug molecule structures in the dataset, encode the SMILES sequences of drug molecules using a Molformer model and then extract drug sequences using a CNN. The process involves several steps: first, converting the SMILES sequence of the drug molecule into a graph structure; then, extracting topological features from the graph structure using a Transformer-based graph convolutional network; finally, fusing the drug sequence features with the topological features to obtain the drug features; second, inputting the protein features and drug features into a two-branch multi-head cross-attention mechanism to achieve bidirectional interaction between the protein and drug features, and using a Mamba layer for deep feature extraction; finally, obtaining multimodal fusion features through concatenation, pooling, and weighted fusion; and third, using a multilayer perceptron regression module to perform regression prediction on the multimodal fusion features to obtain the binding affinity between the drug and the target protein. The computer device 110 can be, but is not limited to, various personal computers, laptops, smartphones, robots, tablets, etc.

[0053] In one embodiment, such as Figure 2 As shown, a drug-target affinity prediction method based on gene ontology guidance and multimodal attention is provided, including the following steps:

[0054] Step 202: Collect affinity datasets containing drug molecule structures and target protein sequences, and perform completion operations on target proteins that lack gene ontology GO functional annotations in the dataset.

[0055] Computer equipment can collect affinity datasets containing drug molecule structures and target protein sequences for later use.

[0056] Specifically, in one embodiment, the provided drug-target affinity prediction method based on gene ontology guidance and multimodal attention may further include a completion operation, which includes: finding target proteins in the dataset that lack gene ontology GO functional annotations; retrieving gene ontology GO functional annotations for each target protein from the UniProt database; and using the NetGO tool to complete the missing gene ontology GO functional annotations for the target proteins.

[0057] Computer devices can retrieve Gene Ontology (GO) functional annotation IDs for target proteins from the UniProt database. For any missing GO annotations, tools such as NetGO can be used to complete them, ensuring the integrity of protein functional information and providing a reliable data foundation for subsequent protein functional feature extraction. Specifically, for missing annotations, computer devices can use NetGO 4.0 for prediction and supplementation.

[0058] Step 204: Extract the GO functional annotations of the target proteins in the dataset, encode the GO functional annotations, and then extract the GO feature representations using CNN; use the ESM-2 model to extract sequence features from the target protein sequences; and fuse the GO feature representations with the sequence features through a gating mechanism to obtain the protein features.

[0059] In protein feature extraction, the computer equipment can first obtain the GO functional description of the protein using UniProt and NetGO 4.0 and encode it with BlueBERT. Then, a CNN is used to extract deep GO representations. Simultaneously, a pre-trained ESM-2 model is used to extract sequence features from the amino acid sequence, obtaining sequence feature representations containing rich biological information, which are then convolved. Finally, a gating mechanism is used to fuse the GO information with the sequence features to obtain a more comprehensive protein feature representation. Functional annotation is encoded using a specially trained BlueBERT model, which can generate semantically meaningful feature vectors.

[0060] In one embodiment, a drug-target affinity prediction method based on gene ontology guidance and multimodal attention may further include a process of extracting GO feature representations. The specific process includes: encoding GO function annotations using a pre-trained model BlueBERT to obtain GO feature representations; inputting the GO feature representations into a CNN model, capturing local and global dependencies between GO function annotations through the CNN model, and performing convolution processing to obtain GO feature representations.

[0061] In this embodiment, for protein feature extraction, the GO annotations of the target protein can first be obtained from the UniProt database, and the missing information can be completed using NetGO4.0; then, the corresponding GO functional descriptions can be downloaded from the Gene Ontology database, and these descriptions can be encoded using BlueBERT to obtain GO feature representations. Where L go The number of GO annotations for a fixed protein is given, d is the feature dimension, and R indicates that each element in the matrix is ​​a value in the real number field. It is a dimension L goA real matrix of ×d, used to extract deeper semantic representations from GO features, can be processed using CNNs. go Encode the GO features of the l-th layer as follows: Input layer is defined as The formula for calculating the feature representation of the (l+1)th layer is: in, and Here, σ represents the weights and biases of the l-th convolutional kernel, and σ is the activation function. CNNs can capture the local and global dependencies between GO descriptions. After l layers of convolution, a deep GO feature representation is obtained.

