Cross-modal interactive fusion structure diagram feature extraction method

Abstract

The invention discloses a cross-modal interactive fusion structure diagram feature extraction method, and the method employs a multilayer diagram isomorphic network to extract local neighborhood features and global topological features of drug molecules, thereby improving the capability of a model for capturing the change of a fine structure. A convolutional neural network and a multi-head convolutional self-attention mechanism are integrated to comprehensively capture local features and long-range dependency relationships in a protein sequence; and interaction information between a drug and a target is fully modeled through a cross-modal interaction fusion module based on a gating mechanism, so that more accurate DTI prediction is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computational biology, in particular the technical field of drug-target interaction prediction. BACKGROUND

[0002] Drug-target interaction (DTI) prediction is a key step in drug development; accurate DTI prediction relies on in-depth analysis of the potential features of drugs and targets, and accurate modeling of their complex interactions; however, existing methods are difficult to effectively capture subtle structural differences between drug molecules, and often cannot fully extract long-range dependencies in protein sequences; in addition, the features of drugs and proteins are usually combined through simple splicing, ignoring the cross-modal complementarity between them.

[0003] In the DTI prediction task, sequence models have become one of the research focuses in recent years due to their ability to capture contextual information and long-range dependencies in sequence data. EmbedDTI combines sequence embedding with graph convolutional networks to enhance the representation of molecular structures and build a model to predict the interaction between drugs and targets (JIN Y, LU J, SHI R, et al. EmbedDTI: Enhancing the Molecular Representations via Sequence Embedding and Graph Convolutional Network for the Prediction of Drug-Target Interaction [J / OL]. Biomolecules, 2021, 11(12): 1783. DOI:10.3390 / biom11121783.). DTI prediction models based on the Transformer architecture show that even with only sequence data, deep semantic features of drugs and proteins can be effectively extracted (LIU S, WANG Y, DENG Y, et al. Improved drug–target interaction prediction with intermolecular graph transformer [J / OL]. Briefings in Bioinformatics, 2022, 23(5): bbac162. DOI:10.1093 / bib / bbac162.). CoaDTI further integrates pre-trained encoders and co-attention mechanisms to address the label sparsity problem while extracting complementary features from the original protein sequences and drug SMILES representations (HUANG L, LIN J, LIU R, et al. CoaDTI: multi-modal co-attention based framework for drug–target interaction annotation [J / OL]. Briefings in Bioinformatics, 2022, 23(6): bbac446. DOI:10.1093 / bib / bbac446.). The DrugBAN model uses graph convolutional networks and one-dimensional convolutional networks to encode drug molecular graphs and protein sequences, respectively, and introduces a bilinear attention mechanism to explicitly model the local interaction between drugs and targets, effectively capturing the specific interactions between key molecular fragments and binding sites (BAI P, MILJKOVIĆ F, JOHN B, et al.Interpretable bilinear attention network with domain adaptation improves drug–target prediction [J / OL]. Nature Machine Intelligence, 2023, 5(2): 126-136. DOI:10.1038 / s42256-022-00605-1.); The GIFDTI model constructs an encoding module consisting of three parts: drug, target, and interaction. It uses CNN and Transformer together to effectively integrate local and global features and model the interaction relationships between different categories (ZHAO Q, DUAN G, ZHAO H, et al. GIFDTI: Prediction of Drug-Target Interactions Based on Global Molecular and Intermolecular Interaction Representation Learning [J / OL]. IEEE / ACM Transactions on Computational Biology and Bioinformatics, 2023, 20(3): 1943-1952. DOI:10.1109 / TCBB.2022.3225423.). .

[0004] Although deep learning-based DTI prediction methods have shown significant advantages in latent feature modeling and capturing complex relationships, existing methods still have the following problems in practical applications due to the structural complexity and modal heterogeneity of the task itself: (1) Some drug molecules have slight differences in chemical composition and topological structure, which usually lead to significant differences in their biological activities; current methods often have difficulty effectively capturing these slight structural differences when dealing with drug subgraphs with similar structures, which limits the accurate representation of drug molecule features; (2) The function and structure of proteins are jointly determined by local information and long-range dependencies. Long-range dependencies between distant residues can affect the binding stability and affinity between drugs and targets; existing models focus more on local information when extracting protein features, which limits the comprehensive capture of drug-target interaction features; (3) There may be specific binding modes between key structural domains and target proteins in drug compounds. Existing studies usually directly splice the feature representations of drugs and proteins, failing to fully consider the cross-modal complementarity between drugs and proteins, lacking fine modeling of complex intermolecular interactions in biology, and failing to effectively identify the part of drug molecules that contributes the most to the interaction with target proteins. SUMMARY

[0005] The purpose of the present application is to solve the problems in the prior art, and propose a structural graph feature extraction method based on cross-modal interaction fusion, which uses a multi-layer graph isomorphism network to extract the local neighborhood features and global topological features of drug molecules, thereby improving the ability of the model to capture subtle structural changes; integrates a convolutional neural network and a multi-head convolutional self-attention mechanism to comprehensively capture the local features and long-range dependencies in the protein sequence; and fully models the interaction information between the drug and the target through a cross-modal interaction fusion module based on a gating mechanism, thereby achieving more accurate DTI prediction.

[0006] To achieve the above purpose, the present application proposes a structural graph feature extraction method based on cross-modal interaction fusion, comprising the following steps: S1, input the SMILES string of the drug molecule into the drug feature encoding module to generate a structured representation suitable for graph neural network processing; and input the amino acid sequence of the protein into the protein feature encoding module to obtain the initial feature representation of the protein; S2, further process the drug features through a biomolecule graph feature extraction module, and use a multi-level graph isomorphism network to extract the local neighborhood information and global topological features of the drug molecule; the protein features are processed by a protein cross-residue feature extraction module, and the local features and long-range dependencies of the protein sequence are captured through a convolutional neural network and a multi-head convolutional self-attention mechanism; S3, input the feature representations of the drug and the protein into the cross-modal interaction fusion module, and adaptively model the deep-level interaction relationship between the drug and the target through a gating mechanism; S4, the fused features are processed through a pooling operation and a feedforward neural network to generate the final drug-target interaction prediction result.

