A dual-flow collaborative attention and sparse feature fusion drug-target interaction prediction method
By fusing dual-stream collaborative attention with sparse features, and combining multi-scale spatial attention and channel-enhanced attention mechanisms, this method addresses the problem of insufficient substructure feature capture in drug-target interaction prediction in existing methods, thereby improving prediction accuracy and robustness and shortening the drug screening cycle.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- YANGTZE DELTA REGION INST (QUZHOU) UNIV OF ELECTRONIC SCI & TECH OF CHINA
- Filing Date
- 2025-08-13
- Publication Date
- 2026-06-16
AI Technical Summary
Existing methods are unable to effectively capture the substructure discrimination features and key binding region features of drug-target interaction pairs, resulting in insufficient accuracy in drug-target interaction prediction.
We employ a dual-stream collaborative attention and sparse feature fusion approach. We extract substructural features of drugs and targets through multi-scale spatial attention and channel-enhanced attention mechanisms, suppress noise using sparse self-attention mechanisms, and model feature interactions between drugs and targets using cross-attention mechanisms to improve prediction performance.
It enhances the ability to identify subtle differences in biomolecular structures, improves the accuracy and robustness of drug-target interaction prediction, shortens the drug screening cycle, reduces dependence on high-purity experimental data, and improves the generalization ability of cross-dataset prediction.
Smart Images
Figure CN120954489B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computational biology technology, specifically relating to a drug-target interaction prediction method that integrates two-stream collaborative attention with sparse features. Background Technology
[0002] Accurate identification of drug-target interactions (DTIs) is crucial for therapy development, laying the foundation for new drug discovery and drug reuse. While traditional biochemical methods for validating DTIs have demonstrated high reliability, their time-consuming nature and significant resource investment limit their widespread application. Researchers have subsequently proposed novel methods such as molecular docking and molecular dynamics simulations. These methods typically require high-quality experimental data, and the accuracy of the prediction models decreases when the three-dimensional (3D) structural information of the target is unknown. Therefore, developing efficient computational methods for DTI prediction remains a key challenge in the field of drug discovery.
[0003] The development of deep learning technology has driven its widespread application in drug-target interaction (DTI) prediction. Unlike traditional methods that rely on predefined descriptors, deep learning methods can automatically learn the characterization of drugs and target proteins directly from heterogeneous data. For example, DeepDTA developed a novel convolutional neural network (CNN) to extract discriminative features from drug-protein sequences and improved prediction accuracy through data augmentation strategies and architectural complexity optimization. DeepCDA, based on the DeepDTA framework, applies a long short-term memory (LSTM) layer after each CNN layer to effectively simulate interactions between biological sequences to predict DTIs. Similarly, DeepConv-DTI utilizes multilayer perceptrons (MLPs) combined with one-dimensional CNN layers to extract extended connection fingerprint (ECFP) embeddings of drugs and target amino acid sequence features. Furthermore, DTI-RCNN integrates gene-drug association information and combines CNNs with LSTMs for DTI prediction. However, these methods rely solely on sequence data, leading to the loss of structural information and insufficient molecular characterization.
[0004] Graph Neural Networks (GNNs) have emerged as a promising approach for drug-target interaction (DTI) prediction, leveraging their ability to model molecular relationships while integrating structural features of both the target and the drug. GraphDTA extracts generative graph representations from atomic features of drug sequences and generates drug embeddings using four GNN variants. MGraphDTA improves prediction accuracy in DTA tasks by integrating the drug's 3D conformation with the target's secondary structure to achieve multi-scale representation learning. TDGraphDTA integrates a diffusion mechanism before refining the graph structural representation, enhancing the model's ability to extract discriminative features. Meanwhile, CE-DTI repositions interaction prediction as a node classification task, with GNNs learning drug-target representations from multi-source data through automatic graph generation. However, these methods independently model drug and target structures, failing to capture the contribution of their respective internal interactions.