[0062] In one embodiment, a drug-target affinity prediction method based on gene ontology guidance and multimodal attention may further include a process of extracting sequence features and performing feature fusion. The specific process includes: extracting sequence features from the target protein sequence using an ESM-2 model and performing convolution processing to obtain the final sequence features; performing average pooling along the annotation dimension on the GO feature representation to obtain a compact vector, and calculating a gate control vector based on the compact vector through linear transformation and Sigmoid activation; and applying the gate control vector to the final sequence features for fusion to obtain protein features.

[0063] Computer equipment can use ESM-2 to extract protein sequence features, and obtain... Where L s Given a fixed sequence length (1024), d is the feature dimension, and R represents that each element in the matrix is ​​a value in the real number field. It is a dimension L s A real matrix of size ×d, the initial input of the sequence encoder is denoted as . Similar to the GO encoder, two CNN layers are used to encode the sequence features. The features of the l-th layer are represented as follows: Its update rules are as follows: in, and These are the learnable parameters of the l-th layer CNN. This design enables the model to capture local motifs and contextual dependencies from protein sequences, thereby obtaining highly discriminative feature representations.

[0064] To integrate functional annotations into sequence-based feature representations, in this embodiment, the GO feature matrix can be first processed. Performing average pooling along the annotation dimension yields a compact vector: L go The vector f represents the number of GO annotations fixed for a protein. go ∈R dThe semantic information of all GO annotations is aggregated. Then, the computer can compute a gating vector: m = σ(Wf) using linear transformation and sigmoid activation. go +b); where W∈R d×d b∈R d The gating vector m is a trainable parameter used to adjust the importance of each dimension of the sequence features.

[0065] Finally, by applying the gating vector to the sequence features, the integrated protein feature representation is obtained. Here, ⊙ represents element-wise (Hadamard) multiplication, m is broadcast along the sequence length dimension, and the fusion mechanism can adaptively utilize GO-based semantic signals to adjust sequence features, thereby enhancing the expressive power of protein feature representation.

[0066] Step 206: Based on the drug molecule structure in the dataset, the SMILES sequence of the drug molecule is encoded using the Molformer model and then extracted using CNN to obtain drug sequence features; the SMILES sequence of the drug molecule is converted into a graph structure, and the topological structure features in the graph structure are extracted using a Transformer-based graph convolutional network; the drug sequence features and the topological structure features are fused to obtain the drug features.

[0067] For drug feature extraction, a dual-channel modeling strategy can be adopted: on the one hand, a pre-trained Molformer model is used to encode the SMILES sequence and then deep-mining local patterns using a CNN to extract the drug sequence features of the molecules; on the other hand, the drug molecules are transformed into a graph structure, introducing atomic attributes and edge features, and a Transformer-based graph convolutional network (TransConv) is used to extract topological features. The features from both channels are then deep-encoded by a one-dimensional convolutional neural network, ultimately obtaining a drug feature representation that integrates global semantics and topological structure, providing a high-quality feature representation for subsequent interaction modeling.

[0068] Specifically, when employing a dual-channel modeling strategy, i.e., a multimodal extraction strategy, for the extraction of drug sequence features, the computer equipment can first use a pre-trained large model Molformer based on the Transformer architecture to encode the drug SMILES sequence, obtaining an initial sequence feature representation F. ds ∈R L×d Let represent the d-dimensional feature representation of a sequence of length L. To further extract local structural patterns, a two-layer one-dimensional convolutional neural network can be used for processing: in, and These are the parameters of the l-th convolutional kernel.