[0007] As a preferred, the working steps of the drug feature encoding module are as follows: S1.1, first parse the SMILES string into a drug molecule graph using the open source toolkit RDKit Each atom node (corresponding to an atom in the drug molecule graph, the edge represents the connection relationship between two atoms; next, use DeepChem to generate a 78-dimensional binary feature vector for each atom node in the drug molecule graph ; S1.2, then introduce an adjacency matrix to represent the connection relationship between atoms in the drug molecule, wherein if two atoms are connected, the corresponding matrix element is 1; otherwise, it is 0; S1.3、Finally, we concatenate the feature vectors of each atom to form a unified feature representation Then, we normalize the features as the input of the final drug feature extraction module.

[0008] As preferred, in the protein feature encoding module, the length of all protein sequences is standardized to 1200; when the length of a protein sequence is less than 1200, it is completed by zero padding; when the length exceeds 1200, the first 1200 amino acid characters are truncated; if some letters are not in the defined letter set during encoding, these characters will be ignored and treated as padding characters; in order to enhance the ability of feature representation, integer encoding is input into the embedding layer to map each integer to a high-dimensional continuous vector, thereby generating a feature matrix of the protein sequence.

[0009] As preferred, the working steps of the biomolecule graph feature extraction module are as follows: S2A.1、The biomolecule graph feature extraction module processes the direct neighbor and indirect neighbor features in the molecular graph respectively by using a multi-channel structure; each channel contains multiple GIN layers to update the node features in a recursive manner; the first channel of the biomolecule graph feature extraction module takes the node features and the direct adjacency matrix as input, focusing on extracting local topological structure features; the second channel takes the node features and the indirect adjacency matrix as input, capturing multi-hop neighborhood information between nodes; the indirect adjacency matrix is constructed based on the direct adjacency matrix A and the multi-hop adjacency relationship is calculated by the following formula: (2); S2A.2、After completing the multi-channel feature extraction, the biomolecule graph feature extraction module adopts a feature concatenation fusion strategy to integrate the outputs of the two channels; the local topological structure features and multi-hop neighborhood features extracted by the two channels are concatenated to obtain a more comprehensive and comprehensive node feature representation.

[0010] As preferred, the protein cross-residue feature extraction module extracts local amino acid features by convolution and uses SE blocks to adaptively adjust channel weights to enhance the expression of important features, while introducing a multi-head convolution self-attention mechanism to comprehensively capture local and global features in the protein sequence.

[0011] As preferred, the specific working steps of the protein cross-residue feature extraction module are as follows: S2B.1、First, local amino acid features in the protein sequence are extracted by combining convolution operation and SE blocks: The input integer encoded protein sequence is processed by a continuous one-dimensional convolution and a LeakyReLU activation function to capture local dependencies in the protein sequence; the convolution operation can be represented as: (3); The SE block is introduced to adaptively adjust the weight of each channel feature, thereby enhancing the expression of important features and suppressing redundant information and low correlation features in the channel dimension to some extent; S2B.2, Finally, in order to capture the long-range dependencies in the protein sequence, we introduce a multi-head convolution self-attention mechanism to improve the model's ability to capture long-range information interactions in the protein sequence by combining convolution and self-attention mechanisms; Given the input feature , the calculation process of the multi-head convolution self-attention mechanism is as follows: (4); (5);

[0012] After the above process, the final output is obtained as the feature representation of the protein sequence.

[0013] As a preferred, the cross-modal interaction fusion module captures the bidirectional cross-modal feature interaction between the drug and the protein through a bidirectional cross-attention mechanism, and uses a gating mechanism to weight and fuse the features.

[0014] As a preferred, in the bidirectional cross-attention mechanism, in order to ensure the stability of the training, the features of the drug and the protein are first subjected to layer normalization, (7); Then the multi-head cross-attention is used to calculate the attention features of the drug to the protein and the attention features of the protein to the drug respectively; the multi-head attention mechanism allows the model to capture the interaction between the drug and the protein features from different subspaces in parallel. For the attention calculation of the two interaction directions, the following multi-head attention formula is used: (8); When the drug feature is used as the query vector and the protein feature is used as the key and value, the cross-attention feature is represented as: (9);

[0015] Correspondingly, when the protein feature is used as the query vector and the drug feature is used as the key and value, the cross-attention feature is represented as: (10).

[0016] As preferred, when the gating mechanism is used for weighted fusion, firstly, the output features are calculated by a fully connected layer and a sigmoid function to obtain the fusion weights of drug features and protein features, and then the two fusion weights are used to weight the drug and protein features element by element: (11).

[0017] Subsequently, the features after gating fusion are reduced in dimension by average pooling; the final pooled feature representation can be directly input to the subsequent feedforward network for prediction of drug-target interactions.

[0018] The application also proposes a drug-target interaction prediction model, which executes the cross-modal interaction fusion structure graph feature extraction method, including a biomolecule graph feature extraction module, a protein cross-residue feature extraction module and a cross-modal interaction fusion module, wherein the biomolecule graph feature extraction module extracts local neighborhood information and global topological features of drug molecules in parallel using a multi-level graph isomorphism network; the protein cross-residue feature extraction module comprehensively captures local features and long-range dependencies of protein sequences through a convolutional neural network and a multi-head convolutional self-attention mechanism; and the cross-modal interaction fusion module realizes deep-level interaction fusion of drug and protein features through a gating mechanism.

[0019] The application has the following beneficial effects: firstly, the application designs a biomolecule graph feature extraction module capable of distinguishing subtle structural differences in view of the limitations of existing GNN models in capturing subtle chemical and topological structural differences; the module uses a multi-level graph isomorphism network to process local neighborhood information and global topological relationships in the drug molecule graph, thereby effectively enhancing the model's ability to capture subtle structural differences in drugs; secondly, in order to overcome the shortcomings of existing models in extracting protein long-range dependencies, we design a protein cross-residue feature extraction module that enhances the ability to extract protein long-range dependencies; this module not only retains the advantages of convolutional neural networks in extracting local features, but also accurately captures long-range dependencies across residues through the introduction of a multi-head convolutional self-attention mechanism, thereby improving the comprehensiveness of protein sequence feature extraction; finally, we propose a cross-modal interaction fusion module that can enhance the modeling ability of deep-level interaction between drugs and targets, aiming to solve the problem that traditional splicing methods cannot effectively capture deep-level interaction between drugs and targets; this module can more comprehensively consider the cross-modal complementarity between drugs and targets, avoiding information redundancy and loss of key features caused by simple splicing.