[0005] In recent years, attention-based deep learning models have demonstrated strong performance and significant application potential in drug-target interaction (DTI) prediction. For example, MATT_DTI utilizes a multi-head attention strategy, projecting drug embeddings as queries and using target sequences as key-value pairs for cross-modal alignment. DrugBAN models DTI prediction as a multimodal fusion task, capturing local structural dependencies through a bilinear attention mechanism. Meanwhile, TransformerCPI encodes drug-target pairs as bimodal sequences, overcoming the long-range dependency limitations of recurrent neural networks (RNNs) through multi-head gated cross-attention. HyperAttentionDTI employs a hierarchical attention mechanism, achieving precise localization of binding sites by dynamically adjusting bilayer weights. MI-DTI combines attention heatmap visualization, providing an intuitive spatial matrix for key intermolecular binding regions. However, these methods often fail to capture the key binding regions of drug-target interaction pairs, generating redundant representations. Despite the strong predictive performance of existing methods, their molecular characterization capabilities are still limited by sequence or graph structure data, failing to capture substructure discriminative features. Furthermore, the critical binding region of DTI, which is essential for improving prediction performance, was not adequately captured.
[0006] In summary, in order to address the problem that existing methods cannot capture the substructure discrimination features and the key binding region features of drug-target interaction pairs, it is essential to propose a new processing method for accurate prediction of drug-target interactions. Summary of the Invention
[0007] The purpose of this invention is to address the problem that existing methods cannot capture the substructure discrimination features and the key binding region features of drug-target interaction pairs, and proposes a drug-target interaction prediction method that fuses dual-stream collaborative attention and sparse features.
[0008] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a drug-target interaction prediction method that fuses dual-stream collaborative attention and sparse features, the method specifically including the following steps:
[0009] Step S1: Obtain drug and target data sample pairs, including positive sample pairs and negative sample pairs;
[0010] Step S2: Process the drug data and target data in each data sample pair separately to obtain the processed drug and target data sample pairs;
[0011] Step S3: Extract the drug structure features and target amino acid sequence features of each processed drug and target data sample pair;
[0012] Step S4: Extract the spatial features of the drug based on the drug structure features of the data sample pair, then extract the channel features of the drug based on the spatial features of the drug, and process the extracted drug channel features to obtain enhanced drug channel features.
[0013] Based on the structural features of the target amino acid sequence of the data sample pair, the spatial features of the target amino acid sequence are extracted. Then, based on the spatial features of the target amino acid sequence, the channel features of the target amino acid sequence are extracted. The extracted channel features of the target amino acid sequence are then processed to obtain the enhanced channel features of the target amino acid sequence.
[0014] Step S5: Sparsify the enhanced drug channel features using a sparse self-attention mechanism to obtain sparse drug features. The enhanced target amino acid sequence channel features were sparsified using a sparse self-attention mechanism to obtain sparse target amino acid sequence features. ;
[0015] Then, the sparse drug signature and the sparse target amino acid sequence signature are cross-fused to obtain the cross-fused drug signature. and the target amino acid sequence characteristics after cross-fusion ;
[0016] Step S6: Based on the enhanced drug channel feature D, the enhanced target amino acid sequence channel feature P, and the cross-fused drug feature... and the target amino acid sequence characteristics after cross-fusion The joint representation of the sample pair consisting of drug and target is obtained. ;
[0017] Step S7: Combine the representations Input the MLP, predict the interaction probability of the sample pair composed of drug and target through the MLP, calculate the loss function based on the predicted drug-target interaction probability and the label of the sample pair composed of drug and target, and adjust the parameters of each layer in reverse according to the loss function until the loss function converges and continue to execute step S8.
[0018] Step S8: Process the drug-target pair to be predicted using steps S2 to S6 to obtain the joint representation of the drug-target pair to be predicted. Input the joint representation of the drug-target pair to be predicted into the MLP to obtain the interaction probability of the drug-target pair to be predicted.
[0019] If the probability of drug-target interaction is greater than the threshold, then the drug-target pair to be predicted has an interaction; otherwise, the drug-target pair to be predicted does not have an interaction.
[0020] The beneficial effects of this invention are:
[0021] This invention employs a dual-stream collaborative attention strategy combining multi-scale spatial attention and channel-enhanced attention to collaboratively capture discriminative features of substructures. The multi-scale spatial attention mechanism utilizes multi-branch convolutional layers to adaptively integrate spatial substructure features, enhancing the molecular representation of each substructure; the channel-enhanced attention mechanism alleviates the inconsistency of substructure features. The sparse attention mechanism can highlight key features while suppressing noise, while the cross-attention mechanism improves the extraction capability of key binding region features by modeling the feature interactions between the drug and the target, thereby enhancing the predictive performance of drug-target interactions.