[0069] In one embodiment, a drug-target affinity prediction method based on gene ontology guidance and multimodal attention may further include a process of extracting drug graph structure features. The specific process includes: using RDKit to convert the SMILES sequence of the drug molecule into a graph structure; using TransConv in a Transformer-based graph convolutional network to encode the graph structure; using an attention mechanism to aggregate neighbor information based on the encoded graph structure, introducing edge features, combining nonlinear activation and batch normalization, and forming an input tensor from all node features in the encoded graph structure through zero padding to obtain the topological structure features in the graph structure.

[0070] Computer devices can use Rdikt to convert drug sequences into drug graph structures and use TransConv for feature encoding. in, The input feature of node i in the (l-1)th layer is represented; N(i) represents the set of neighboring nodes of node i; e ij Let represent the edge features between node i and its neighbor j. Neighbor information is aggregated through an attention mechanism, and edge features are introduced, combined with non-linear activation and batch normalization to enhance the model's expressive power and training stability. All node features are zero-padding to form a unified input tensor. After l-layer encoding, a graph-structured drug feature representation is obtained. Used to support subsequent prediction tasks.

[0071] Step 208: Input the protein features and drug features into the two-branch multi-head cross-attention mechanism respectively. The two-branch multi-head cross-attention mechanism realizes the bidirectional interaction between protein features and drug features. The Mamba layer is used for deep feature extraction. Multimodal fusion features are obtained by splicing, pooling and weighted fusion.

[0072] To fully capture the complex interactions between drug and target, this embodiment designs an innovative multimodal interaction module. Specifically, it combines a cross-attention mechanism with a feature fusion strategy based on a Mamba layer of a structured state-space model (SSM). This not only captures the interdependencies between protein and drug features but also enhances the model's ability to model long sequences, thereby improving prediction performance. The multimodal interaction module includes a two-branch, multi-head cross-attention mechanism to handle the interactions between protein features and drug semantic features, and between protein features and drug structural features, respectively. Simultaneously, the Mamba module is introduced to enhance long sequence modeling capabilities, effectively capturing long-distance dependencies through the state-space model. Finally, protein and drug features are weighted and fused, and affinity prediction is performed using a multilayer perceptron. The entire model is trained end-to-end, using mean squared error as the loss function for optimization.

[0073] That is, the computer device can input the obtained drug and protein multimodal feature representations into the cross-attention module to realize bidirectional interaction between protein and drug (sequence and graph) features, then strengthen long-range dependencies through the Mamba layer, and finally integrate multimodal information through splicing, pooling and weighted fusion for DTA prediction.

[0074] In one embodiment, a drug-target affinity prediction method based on gene ontology guidance and multimodal attention may further include a feature interaction process, specifically including: performing a linear transformation on protein features and drug features to obtain corresponding query, key, and value vectors; performing cross-attention calculation on protein features and drug sequence features based on query, key, and value vectors to obtain a first fusion result; performing cross-attention calculation on protein features and topological structure features based on query, key, and value vectors to obtain a second fusion result; and inputting the first fusion result and the second fusion result into a Mamba layer based on a structured state-space model for deep feature extraction.

[0075] In the cross-attention mechanism, the integrated protein features are first analyzed. Drug sequence characteristics And topological features, i.e., drug map features A linear transformation is performed to obtain the corresponding query (Q), key (K), and value (V) vectors. The formula for calculating the cross-attention between protein and drug sequence features is: Among them, Q p K p V p Q represents the query, key, and value vectors of protein features, respectively. ds K ds V ds This represents a query, key, and value vector representing drug sequence features. Similarly, the cross-attention between protein features and drug graph features is defined as: Among them, Q dg K dg V dg This represents the query, key, and value vector obtained from the drug graph features, where p s and p g This represents the updated protein features after incorporating drug sequence information and graph information, d s and d g This represents the updated drug sequence and drug map features after incorporating protein information, a s ,a' s a g ,a' g These are the corresponding attention weights. Then, the output p of the cross-attention module... s pg d s d g It will be input into the Mamba layer based on the structured state-space model for deep feature extraction.