[0020] The features and advantages of the present invention will be described in detail through embodiments and in conjunction with the accompanying drawings. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the drug-target interaction prediction model of the present invention; Figure 2 This is a figure showing the ablation experiment results of this invention on the BindingDB, BioSNAP, and DAVIS datasets; Figure 3 This is a graph showing the performance of the present invention on the BindingDB, BioSNAP, and DAVIS datasets under different learning rate thresholds. Figure 4 This is a performance graph showing the results of this invention using different numbers of attention heads on the BindingDB, BioSNAP, and DAVIS datasets; Figure 5 This is a performance comparison of the present invention under random partitioning and cold pairing partitioning conditions on the Human dataset; Figure 6 and Figure 7 This is a diagram showing the docking results of a molecular docking experiment. Detailed Implementation

[0022] Problem Definition In DTI prediction tasks, the goal is to predict whether an interaction exists between a given drug and a target protein; the target protein is typically represented as an amino acid sequence. ,in Representing one of the 25 standard amino acids; to facilitate model processing, an embedding layer is typically used to map it to a latent feature space; drugs are encoded using Simplified Molecular Input Line Representation (SMILES), which records the atomic and bond information of the molecule in the form of a one-dimensional string; to enable the model to more effectively capture the structural information of the drug, the SMILES string is usually converted into a corresponding two-dimensional molecular diagram. In this model, V represents atoms (nodes) in the drug molecule, and E represents chemical bonds (edges). By constructing an adjacency matrix to capture the structural information of the drug molecule, the model can more effectively model the chemical properties of the drug. Based on the above representation method, the core of DTI prediction is to learn a mapping function M, which maps the joint feature space P×G of the drug molecule graph G and the target protein sequence P to the interaction probability score. This is to measure the likelihood of interaction between the drug and the target protein.

[0023] Model Overview like Figure 1As shown, the drug-target interaction prediction model consists of three main parts: BioGraphNet, Protein Cross-Residue Feature Extraction (PCRFE), and Cross-Modal Interaction Fusion (CMIF).

[0024] BioGraphNet utilizes a multi-level graph isomorphic network (GIN) to extract local neighborhood information and global topological features of drug molecules in parallel; PCRFE comprehensively captures local features and long-range dependencies of protein sequences through a convolutional neural network (CNN) and a multi-head convolutional self-attention mechanism; CMIF achieves deep interaction and fusion of drug and protein features through a gating mechanism; finally, the fused features are processed by pooling and feedforward neural networks to output DTI prediction results.

[0025] First, the SMILES string of the drug molecule is input into the drug feature encoding module to generate a structured representation suitable for graph neural network processing. The amino acid sequence of the protein is input into the protein feature encoding module to obtain the initial feature representation of the protein. Next, the drug features are further processed by BioGraphNet, which uses a multi-level graph isomorphic network (GIN) to extract local neighborhood information and global topological features of the drug molecule. The protein features are processed by protein cross-residue feature extraction (PCRFE), which uses a convolutional neural network (CNN) and a multi-head convolutional self-attention mechanism to capture local features and long-range dependencies of the protein sequence. Subsequently, the feature representations of the drug and protein are input into cross-modal interaction fusion (CMIF), which adaptively models the deep interaction between the drug and the target through a gating mechanism. Finally, the fused features are processed by pooling operations and a feedforward neural network to generate the final drug-target interaction prediction results.

[0026] The feature encoding module consists of two independent modules: drug feature encoding and protein feature encoding. Drug feature encoding converts the drug's SMILES string into a standardized numerical feature representation, while protein feature encoding converts the protein's amino acid sequence into a corresponding feature vector.

[0027] Drug Feature Encoding Module Drug feature encoding is a method that converts the SMILES strings of a drug into a structured input representation suitable for GNN processing. By systematically extracting and encoding the chemical information of the molecule, we ensure that the generated feature vectors and graph structures can accurately reflect the topological structure and chemical properties of the molecule. We use a molecular graph modeling method to extract atomic properties and their connections from the drug molecule graph. Drug molecules are encoded using chemical language representations (SMILES).

[0028] Specifically, we first use the open-source toolkit RDKit to parse the SMILES string into a drug molecule diagram. , in It is a collection of atomic nodes. It is a collection of chemical bonds; each atomic node (in () corresponds to an atom in the drug molecule diagram, edge This represents the connection between two atoms; next, the DeepChem tool is used to represent each atom node in the drug molecule diagram. Generate a 78-dimensional binary feature vector. This feature vector encodes a variety of chemical properties of an atom, including but not limited to the number of implicit hydrogen atoms, formal charge, total number of hydrogen atoms, atom hybridization state, number of unpaired electrons, atom type, atom mass, and whether the atom belongs to an aromatic structure; these binary features ensure that the chemical properties of each atom are fully represented.

[0029] To represent the connectivity between atoms in a drug molecule, we introduce an adjacency matrix. In this matrix, if two atoms are connected, the corresponding matrix element is 1; otherwise, it is 0. In this way, we can effectively describe the connections between atoms in a drug molecule. Finally, we define the feature vector of each atom... When pieced together, they form a unified feature representation. ,in This is the dimension of the feature vector; subsequently, we normalize the features to serve as input for the final drug feature extraction module.

[0030] Protein feature coding module Proteins are macromolecules composed of multiple amino acids linked by peptide bonds, possessing complex functions and structures. Typically, protein sequences are represented by letter symbols, with each letter representing a specific amino acid (e.g., "A" for alanine, "D" for aspartic acid, and "F" for phenylalanine). In protein feature coding, we define a set of letters containing common amino acids and assign a unique integer value to each amino acid. Using this set of letters, a protein sequence can be converted into a corresponding integer-coded sequence.