[0022] The method of the present invention has the following advantages:
[0023] 1. Enhance the model's ability to identify subtle differences in biomolecular structures, improve the biological relevance of feature representations through the synergistic effect of multiple mechanisms, and make the prediction results more consistent with the laws of real molecular interactions.
[0024] 2. It reduces the reliance on high-purity experimental data. The sparse attention mechanism enhances the model's robustness in complex biological sample data by suppressing noise, thus reducing data preprocessing costs.
[0025] 3. Shortening the drug screening cycle: The accurate modeling of cross-attention pairs’ feature interactions accelerates the screening efficiency of potential drug-target pairs, providing more efficient computational support for early drug development.
[0026] 4. Expand the applicability of the invention in multiple scenarios. The collaborative attention strategy can be flexibly adapted to different molecular data types, including small molecule compounds and protein sequences, and improve the generalization ability of cross-dataset prediction. Attached Figure Description
[0027] Figure 1 This is a flowchart of a drug-target interaction prediction method based on the fusion of dual-stream collaborative attention and sparse features according to the present invention;
[0028] Figure 2a This is a comparison chart of the ROC curve prediction performance obtained by the present invention and other comparative methods through a five-fold cross-validation experiment;
[0029] Figure 2b This is a comparison chart of the PRC curve prediction performance obtained by the present invention and other comparative methods through a five-fold cross-validation experiment;
[0030] Figure 3 This is a comparison chart of the prediction performance of AUROC and AUPR of the present invention and other comparative methods in different scenarios. Detailed Implementation
[0031] Specific Implementation Method 1: The drug-target interaction prediction method fused with dual-stream collaborative attention and sparse features described in this implementation method specifically includes the following steps:
[0032] Step S1: Obtain drug and target data sample pairs, including positive sample pairs and negative sample pairs;
[0033] Step S2: Process the drug data and target data in each data sample pair separately to obtain the processed drug and target data sample pairs;
[0034] Step S3: Extract the drug structure features and target amino acid sequence features of each processed drug and target data sample pair;
[0035] Step S4: Extract the spatial features of the drug based on the drug structure features of the data sample pair, then extract the channel features of the drug based on the spatial features of the drug, and process the extracted drug channel features to obtain enhanced drug channel features.
[0036] Based on the structural features of the target amino acid sequence of the data sample pair, the spatial features of the target amino acid sequence are extracted. Then, based on the spatial features of the target amino acid sequence, the channel features of the target amino acid sequence are extracted. The extracted channel features of the target amino acid sequence are then processed to obtain the enhanced channel features of the target amino acid sequence.
[0037] Step S5: Sparsify the enhanced drug channel features using a sparse self-attention mechanism to obtain sparse drug features. The enhanced target amino acid sequence channel features were sparsified using a sparse self-attention mechanism to obtain sparse target amino acid sequence features. ;
[0038] Then, the sparse drug signature and the sparse target amino acid sequence signature are cross-fused to obtain the cross-fused drug signature. and the target amino acid sequence characteristics after cross-fusion ;
[0039] Step S6: Based on the enhanced drug channel feature D, the enhanced target amino acid sequence channel feature P, and the cross-fused drug feature... and the target amino acid sequence characteristics after cross-fusion The joint representation of the sample pair consisting of drug and target is obtained. ;
[0040] Step S7: Combine the representations Input the MLP, predict the interaction probability of the sample pair composed of drug and target through the MLP, calculate the loss function based on the predicted drug-target interaction probability and the label of the sample pair composed of drug and target, and adjust the parameters of each layer in reverse according to the loss function until the loss function converges and continue to execute step S8.
[0041] Step S8: Process the drug-target pair to be predicted using steps S2 to S6 to obtain the joint representation of the drug-target pair to be predicted. Input the joint representation of the drug-target pair to be predicted into the MLP to obtain the interaction probability of the drug-target pair to be predicted.
[0042] If the probability of drug-target interaction is greater than a threshold (in this invention, the threshold value is set to 0.5), then the drug-target pair to be predicted has an interaction; otherwise, the drug-target pair to be predicted does not have an interaction.