[0076] In one embodiment, the structured state space model includes a structured state space model module, a fully connected layer, a convolutional layer, and a pooling layer. The provided drug-target affinity prediction method based on gene ontology guidance and multimodal attention may further include a deep feature extraction process. Specifically, this process includes: defining transformation functions and determining activation functions based on the structured state space model module, the fully connected layer, the convolutional layer, and the pooling layer; and using the transformation functions and activation functions to extract deep features from the data input to the structured state space model, obtaining deep protein features and deep drug features.

[0077] In this embodiment, the transformation function is defined as: Φ(x)=fc(SSM(σ(conv(fc(p(x)))))⊙σ(fc(p(x)))); where SSM(·) represents the structured state space model module, used to efficiently capture long-range dependencies; fc(·), conv(·), and p(·) represent fully connected layer, one-dimensional convolution, and max pooling operations, respectively; σ(·) is the SiLU activation function. Using the defined transformation function, four deep feature representations can be obtained: Here, the superscripts s and g correspond to features from sequence and graph modalities, respectively. Deep feature representation of drugs that incorporate protein information, while This represents the deep feature representation of proteins that incorporate drug sequence or graph information. These rich feature representations provide more discriminative and semantically rich features, laying a solid foundation for subsequent multi-view fusion and interactive modeling, significantly improving the representational power of long sequence features and promoting more informative and complementary fusion.

[0078] In one embodiment, a drug-target affinity prediction method based on gene ontology guidance and multimodal attention may further include a feature fusion process. Specifically, this process includes: concatenating deep protein features and deep drug features into modal features; using max pooling to obtain the fused representations of protein and drug sequence features and the fused representations of protein and drug graph features; calculating the mean of the fused representations of protein and drug sequence features and the fused representations of protein and drug graph features, respectively, and using the ratio of the means as an adjustment coefficient; using the adjustment coefficient to weight the deep protein features and deep drug features, and introducing a fusion layer for nonlinear interaction to obtain multimodal fused features.

[0079] In the feature fusion stage, the two modal features of protein and drug are concatenated, and max pooling is used to obtain two interactive representations: Among them, f s This represents the fused protein and drug sequence features; Concat(·) represents the splicing operation; f g This represents the fused protein and drug graph features. To adaptively balance the contributions of these two modalities before final fusion, a fusion module FSR is introduced in this embodiment, specifically defined as the ratio of the absolute means of sequence features to graph features: Where mean(·) represents the operation of calculating the mean of the tensor elements, and the scalar is used as an adjustment coefficient to weight the graph feature representation: f' = mean(Concat(f s ,FSR⊙f g Next, a fusion layer is introduced to perform nonlinear interaction between these two types of complementary information, defined as: f = ω(Concat(f g ,f s ,σ(f')⊙f g ,σ(f')⊙f s ); where σ(·) represents the Sigmoid activation function, ω(·) is a linear transformation implemented through a fully connected layer, and the final fused feature f is input into the classifier to predict drug-target interactions.

[0080] Step 210: The regression module based on the multilayer perceptron performs regression prediction on the multimodal fusion features to obtain the binding affinity between the drug and the target protein.

[0081] In one embodiment, the computer device may be trained based on a regression module of a multilayer perceptron, including multilayer linear transformations and nonlinear activations, by minimizing the mean squared error loss.

[0082] The computer device uses a regression module based on a multilayer perceptron (MLP) to perform regression prediction on the fused feature f, outputting the binding affinity between the drug and the target. This mainly involves multilayer linear transformation and nonlinear activation, and its structure can be represented as follows: Where W1, W2, and W3 are weight matrices, b1, b2, and b3 are bias terms, φ(·) represents a nonlinear activation function (such as ReLU), and y is the predicted binding affinity. Training is performed by minimizing the mean squared error (MSE) loss.

[0083] In this embodiment, six widely accepted evaluation metrics were selected: RMSE, MAE, Pearson correlation coefficient, Spearman correlation coefficient, consistency index (CI), and r. m 2These metrics can measure the predictive performance from different perspectives, such as accuracy, relevance, and ranking ability.