[0031] To ensure the consistency of input data, the length of all protein sequences is normalized to 1200. When the protein sequence length is less than 1200, it is padded with zeros; when the length exceeds 1200, the first 1200 amino acid characters are truncated. During the encoding process, if some letters are not in the defined letter set, these characters are ignored and treated as padding characters. To enhance the feature representation capability, integer codes are input into the embedding layer, mapping each integer to a high-dimensional continuous vector, thereby generating the feature matrix of the protein sequence; this matrix is ​​represented as... :in, The feature matrix representing proteins, This represents the dimension of each amino acid embedding vector.

[0032] Biomolecular graph feature extraction module Graph Isomorphism Networks (GINs), proposed by Xu et al. in 2018, aim to improve the ability of traditional graph neural networks to distinguish different graph structures (XU K, HU W, LESKOVEC J, et al. How Powerfulare Graph Neural Networks?[A / OL]. arXiv, 2019[2025-03-22]. http: / / arxiv.org / abs / 1810.00826. DOI:10.48550 / arXiv.1810.00826.). GINs show significant advantages in processing non-Euclidean spatial data such as molecular graphs, especially in distinguishing structurally similar drug subgraphs. Their theoretical basis stems from the classic Weisfeiler-Lehman (WL) graph isomorphism test algorithm, which is commonly used for graph structure discrimination problems (SHERVASHIDZENINO, SCHWEITZERPASCAL, JAN van L, et al. Weisfeiler-Lehman). GraphKernels [J / OL]. The Journal of Machine Learning Research, 2011 [2025-03-22]. https: / / dl.acm.org / doi / 10.5555 / 1953048.2078187. DOI:10.5555 / 1953048.2078187.); The WL algorithm recursively updates node labels, combining node features with their neighbor features to gradually enhance the ability to distinguish different graph structures; GIN improves upon this, especially in feature aggregation. By integrating neighbor information, GIN adaptively extracts key features from the graph, thereby improving the ability to represent complex graph structures; its core innovation lies in using a multilayer perceptron (MLP) to flexibly update features and through learnable parameters. The aggregation method of node features and neighbor features is adaptively adjusted to enhance the representational power of nodes; the feature update formula is as follows: (1); in, This represents the feature of node v at layer l. Let v be the set of neighbors of node v. These are learnable parameters used to balance the node's own features. Characteristics of neighbors The weights; although GIN performs well in molecular graph modeling, it still has certain limitations in distinguishing structurally similar or identical subgraphs and in modeling cross-level topological relationships.

[0033] To address this, we designed a biomolecular graph feature extraction module (BioGraphNet) to more comprehensively capture the feature information of drug molecule graphs and overcome the limitations of traditional graph neural networks in distinguishing structurally similar subgraphs and modeling multi-level dependencies. BioGraphNet improves its ability to distinguish structurally similar drug subgraphs by using a multi-channel structure to process the features of direct and indirect neighbors in the molecular graph separately. Each channel contains multiple GIN layers that recursively update node features. The first channel of BioGraphNet takes node features and the direct adjacency matrix as input and focuses on extracting local topological features. The second channel takes node features and the indirect adjacency matrix as input and captures multi-hop neighborhood information between nodes. The indirect adjacency matrix is ​​constructed based on the direct adjacency matrix A, and the multi-hop adjacency relationship is calculated using the following formula: (2);

[0034] Where A is the direct adjacency matrix, and K is the maximum number of hops. This represents the k-th matrix exponentiation operation of A; in the result, all elements greater than 0 are set to 1, forming a binary concatenated representation. This enhances the expression of remote dependencies.

[0035] After completing multi-channel feature extraction, BioGraphNet employs a feature concatenation fusion strategy to integrate the outputs of the two channels. We concatenate the local topological features extracted from the two channels with multi-hop neighborhood features to obtain richer and more comprehensive node feature representations. This fusion method can effectively preserve structural information from different topological ranges without introducing additional model complexity. Furthermore, the feature concatenation strategy facilitates subsequent model learning and optimization, enabling BioGraphNet to capture both fine-grained local structural patterns and reflect global topological relationships between nodes when dealing with complex drug molecule topologies, thereby enhancing the discriminative power of the final feature representation to a certain extent.

[0036] Protein cross-residue feature extraction module Protein sequences contain complex dependencies, including local structural features (such as secondary structures like α-helices and β-sheets) and long-range spatial interactions (such as hydrophobic interactions or hydrogen bonds in tertiary structures). While traditional convolutional neural networks (CNNs) can effectively capture local patterns, they are limited by fixed receptive fields and struggle to model long-range dependencies. To overcome this limitation, we designed a protein cross-residue feature extraction module (PCRFE). This module extracts local amino acid features through convolution and uses SE blocks to adaptively adjust channel weights to enhance the expression of important features. Simultaneously, we introduce a multi-head convolutional self-attention mechanism, which can effectively model long-range dependencies, thereby comprehensively capturing both local and global features in protein sequences.

[0037] First, local amino acid features in protein sequences are extracted by combining convolution operations with SE blocks. The input integer-encoded protein sequence is processed by consecutive one-dimensional convolutions and the Leaky ReLU activation function to capture local dependencies in the protein sequence. The convolution operation can be represented as: (3);

[0038] Where S is the input integer encoding protein sequence, S' represents a one-dimensional convolution operation with a kernel size of k, and S' is the feature map obtained through the convolution operation.

[0039] To further enhance the representation of key features, we introduce SE blocks to adaptively adjust the weights of each channel feature, thereby enhancing the expression of important features while suppressing redundant information and low-relevance features in the channel dimension to some extent.

[0040] Since convolution operations cannot capture the relative positional relationships of sequence elements, this may lead to a lack of understanding of the relative positions of elements when dealing with tasks with order dependencies, such as protein sequences. We introduce positional encoding based on sine and cosine functions to add positional information to each sequence element, enabling the model to learn the relative order of elements in the sequence, thereby improving the model's ability to model long-range dependencies.

[0041] Finally, to capture long-range dependencies in protein sequences, we introduce a multi-head convolutional self-attention mechanism. By combining convolution and self-attention, we improve the model's ability to capture long-range information interactions in protein sequences; given input features... The computation process of the multi-head convolution self-attention mechanism is as follows: (4); (5);

[0042] in, , , Generate query, key, and value convolution kernel parameters for one-dimensional convolution. For the dimension of the key, It is a scaling factor. It is a linear projection matrix used to merge multi-head outputs.