[0043] Specific Implementation Method Two: This implementation method is a further limitation of Specific Implementation Method One. The drug data inside the sample pair is drug SMILES string data, and the target data inside the sample pair is target amino acid sequence data.
[0044] The other steps and parameters are the same as in Specific Implementation Method 1.
[0045] Specific Implementation Method Three: This implementation method further defines Specific Implementation Method Two. The processing of drug data within each data sample pair specifically involves:
[0046] The drug SMILES string data was converted into an atomic interaction diagram using the open-source dgl tool.
[0047] The other steps and parameters are the same as in Specific Implementation Method Two.
[0048] Specific Implementation Method Four: This implementation method further defines Specific Implementation Method Three. The processing of the target data within each data sample pair specifically involves:
[0049] Step S21: Divide the target amino acid sequence into several fragments of equal length. If there are identical fragments among the divided fragments, only one identical fragment is retained. If there are no identical fragments among the divided fragments, all the divided fragments are retained.
[0050] Step S22: Convert each retained segment into a digital vector, and use the obtained digital vectors to form an embedding matrix.
[0051] The other steps and parameters are the same as in Specific Implementation Method 3.
[0052] Specific Implementation Method Five: This implementation method is a further limitation of Specific Implementation Method Four. The specific process of step S3 is as follows:
[0053] Step S31: For drugs: Extract drug structural features from the atomic interaction diagram using the GCN model; specifically:
[0054]
[0055] in: This represents an adjacency matrix with self-connections; In the GCN model, the first... Layer weight matrix; In the GCN model, the first... Layer bias vector; It is the first in the GCN model The output of the layer, , The output of the atomic interaction diagram of the drug after the embedding operation; It is the first in the GCN model Layer output; Represents the nonlinear activation function ReLU;
[0056] The output of the last layer in the GCN model is used as the drug structure feature;
[0057] Step S32: For the target: Use the MKCNN model to process the embedding matrix and extract the features of the target amino acid residue sequence; specifically:
[0058]
[0059] in: Indicates the batch normalization layer; The MKCNN model represents the first... One convolutional layer; Indicates the first The kernel size of each convolutional layer; The MKCNN model represents the first... The weight matrix of each convolutional layer; The MKCNN model represents the first... The bias vectors of each convolutional layer; In the MKCNN model, the first... The output of each convolutional layer; In the MKCNN model, the first... The output of each convolutional layer;
[0060] The output of the last ReLU activation function layer is used as a feature of the target amino acid residue sequence.
[0061] The other steps and parameters are the same as in Specific Implementation Method Four.
[0062] Specific Implementation Method Six: This implementation method is a further limitation of Specific Implementation Method Five. The specific process of step S4 is as follows:
[0063] Step S41: Perform dimensional transformation on the drug structural features:
[0064] The structural characteristics of the drug are denoted as , ,in, Represent real numbers, Indicates batch size, Indicates the number of nodes. Let be the size of the drug mapping dimension; Regarding the structural characteristics of drugs Perform dimensional transformation to obtain the transformed features. , , , Indicates the actual number of atoms in the drug;
[0065] Step S42: Transform the features along the height and width directions. Decomposition yields spatial substructures along the width direction. and spatial substructures in the height direction ;
[0066] and features Divided into Each spatial substructure will feature Divided into Each spatial substructure includes [number] spatial substructures. One channel;
[0067]
[0068]
[0069] in, Indicates along The first axis assignment obtained Each spatial substructure feature Indicates along The first axis assignment obtained Each spatial substructure feature; =1,2,…, ;
[0070] Step S43: Extract the spatial feature matrix of each spatial substructure obtained in step S42 by depthwise separable convolution;
[0071]
[0072]
[0073] in, express The spatial feature matrix, express The spatial feature matrix, This represents a depthwise separable convolutional layer;
[0074] The spatial attention matrix is obtained by aggregating and grouping the spatial feature matrices of the spatial substructures:
[0075]
[0076]
[0077] in, Indicates will Add;
[0078] Indicates will Add;
[0079] This indicates that group normalization is performed along the width direction;
[0080] This represents performing group normalization along the height direction;
[0081] Indicates along The spatial attention matrix obtained from the axis;
[0082] Indicates along The spatial attention matrix obtained from the axis;
[0083] Step S44, will , and Multiplying yields the result of multiplication. The result of multiplication The drug's channel features are extracted by sequentially passing through a global average pooling layer and a depthwise separable convolutional layer.