[0084] In one embodiment, a drug-target affinity prediction method based on gene ontology guidance and multimodal attention can be implemented using a GoMA-DTA model, such as... Figure 3 As shown, the GoMA-DTA model framework can include: A) a functional annotation completion module; B) a feature extraction and fusion module; C) a cross-attention module; D) a Mamba module; and E) a multimodal feature fusion module. Specifically, the functional annotation completion module can use the NetGO tool to complete the missing GO functional annotations for target proteins in the gene ontology, retrieving GO functional annotations for each target protein from the UniProt database for subsequent data processing. The feature extraction and fusion module can use a pre-trained ESM-2 model to encode the input amino acid sequence, obtaining a sequence representation rich in biological information; simultaneously, the GO functional annotations supplemented by the UniProt database are encoded using a specially trained BlueBERT model to generate semantically meaningful feature vectors; finally, a gated fusion mechanism is designed to adaptively fuse sequence features and functional annotation features to form the final protein feature representation. It is also used to process SMILES sequences using the Molformer model to extract semantic features of molecules; the molecules are transformed into graph structures, and topological features are extracted using a Transformer-based graph convolutional network (TransConv). The features of the two channels are respectively deep encoded by a one-dimensional convolutional neural network to obtain drug features. The cross-attention module can input protein features and drug features into a two-branch multi-head cross-attention mechanism, which realizes bidirectional interaction between protein features and drug features. The Mamba module is used for deep feature extraction, and the multimodal feature fusion module concatenates, pools, and weights the features to obtain multimodal fused features, thereby realizing the prediction of binding affinity between drugs and target proteins.

[0085] In this application, a large-scale pre-trained model, ESM2, is used to extract protein sequence features. Simultaneously, functional vectors based on GeneOntology (GO) are combined to explicitly integrate protein functional annotation information into the model, thereby enhancing the biological guidance of protein feature representation. For drug feature modeling, a large-scale pre-trained model, Molformer, is used to extract features from SMILES sequences. The TransConv module is then used to extract structural information from the drug molecule map, achieving a deep fusion of drug semantics and chemical structure. A multi-head cross-attention mechanism and an efficient Mamba module are designed to capture complex drug-target multimodal interaction features. Finally, affinity prediction is achieved through feature fusion and a multilayer perceptron. This innovative architecture overcomes the limitations of traditional methods in feature extraction and prediction performance, achieving efficient and accurate prediction of drug-target affinity.

[0086] This application achieves comprehensive multidimensional expression of drug and protein features by fusing drug molecular graphs, SMILES sequence features, and protein sequence and GO functional annotation features. It effectively captures the complex interactions between the two by utilizing cross-attention and the Mamba structure. Finally, it accurately predicts drug-protein binding affinity through a regression network. Specifically, this application is the first to introduce protein Gene Ontology (GO) functional annotation features into the drug-target affinity prediction task, combining them with protein sequence embeddings generated by the large-scale pre-trained model ESM2 to achieve explicit biological guidance for key functional regions of proteins. For drugs, it uses a Molformer architecture based on the Transformer architecture to encode SMILES sequences and combines this with the TransConv module to extract molecular graph structural features, achieving multimodal fusion of drug semantics and structural information. Through multi-head cross-attention mechanisms and the Mamba module, it deeply mines the complex and multi-level interaction features between drugs and proteins. This overcomes the limitations of traditional methods in feature extraction, achieving efficient and accurate prediction of drug-target affinity, demonstrating the ability to learn complex relationship patterns from large-scale data, and improving the accuracy and efficiency of affinity prediction.

[0087] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0088] In one embodiment, such as Figure 4 As shown, a drug-target affinity prediction system based on gene ontology guidance and multimodal attention is provided, including: a data collection module 410, a protein feature extraction module 420, a drug feature extraction module 430, a multimodal feature fusion module 440, and a regression prediction module 450, wherein:

[0089] The data collection module 410 is used to collect affinity datasets containing drug molecule structures and target protein sequences, and to complete the target proteins in the dataset that lack gene ontology GO functional annotations.