[0043] The output of the multi-head convolutional self-attention layer is further processed by a feedforward network to enhance feature representation capabilities. Furthermore, each sub-layer employs residual connections and layer normalization to stabilize training and improve gradient flow. The final output is obtained after these processes. , as a characteristic representation of protein sequence.

[0044] Cross-modal interaction fusion module Drug-target interactions occur at drug-protein binding sites, which are typically located in the pocket regions of proteins. These interactions determine the biological effects of drugs; therefore, accurately identifying these binding sites is crucial for drug discovery. In many existing studies, a common practice is to directly splice the features of drugs and proteins and process them through fully connected layers. However, this simple splicing method fails to effectively model the complex interactions between drugs and targets, making it difficult for models to accurately identify key features that play a decisive role in binding.

[0045] The core idea of ​​attention mechanisms is to quickly filter out high-value information using limited attention resources to address information redundancy in large-scale data. Gating mechanisms selectively transmit and fuse the most relevant features by controlling information flow, thereby removing redundant information. Based on these two methods, we designed a cross-modal interaction fusion (CMIF). This method captures the bidirectional cross-modal feature interactions between drugs and proteins through a bidirectional cross-attention mechanism, and uses a gating mechanism to perform weighted fusion of these features, thereby more accurately capturing the complex interaction relationships between drugs and proteins.

[0046] Given drug feature matrix and protein feature matrix The inputs are fed into feedforward neural networks, which map the input features into a new feature space, thereby capturing their higher-order nonlinear features; drug features and protein characteristics They are obtained through the following processes respectively: (6).

[0047] Next, we use a bidirectional cross-attention mechanism to capture the cross-modal dependency between drugs and proteins. When using a single cross-attention mechanism, protein features are typically used as the query, and drug features as the key and value. The limitation of this approach is that it can only capture the dependency of protein features on drug features, ignoring the influence of the drug on protein features. The bidirectional cross-attention mechanism, however, calculates drug attention separately... Interaction with protein attention This allows for a more comprehensive modeling of the complex relationship between drugs and proteins.

[0048] In the bidirectional cross-attention mechanism, to ensure training stability, the features of the drug and protein are first subjected to layer normalization: (7).

[0049] Next, multi-head cross-attention was used to calculate the attentional characteristics from drug to protein separately. Attention characteristics of proteins to drugs Multi-head attention mechanisms allow models to capture the interactions between drug and protein features in parallel from different subspaces. For attention calculations in two interaction directions, the following multi-head attention formula is used: (8); The computation of each attention head involves performing a linear transformation on the inputs Q, K, and V to generate a weighted feature representation; when using drug features... Protein features as query vectors When used as keys and values, the cross-attention feature is represented as follows: (9); This feature reflects how drug features selectively focus on key regions of protein features within the drug feature space, thereby characterizing the interaction between drugs and proteins.

[0050] Accordingly, when characterized by protein features As a query vector, using drug features as... When the key and value are the same, the cross-attention feature is represented as: (10).

[0051] After obtaining the output features using the bidirectional cross-attention mechanism, although the interaction dependency between the drug and the protein has been captured, these two features may still contain redundant information. Therefore, we introduce a gated fusion mechanism that can dynamically adjust the fusion weights of the drug and protein features and selectively weight the most relevant information.

[0052] First, the output features are used to calculate fusion weights through a fully connected layer and a sigmoid function. and Then, drug and protein features were weighted element-wise using two fusion weights: (11);

[0053] This gating fusion method ensures that drug and protein features are weighted according to their importance, adaptively selecting the most meaningful features.

[0054] Subsequently, we used average pooling to reduce the dimensionality of the gated fusion features; the final pooled feature representation can be directly input into the subsequent feedforward network for the prediction of drug-target interactions.

[0055] Experimental Section benchmark dataset We evaluated the model performance using the public datasets BindingDB, BioSNAP, Davis, and Human.

[0056] BindingDB is a database that aggregates data on interactions between small molecule drugs and proteins. We used the reference (ZHAO Q, DUAN G, ZHAO H, et al. GIFDTI: Prediction of Drug-Target Interactions Based on Global Molecular and Intermolecular Interaction Representation Learning[J / OL]. IEEE / ACM Transactions on Computational Biology and Bioinformatics, 2023, 20(3): 1943-1952). The preprocessed version is from DOI:10.1109 / TCBB.2022.3225423.; the BioSNAP dataset is from the DrugBank database and contains 4510 drugs and 2181 proteins, covering experimentally confirmed positive interactions and randomly sampled negative interactions; the Davis dataset includes 72 drugs and 379 proteins, with a total of 8587 interaction pairs; the Human dataset includes 2726 drugs and 2001 proteins, with a total of 6728 interaction pairs, focusing on human targets; detailed dataset statistics are shown in Table 1.

[0057] Table 1. Statistical Information of the Dataset

[0058]

[0059] Evaluation indicators We used five widely used evaluation metrics: Area under the receiver operating characteristic curve (AUROC), Area under the precision-recall curve (AUPRC), accuracy, sensitivity, and specificity; these metrics comprehensively evaluated the performance of the classification model from multiple perspectives.

[0060] The specific definitions of these indicators are as follows: AUROC is used to measure the overall ability of a model to distinguish between positive and negative samples. It is defined as the area under the receiver operating characteristic (ROC) curve. The AUROC definition is as follows: (12); AUPRC evaluates a model's performance on imbalanced datasets by calculating the area under the precision-recall curve, particularly its effectiveness in identifying minority class samples; AUPRC is the integral of precision with respect to recall. (13); Accuracy measures the overall consistency between predicted results and true labels, and is defined as: (14); Sensitivity assesses the model's ability to correctly identify positive samples, and is defined as: (15); Specificity assesses a model's ability to correctly identify negative samples, and is defined as: (16).

[0061] Experimental setup The experiment was conducted on a server running Ubuntu 20.04.