[0084]
[0085] in, This indicates a global average pooling operation. This indicates the extracted drug pathway characteristics;
[0086] Step S45: Process the extracted drug channel features using a single-head self-attention mechanism to obtain enhanced drug channel features. .
[0087] The other steps and parameters are the same as in Specific Implementation Method 5.
[0088] Specific Implementation Method Seven: This implementation method is a further limitation of Specific Implementation Method Six. The specific process of step S45 is as follows:
[0089]
[0090]
[0091]
[0092] in, , and These represent queries generated through linear mapping. ,key Sum , yes The dimension, where the superscript T indicates transpose. This indicates a global average pooling operation. It is an enhanced drug pathway feature.
[0093] The other steps and parameters are the same as in Specific Implementation Method Six.
[0094] Specific Implementation Method Eight: This implementation method further defines Specific Implementation Method Seven. It utilizes a sparse self-attention mechanism to sparsify the enhanced drug channel features, resulting in sparse drug features. The enhanced target amino acid sequence channel features were sparsified using a sparse self-attention mechanism to obtain sparse target amino acid sequence features. Specifically:
[0095] Taking the sparsification of enhanced drug channel features as an example:
[0096]
[0097]
[0098] in, It is the weight of the i-th attention head;
[0099] It is a Hadama pile;
[0100] , and This represents the query, key, and value of D after the i-th head mapping;
[0101] This indicates the percentage of element values retained in each row of D. , This indicates that a binary mask operation is performed based on the index corresponding to the retained element value;
[0102] Indicates the total number of heads of attention;
[0103] Indicates the dimension of D;
[0104] This indicates a sparse drug characteristic.
[0105] Similarly, sparse target amino acid sequence characteristics were obtained. .
[0106] The other steps and parameters are the same as in Specific Implementation Method Seven.
[0107] Specific Implementation Method Nine: This implementation method further defines Specific Implementation Method Eight. The method involves cross-fusing sparse drug characteristics and sparse target amino acid sequence characteristics to obtain cross-fused drug characteristics. and the target amino acid sequence characteristics after cross-fusion Specifically:
[0108] Sparse drug features are transformed by linear transformation Mapped to query s and keys By using linear transformation, the sparse target amino acid sequence features are transformed. Mapped to query s and keys ;
[0109]
[0110]
[0111]
[0112]
[0113] in, express Dimensions express Dimensions.
[0114] The other steps and parameters are the same as in Specific Implementation Method 8.
[0115] Specific Implementation Method Ten: This implementation method is a further limitation of Specific Implementation Method Nine. The specific process of step S6 is as follows:
[0116]
[0117] in, and They are Obtained through average pooling operation Indicates multiplication. This indicates feature splicing.
[0118] The other steps and parameters are the same as in Specific Implementation Method Nine.
[0119] Example
[0120] like Figure 1 As shown, this embodiment proposes a drug-target interaction prediction method that fuses two-stream collaborative attention with sparse features. The drug-target interaction in this invention is represented by the probability of drug-target interaction. The method specifically includes the following steps:
[0121] Step S1: Obtain drug and target data sample pairs by collection or download. The sample pairs include positive sample pairs and negative sample pairs. The drug data in the sample pair is the drug SMILES string data, and the target data in the sample pair is the target amino acid sequence data.
[0122] The acquired drug and target data came from three datasets: BindingDB, BioSNAP, and Human. Each dataset provides unique features that support robust evaluation.
[0123] BindingDB is a public database containing experimentally measured binding affinity data for drug-target interactions, focusing on pharmacologically relevant target-ligand complexes. Drug-target pairs with affinity greater than 0.5 are designated as positive pairs, labeled 1, while those with affinity less than or equal to 0.5 are designated as negative pairs, labeled 0.
[0124] The BioSNAP dataset integrates the drug database and the KEGG drug dataset, containing 13,830 positive samples with randomly paired positive factors and negative samples with paired negative factors. That is, it includes both drug-target data pairs with binding affinity greater than or equal to 0.5 and drug-target data pairs with binding affinity less than 0.5.
[0125] The Human dataset is specifically designed for studying highly reliable negative sample pairs.