[0090] The protein feature extraction module 420 is used to extract the GO functional annotations of target proteins in the dataset, and after encoding the GO functional annotations, extract the GO feature representations through CNN; use the ESM-2 model to extract sequence features from the target protein sequence; and fuse the GO feature representations with the sequence features through a gating mechanism to obtain the protein features.

[0091] The drug feature extraction module 430 is used to encode the SMILES sequence of drug molecules using a Molformer model based on the drug molecule structure in the dataset, and then extract the drug sequence features through a CNN; the SMILES sequence of drug molecules is converted into a graph structure, and the topological structure features in the graph structure are extracted using a Transformer-based graph convolutional network; the drug sequence features and the topological structure features are fused to obtain the drug features.

[0092] The multimodal feature fusion module 440 is used to input protein features and drug features into the two-branch multi-head cross attention mechanism, respectively. The two-branch multi-head cross attention mechanism realizes the bidirectional interaction between protein features and drug features, and uses the Mamba layer for deep feature extraction. Multimodal fused features are obtained through splicing, pooling and weighted fusion.

[0093] The regression prediction module 450 is used to perform regression prediction on multimodal fusion features based on the multilayer perceptron regression module to obtain the binding affinity between the drug and the target protein.

[0094] In one embodiment, the data collection module 410 is further configured to locate target proteins in the dataset that lack gene ontology GO functional annotations; retrieve gene ontology GO functional annotations for each target protein from the UniProt database; and complete the missing gene ontology GO functional annotations for the target proteins using the NetGO tool.

[0095] In one embodiment, the protein feature extraction module 420 is further configured to encode GO function annotations using a pre-trained model BlueBERT to obtain GO feature representations; input the GO feature representations into a CNN model, capture the local and global dependencies between GO function annotations through the CNN model, and obtain GO feature representations after convolution processing.

[0096] In one embodiment, the protein feature extraction module 420 is further configured to extract sequence features from the target protein sequence using the ESM-2 model and perform convolution processing to obtain the final sequence features; perform average pooling on the GO feature representation along the annotation dimension to obtain a compact vector, and calculate a gate vector based on the compact vector through linear transformation and Sigmoid activation; and apply the gate vector to the final sequence features for fusion to obtain the protein features.

[0097] In one embodiment, the drug feature extraction module 430 is further configured to convert the SMILES sequence of the drug molecule into a graph structure using RDKit, and to perform feature encoding processing on the graph structure using TransConv in the Transformer-based graph convolutional network; based on the encoded graph structure, attention mechanism is used to aggregate neighbor information, introduce edge features, combine nonlinear activation and batch normalization, and all node features in the encoded graph structure are zero-padding to form an input tensor to obtain the topological features in the graph structure.

[0098] In one embodiment, the multimodal feature fusion module 440 is further configured to perform linear transformation on protein features and drug features to obtain corresponding query, key, and value vectors; perform cross-attention calculation on protein features and drug sequence features based on query, key, and value vectors to obtain a first fusion result; perform cross-attention calculation on protein features and topological structure features based on query, key, and value vectors to obtain a second fusion result; and input the first fusion result and the second fusion result into a Mamba layer based on a structured state space model for deep feature extraction.

[0099] In one embodiment, the structured state space model includes a structured state space model module, a fully connected layer, a convolutional layer, and a pooling layer; the multimodal feature fusion module 440 is further used to define a transformation function and determine an activation function based on the structured state space model module, the fully connected layer, the convolutional layer, and the pooling layer; using the transformation function and the activation function, deep feature extraction is performed on the data input to the structured state space model to obtain deep protein features and deep drug features.