[0062] The server configuration included a Xeon(R) Platinum 8362 processor and an NVIDIA GeForce RTX 3090 graphics card; the experimental software environment used Python 3.8.10 and PyTorch 1.11.0, and depended on the following key libraries: DGL 0.9.1, Numpy 1.24.3, Pandas 1.5.3, Scikit-learn 1.0.2, and RDKit2021.03.1; to ensure the reproducibility of the results, all experiments were conducted in this environment; the key hyperparameter configurations of the model are detailed in Table 2.

[0063] Table 2 Key Hyperparameter Settings

[0064] .

[0065] Baseline method To comprehensively evaluate the performance of our model, we selected five existing representative models as baseline methods to verify the effectiveness of our model: DeepConv-DTI: DeepConv-DTI uses a convolutional neural network (CNN) to extract local features of protein sequences, while employing extended connection fingerprinting (ECFP4) to encode drug molecule features and a fully connected network (FCN) to predict drug-target interactions.

[0066] GraphDTA: GraphDTA represents drugs as molecular graphs based on graph neural networks. It extracts features of drugs and proteins through GNN and CNN respectively, and finally realizes the prediction of drug affinity to target.

[0067] MolTrans: MolTrans extracts substructural features of drugs and proteins using Transformer and constructs these features into an interaction graph for further prediction of drug-target interactions.

[0068] DrugBAN: DrugBAN uses a bilinear attention network (BAN) to explicitly capture the local interactions between drugs and targets, and improves the predictive performance on new drug-target pairs by aligning data from different distributions through conditional domain adversarial learning.

[0069] BINDTI: BINDTI uses a graph convolutional network (GCN) to encode drug features and combines it with ACmix to extract protein features. It then performs feature fusion through a bidirectional intent network and finally uses a multilayer perceptron to predict drug-target interactions.

[0070] DeFuseDTI: DeFuseDTI employs a dual-branch encoder architecture to extract local aggregation features and drug atom and target residue features separately; and achieves drug and target feature extraction through a multi-view fusion attention module. The interactive learning of features generates a joint representation which is then used to decode the prediction results.

[0071] Performance comparison with different models To evaluate the performance of the BPCI-DTI model, we compared it with six mainstream methods. All experiments were based on the widely used standardized DTI dataset, which was randomly divided into training, validation, and test sets in a 7:1:2 ratio. To ensure the reproducibility of the experimental results, we conducted five independent experiments on each dataset under different random seeds to reduce the impact of experimental bias and randomness. Finally, based on the model parameters that performed best on the validation set, we performed a final performance evaluation on the test set. The experimental results are detailed in Table 3, which lists the performance of each model on different datasets.

[0072] Table 3 shows the performance comparison on the BindingB, BioSNAP, and DAVIS datasets.

[0073]

[0074] BPCI-DTI achieved high AUROC and AUPRC values ​​in most experiments, significantly outperforming other baseline methods. In particular, on the BioSNAP and Davis datasets, BPCI-DTI's AUROC and AUPRC values ​​were significantly higher than other baseline methods. However, on the BindingDB dataset, although BPCI-DTI improved in AUROC and AUPRC values, the performance improvement was relatively limited compared to the BioSNAP and Davis datasets. This may be related to the diversity and complexity of the BindingDB dataset, which contains multiple types of targets, resulting in greater data heterogeneity.

[0075] ablation experiment To evaluate the contribution of each module to the model performance, we conducted ablation experiments; we designed four variations of the ablation experiment to evaluate the independent contribution of each module in the DTI prediction task; firstly, The variant replaces BioGraphNet with standard GIN to simplify the drug feature extraction process and evaluate BioGraphNet's contribution to the overall model performance; secondly, The variant replaces PCRFE with a standard CNN, removes the SE block and multi-head convolutional self-attention to evaluate the independent role of CNN in protein feature extraction; furthermore, to evaluate the contribution of multi-head convolutional self-attention to protein feature extraction, we designed... A variant, removing the CNN components and SE block from PCRFE, was used to verify the effectiveness of the multi-head convolutional self-attention mechanism; finally, we designed... A variant, replacing the CMIF module with a simpler splicing method, was used to evaluate the contribution of CMIF in the feature fusion process. The ablation experimental results are shown in Table 4. Figure 2As shown, experimental results indicate that removing any module leads to varying degrees of performance degradation in the model. Among these, removing the CNN component and the SE block from the protein feature extraction module resulted in the most significant decreases in AUROC and AUPRC, demonstrating the crucial role of CNN in extracting local protein structural features. Furthermore, removing the CMIF module also caused a significant performance degradation, further illustrating its importance in modeling the complex interactions between drugs and targets. In contrast, removing the multi-head convolutional self-attention from the BioGraphNet and PCRFE modules had relatively smaller impacts on performance, but still showed varying degrees of performance degradation, validating the important role of these two modules in the feature extraction process. It is worth noting that… The performance degradation of the variants was relatively small, but relying solely on CNNs for local feature extraction made it difficult to fully capture long-range dependencies in protein sequences. Overall, the complete model achieved the best performance, verifying that the synergistic effect between modules is crucial for improving the overall performance of the model.

[0076] Table 4 Ablation experimental results on BindingDB, BioSNAP, and DAVIS datasets

[0077] .

[0078] Parameter sensitivity analysis To evaluate the impact of hyperparameters on model performance, we conducted a parameter sensitivity analysis. We mainly studied two key hyperparameters: the learning rate and the number of attention heads. These two hyperparameters have a significant impact on the model training process and the final performance.

[0079] The learning rate is one of the key hyperparameters affecting model training efficiency and performance. To determine the optimal learning rate, we evaluated five different learning rate configurations: 5e-05, 2e-05, 1e-05, 2e-04, and 1e-04, and compared model performance using metrics such as AUROC. Figure 3 As shown, the model performs best across all performance metrics when the learning rate is 5e-05. As the learning rate decreases, especially when set to 2e-05 and 1e-05, the model performance declines. This indicates that an excessively small learning rate slows down the training process and fails to effectively capture complex features. Furthermore, when the learning rate is increased to 2e-04 and 1e-04, the model performance further declines. Therefore, we choose 5e-05 as the optimal learning rate configuration.