[0100] In one embodiment, the multimodal feature fusion module 440 is further configured to perform modal feature concatenation on the deep protein features and deep drug features, and use max pooling to obtain the fused representation of protein and drug sequence features and the fused representation of protein and drug graph features; perform mean operation on the fused representation of protein and drug sequence features and the fused representation of protein and drug graph features respectively, and use the ratio of the means as the adjustment coefficient; use the adjustment coefficient to perform weighting operation on the deep protein features and deep drug features, and introduce a fusion layer for nonlinear interaction to obtain multimodal fused features.

[0101] In one embodiment, the regression prediction module 450 is further configured to include a multilayer perceptron-based regression module comprising multilayer linear transformations and nonlinear activations, and to be trained by minimizing mean squared error loss.

[0102] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a drug-target affinity prediction method based on gene ontology guidance and multimodal attention. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device casing, or an external keyboard, touchpad, or mouse.

[0103] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0104] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement steps of a drug-target affinity prediction method based on gene ontology guidance and multimodal attention.

[0105] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of a drug-target affinity prediction method based on gene ontology guidance and multimodal attention.

[0106] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0107] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0108] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A drug-target affinity prediction method based on gene ontology guidance and multimodal attention, characterized in that, The method includes: Collect affinity datasets containing drug molecular structures and target protein sequences, and perform completion operations on target proteins in the dataset that lack gene ontology GO functional annotations. GO functional annotations of target proteins in the dataset are extracted, and GO feature representations are obtained by encoding the GO functional annotations and then extracting them using CNN. Sequence features are extracted from the target protein sequences using the ESM-2 model. The GO feature representations are fused with the sequence features through a gating mechanism to obtain protein features. Based on the drug molecule structures in the dataset, the SMILES sequences of the drug molecules are encoded using a Molformer model and then extracted using a CNN to obtain drug sequence features; the SMILES sequences of the drug molecules are converted into a graph structure, and topological features in the graph structure are extracted using a Transformer-based graph convolutional network; the drug sequence features and the topological features are fused to obtain the drug features. The protein features and the drug features are respectively input into a two-branch multi-head cross-attention mechanism to achieve bidirectional interaction between the protein features and the drug features. A Mamba layer is used for deep feature extraction, and multimodal fusion features are obtained through splicing, pooling, and weighted fusion. The regression module based on the multilayer perceptron performs regression prediction on the multimodal fusion features to obtain the binding affinity between the drug and the target protein.

2. The drug-target affinity prediction method based on gene ontology guidance and multimodal attention according to claim 1, characterized in that, The dataset is used to complete the target proteins for missing gene ontology GO functional annotations, including: Locate the target proteins in the dataset that lack GO functional annotations of the gene ontology; Retrieve gene ontology GO functional annotations for each target protein from the UniProt database; Use the NetGO tool to complete the GO functional annotations of genes with missing target proteins.

3. The drug-target affinity prediction method based on gene ontology guidance and multimodal attention according to claim 1, characterized in that, After encoding the GO function annotations, GO feature representations are extracted using a CNN, including: The GO feature annotations are encoded using the pre-trained model BlueBERT to obtain GO feature representations. The GO feature representation is input into a CNN model, which captures the local and global dependencies between GO function annotations. After convolution processing, the GO feature representation is obtained.

4. The drug-target affinity prediction method based on gene ontology guidance and multimodal attention according to claim 1, characterized in that, Sequence features were extracted from the target protein sequence using the ESM-2 model; The GO feature representation is fused with the sequence feature through a gating mechanism to obtain protein features, including: The ESM-2 model was used to extract sequence features from the target protein sequence and then convolutional processing was performed to obtain the final sequence features. The GO feature representation is average pooled along the annotation dimension to obtain a compact vector, and a gate control vector is calculated based on the compact vector through linear transformation and Sigmoid activation; The gating vector is applied to the final sequence features for fusion to obtain protein features.