[0080] The number of attention heads is a key hyperparameter in the multi-head attention mechanism and has a significant impact on model performance. To evaluate the impact of different numbers of attention heads on model performance, we conducted experiments with configurations of 2, 4, and 8 attention heads, respectively. Figure 4 The AUROC curves of the models on the validation set for each configuration are shown. The results indicate that the model with 4 attention heads performs best on the BindingDB and BioSNAP datasets. In contrast, the model with 2 attention heads performs weaker, and although the model with 8 attention heads shows some improvement, it still does not perform as well as the configuration with 4 attention heads. This suggests that a smaller number of attention heads may not be able to fully capture complex features, while an excessive number of attention heads may increase the computational burden without bringing significant performance improvements. Therefore, we choose 4 attention heads as the optimal configuration. It is worth noting that on the DAVIS dataset, the AUROC curves for all attention head configurations exhibit significant fluctuations. This phenomenon is mainly attributed to the small sample size of the DAVIS dataset, which leads to a significant increase in the variance of the evaluation metrics during each model evaluation, resulting in high volatility in the AUROC curves. Small sample datasets typically lead to instability and volatility in model evaluation metrics in machine learning experiments. Although the experimental results for different attention head numbers on the DAVIS dataset show significant fluctuations, the overall trend and final performance can still serve as a reference for the quality of the model.

[0081] Experimental analysis of the cold pairing partitioning strategy on the Human dataset The cold-pairing partitioning strategy ensures that all drugs and proteins in the test set are not present during training, thus preventing predictions on test data from relying solely on known drug or protein features. This strategy avoids model dependence on known data features, thereby more realistically reflecting the model's evaluation results. Previous research has pointed out that the Human dataset exhibits significant ligand bias, which may cause the model to tend to learn specific drug features while ignoring the true interaction patterns between drugs and targets (CHEN L, TAN X, WANG D, et al. TransformerCPI:improving compound–protein interaction prediction by sequence-based deep learning with self-attention mechanism and label reversal experiments[J / OL].

[2025] [2025-04-15]. https: / / dx.doi.org / 10.1093 / bioinformatics / btaa524.). Compared to other datasets, the ligand bias in the Human dataset is more pronounced, making model performance easily overestimated. Therefore, we implement a cold-pairing partitioning strategy on the Human dataset, randomly assigning 5% and 10% of the data to the test set. The DTI was applied to the validation and test sets, and relevant drugs and targets were removed from the training set to ensure that the model could not access the test data.

[0082] Experimental results show that all models exhibit good predictive performance under random partitioning conditions; however, performance generally drops significantly after adopting a cold-pairing partitioning strategy. Our BPCI-DTI model was compared with other models on the Human dataset, and the results are as follows: Figure 5 As shown, although the cold-pairing partitioning strategy leads to an overall performance decrease, the BPCI-DTI model still maintains a high level of performance on the Human dataset, significantly outperforming other models. This verifies the robustness of the BPCI-DTI model under complex data distributions, especially its performance on the Human dataset with ligand bias. This result indicates that the cold-pairing partitioning strategy helps to more accurately evaluate the model's generalization ability and reduce performance overestimation caused by training data feature bias. Compared to random partitioning, the cold-pairing partitioning strategy can more realistically simulate DTI prediction tasks in real-world application scenarios, improving the model's ability to cope with unknown drugs and targets.

[0083] Case Studies Although the BPCI-DTI model demonstrates excellent performance on multiple benchmark datasets, these results alone are insufficient to fully demonstrate its practical application capability in identifying novel DTIs. To further validate the model's reliability, we conducted case studies from both drug and target perspectives. First, we removed the drug-target data used in the case studies during model training. Then, we predicted these drug-target pairs and ranked them based on the predicted interaction probabilities. Finally, we validated all predicted DTI results against data in the DrugBank database, further confirming the reliability of the BPCI-DTI model in predicting drug-target interactions.

[0084] In the drug space, we selected Dasatinib (DB01254) and Haloperidol (DB00502) as research subjects. Dasatinib is a drug that targets tyrosine kinases and is widely used to treat chronic myeloid leukemia (CML) and acute lymphoblastic leukemia (ALL). Haloperidol is a classic antipsychotic drug that is widely used to treat schizophrenia, mania, and other diseases. According to the prediction results of the BPCI-DTI model, as shown in Table 5, Dasatinib and Haloperidol interact with multiple targets, and almost all of the top 10 predicted targets were validated in the DrugBank database. In addition, the BPCI-DTI model also predicted that DB01254 may interact with O00238.

[0085] In the target space, we selected Alpha adrenergic receptor (P35348) and D(2) dopamine receptor (P14416) as research subjects. Alpha adrenergic receptor plays an important role in blood pressure regulation and cardiac function regulation, while D(2) dopamine receptor plays a core role in signal transduction in the nervous system. According to the prediction results of the BPCI-DTI model, as shown in Table 6, among the top 10 predicted target drugs, only the interaction between DB04216 and P14416 was not verified.

[0086] Table 5. Top 10 predicted targets interacting with Dasatinib (DB01254) and Haloperidol (DB00502)

[0087]

[0088] To further verify the interactions predicted by the BPCI-DTI model (i.e., the interaction between DB01254 and O00238, and the interaction between P14416 and DB04216), we conducted molecular docking experiments using AutoDock4. The docking results are as follows: Figure 6 and Figure 7 As shown, the binding energy between DB01254 and O00238 is -6.7 kcal / mol, and the binding energy between P14416 and DB04216 is -7.5 kcal / mol. The magnitude of binding energy is crucial in drug design and pharmacological research, providing an important theoretical basis for drug target selection. These binding energy values ​​indicate that the predicted drug-target pairs have strong binding stability, further validating the reliability of the BPCI-DTI model prediction results.

[0089] Table 6. Top 10 drugs predicted to interact with Alpha Adrenergic Receptor (P35348) and D(2) Dopamine Receptor (P14416)

[0090] .

[0091] In this study, we proposed a novel drug-target interaction prediction model—BPCI-DTI. Experimental results show that BPCI-DTI outperforms existing mainstream models on multiple benchmark datasets and its reliability has been verified in different application scenarios. We believe that BPCI-DTI can not only promote virtual drug screening but also provide strong support for drug reuse and new drug development. Future research can further improve the interpretability of the model and explore its broad applicability in large-scale drug discovery and biological research.

[0092] The above embodiments are illustrative of the present invention and are not intended to limit the present invention. Any simple modifications to the present invention are within the scope of protection of the present invention.

Claims

1. A method for extracting structural graph features through cross-modal interactive fusion, characterized in that, Includes the following steps: S1. Input the SMILES string of the drug molecule into the drug feature encoding module to generate a structured representation suitable for graph neural network processing; and input the amino acid sequence of the protein into the protein feature encoding module to obtain the initial feature representation of the protein. S2. The drug features are further processed by the biomolecular graph feature extraction module, and the local neighborhood information and global topological features of the drug molecules are extracted by using a multi-level graph isomorphic network. Protein features are processed by the protein cross-residue feature extraction module, which captures local features and long-range dependencies of protein sequences through convolutional neural networks and multi-head convolutional self-attention mechanisms. S3. Input the feature representations of drugs and proteins into the cross-modal interaction fusion module, and adaptively model the deep interaction relationship between drugs and targets through a gating mechanism; S4. The fused features are processed by pooling and feedforward neural networks to generate the final drug-target interaction prediction results.

2. The cross-modal interactive fusion structural graph feature extraction method as described in claim 1, characterized in that, The working steps of the drug feature encoding module are as follows: S1.1 First, the SMILES string is parsed into a drug molecule diagram using the open-source toolkit RDKit; then, a 78-dimensional binary feature vector is generated for each atom node in the drug molecule diagram using the DeepChem tool. S1.2 Then, an adjacency matrix is ​​introduced to represent the connection relationship between atoms in the drug molecule. If there is a connection between two atoms, the corresponding matrix element is 1; otherwise, it is 0. S1.3 Finally, we concatenate the feature vectors of each atom to form a unified feature representation, and normalize the features to serve as the input for the final drug feature extraction module.

3. The cross-modal interactive fusion structural graph feature extraction method as described in claim 1, characterized in that: In the protein feature encoding module, the length of all protein sequences is normalized to 1200. When the length of a protein sequence is less than 1200, it is padded with zeros. When the length exceeds 1200, the first 1200 amino acid characters are truncated. If some letters are not in the defined letter set during the encoding process, these characters will be ignored and treated as padding characters. In order to enhance the feature representation capability, integer codes are input into the embedding layer, and each integer is mapped to a high-dimensional continuous vector to generate the feature matrix of the protein sequence.

4. The cross-modal interactive fusion structural graph feature extraction method as described in claim 1, characterized in that, The working steps of the biomolecular graph feature extraction module are as follows: S2A.1: By using a multi-channel structure to process the features of direct and indirect neighbors in the molecular graph, each channel contains multiple GIN layers to update node features recursively. The first channel of the biomolecular graph feature extraction module takes node features and the direct adjacency matrix as input to focus on extracting local topological features; the second channel takes node features and the indirect adjacency matrix as input to capture multi-hop neighborhood information between nodes. S2A.2 After completing the multi-channel feature extraction, the biomolecular graph feature extraction module uses a feature splicing fusion strategy to integrate the outputs of the two channels; the local topological features extracted from the two channels are spliced ​​with the multi-hop neighborhood features to obtain a richer and more comprehensive node feature representation.

5. The cross-modal interactive fusion structural graph feature extraction method as described in claim 1, characterized in that: The protein cross-residue feature extraction module extracts local amino acid features through convolution and uses SE blocks to adaptively adjust channel weights to enhance the expression of important features. At the same time, it introduces a multi-head convolutional self-attention mechanism to comprehensively capture local and global features in the protein sequence.

6. The cross-modal interactive fusion structural graph feature extraction method as described in claim 5, characterized in that: The specific working steps of the protein cross-residue feature extraction module are as follows: S2B.1: Extracting local amino acid features from protein sequences by combining convolution operations with SE blocks: The input integer-encoded protein sequence is processed by successive one-dimensional convolutions and the LeakyReLU activation function to capture local dependencies in the protein sequence. The introduction of SE blocks adaptively adjusts the weights of each channel feature, thereby enhancing the expression of important features while suppressing redundant information and low-correlation features in the channel dimension to a certain extent. S2B.2 To capture long-range dependencies in protein sequences, a multi-head convolutional self-attention mechanism is introduced. By combining convolution and self-attention mechanisms, the model's ability to capture long-range information interactions in protein sequences is improved.

7. The cross-modal interactive fusion structural graph feature extraction method as described in claim 1, characterized in that: The cross-modal interaction fusion module captures the bidirectional cross-modal feature interactions between drugs and proteins through a bidirectional cross-attention mechanism, and uses a gating mechanism to perform weighted fusion of these features.

8. The cross-modal interactive fusion structural graph feature extraction method as described in claim 7, characterized in that: In the bidirectional cross-attention mechanism, in order to ensure the stability of training, the features of drugs and proteins are first processed by layer normalization, and then multi-head cross-attention is used to calculate the attention features from drugs to proteins and the attention features from proteins to drugs respectively. Multi-head attention mechanisms allow models to capture the interactions between drug and protein features in parallel from different subspaces.

9. The cross-modal interactive fusion structural graph feature extraction method as described in claim 7, characterized in that, The specific steps for weighted fusion using a gating mechanism are as follows: First, the output features are processed through a fully connected layer and a sigmoid function to calculate the fusion weights of drug and protein features. Then, these two fusion weights are used to weight the drug and protein features element by element. Subsequently, average pooling is used to reduce the dimensionality of the gated fusion features. The final pooled feature representation can be directly input into the subsequent feedforward network for the prediction of drug-target interactions.

10. A drug-target interaction prediction model, performing the cross-modal interaction fusion structural graph feature extraction method as described in claim 1, characterized in that, It includes a biomolecular graph feature extraction module, a protein cross-residue feature extraction module, and a cross-modal interaction fusion module. The biomolecular graph feature extraction module uses a multi-level graph isomorphic network to extract local neighborhood information and global topological features of drug molecules in parallel. The protein cross-residue feature extraction module captures local features and long-range dependencies of protein sequences through convolutional neural networks and multi-head convolutional self-attention mechanisms. The cross-modal interaction fusion module realizes deep interaction fusion of drug and protein features through a gating mechanism.

Images