5. The drug-target affinity prediction method based on gene ontology guidance and multimodal attention according to claim 1, characterized in that, The SMILES sequence of the drug molecule is converted into a graph structure, and the topological features of the graph structure are extracted using a Transformer-based graph convolutional network, including: The SMILES sequence of the drug molecule was converted into a graph structure using RDKit, and the graph structure was then processed by feature encoding using TransConv in a Transformer-based graph convolutional network. Based on the encoded graph structure, attention mechanism is used to aggregate neighbor information, introduce edge features, combine nonlinear activation and batch normalization, and form the input tensor of all node features in the encoded graph structure through zero padding to obtain the topological features in the graph structure.

6. The drug-target affinity prediction method based on gene ontology guidance and multimodal attention according to claim 1, characterized in that, The bidirectional interaction between protein features and drug features is achieved through the aforementioned dual-branch multi-head cross-attention mechanism, and deep feature extraction is performed using Mamba layers, including: A linear transformation is performed on the protein features and the drug features to obtain the corresponding query, key, and value vectors. Based on the query, key, and value vectors, cross-attention calculation is performed on the protein features and the drug sequence features to obtain a first fusion result; based on the query, key, and value vectors, cross-attention calculation is performed on the protein features and the topological structure features to obtain a second fusion result; The first fusion result and the second fusion result are input into the Mamba layer based on the structured state-space model for deep feature extraction.

7. The drug-target affinity prediction method based on gene ontology guidance and multimodal attention according to claim 6, characterized in that, The method further includes: The structured state space model includes a structured state space model module, a fully connected layer, a convolutional layer, and a pooling layer. Transformation functions are defined and activation functions are determined based on the structured state space model module, fully connected layer, convolutional layer, and pooling layer. Using the transformation function and activation function, deep feature extraction is performed on the data input to the structured state space model to obtain deep protein features and deep drug features.

8. The drug-target affinity prediction method based on gene ontology guidance and multimodal attention according to claim 7, characterized in that, Multimodal fusion features are obtained through splicing, pooling, and weighted fusion, including: The deep protein features and deep drug features are spliced ​​together using modal features, and max pooling is used to obtain the fused representation of protein and drug sequence features and the fused representation of protein and drug graph features. The average values ​​of the protein-drug sequence feature fusion representation and the protein-drug map feature fusion representation are calculated respectively, and the ratio of the average values ​​is used as an adjustment coefficient. The aforementioned adjustment coefficients are used to weight the deep features of proteins and drugs, and a fusion layer is introduced for nonlinear interaction to obtain multimodal fusion features.

9. The drug-target affinity prediction method based on gene ontology guidance and multimodal attention according to claim 1, characterized in that, The regression module based on a multilayer perceptron includes multilayer linear transformation and nonlinear activation, and is trained by minimizing the mean squared error loss.

10. A drug-target affinity prediction system based on gene ontology guidance and multimodal attention, characterized in that, The system includes: The data collection module is used to collect affinity datasets containing drug molecule structures and target protein sequences, and to complete the target proteins in the dataset that lack gene ontology GO functional annotations. The protein feature extraction module is used to extract the GO functional annotations of target proteins in the dataset, encode the GO functional annotations, and extract GO feature representations through CNN; extract sequence features from the target protein sequence using the ESM-2 model; and fuse the GO feature representations with the sequence features through a gating mechanism to obtain protein features. The drug feature extraction module is used to encode the SMILES sequences of drug molecules using a Molformer model based on the drug molecule structure in the dataset, and then extract drug sequence features through a CNN; convert the SMILES sequences of drug molecules into a graph structure, and extract the topological structure features in the graph structure using a Transformer-based graph convolutional network; and fuse the drug sequence features with the topological structure features to obtain the drug features. The multimodal feature fusion module is used to input the protein features and the drug features into the two-branch multi-head cross-attention mechanism, respectively. The two-branch multi-head cross-attention mechanism realizes the bidirectional interaction between the protein features and the drug features, and uses the Mamba layer to perform deep feature extraction. The multimodal fused features are obtained by splicing, pooling and weighted fusion. The regression prediction module is used to perform regression prediction on the multimodal fusion features based on the multilayer perceptron regression module to obtain the binding affinity between the drug and the target protein.

Images