[0126] Step S2: Process the drug data and target data within each data sample pair separately to obtain processed drug and target data sample pairs; wherein:
[0127] For drug data: The drug SMILES string data is converted into an atomic interaction graph using the open-source dgl tool;
[0128] For target data: the target amino acid sequence is converted into an embedding matrix vector using word embedding; specifically:
[0129] Step S21: Divide the target amino acid sequence into several fragments of equal length. If there are identical fragments among the divided fragments, only one identical fragment is retained, and the fragment that appears first is retained. If there are no identical fragments among the divided fragments, all the divided fragments are retained.
[0130] Step S22: Convert each retained fragment into a digital vector. According to the order in which each fragment appears in the target amino acid sequence, use each digital vector as a row of the embedding matrix. Use the obtained digital vectors to form the embedding matrix.
[0131] After processing the drug data and target data within a sample pair, the sample pair consists of an atomic interaction diagram corresponding to the drug and an embedding matrix corresponding to the target.
[0132] Step S3: Extract the drug structure features and target amino acid sequence features of each processed drug and target data sample pair;
[0133] Step S31: For drugs: Extract drug structural features from the atomic interaction diagram using the GCN model; specifically:
[0134]
[0135] in: This represents an adjacency matrix with self-connections, where the atomic interaction graph contains a total of [number missing] atoms. If there are 10 atoms, then the matrix The dimension is × If two atoms are connected on an atomic interaction diagram, then these two atoms are in the matrix. The corresponding element value in the matrix is 1; otherwise, the two atoms are in the matrix. The corresponding element in the matrix has a value of 0, and the matrix... All elements on the diagonal are 1;
[0136] In the GCN model, the first... The weight matrix of the layer, In the GCN model, the first... Layer bias vector;
[0137] It is the first in the GCN model The output of the layer, , The output represents the atomic interaction diagram of the drug after the embedding operation (i.e., after processing with an embedding layer);
[0138] It is the first in the GCN model Layer output;
[0139] Represents the nonlinear activation function ReLU;
[0140] The output of the last layer in the GCN model is used as the drug structure feature;
[0141] Step S32: For the target: Use the MKCNN model to process the embedding matrix and extract the features of the target amino acid residue sequence; specifically:
[0142]
[0143] in: Indicates the batch normalization layer;
[0144] The MKCNN model represents the first... One convolutional layer;
[0145] Indicates the first The kernel size of each convolutional layer;
[0146] The MKCNN model represents the first... The weight matrix of each convolutional layer;
[0147] The MKCNN model represents the first... The bias vectors of each convolutional layer;
[0148] In the MKCNN model, the first... The output of each convolutional layer The embedding matrix obtained in step S22;
[0149] In the MKCNN model, the first... The output of each convolutional layer;
[0150] The output of the last ReLU activation function layer is used as a feature of the target amino acid residue sequence;
[0151] In step S3 of this invention, the embedding representations of the drug and target are finally obtained through steps S31 to S34. It should be noted that the MKCNN model includes four convolutional layers. The MKCNN model used in this invention includes, in sequence, a first convolutional layer with a kernel size of 3×3, a first BN layer, a first ReLU activation function layer, a second convolutional layer with a kernel size of 5×5, a second BN layer, a second ReLU activation function layer, a third convolutional layer with a kernel size of 7×7, a third BN layer, a third ReLU activation function layer, a fourth convolutional layer with a kernel size of 9×9, a fourth BN layer, and a fourth ReLU activation function layer.
[0152] Step S4: Design a dual-stream collaborative attention strategy. Extract the spatial features of the drug based on the drug structure features of the data sample pair, then extract the channel features of the drug based on the spatial features of the drug, and process the extracted drug channel features to obtain enhanced drug channel features.
[0153] Based on the structural features of the target amino acid sequence of the data sample pair, the spatial features of the target amino acid sequence are extracted. Then, based on the spatial features of the target amino acid sequence, the channel features of the target amino acid sequence are extracted. The extracted channel features of the target amino acid sequence are then processed to obtain the enhanced channel features of the target amino acid sequence.
[0154] Step S4 of this invention extracts substructure features from the spatial and channel dimensions to obtain a discriminative representation of the substructure. The following detailed explanation uses a drug as an example:
[0155] Step S41: Perform dimensional transformation on the drug structural features;
[0156] The structural characteristics of the drug are denoted as , ,in, Represent real numbers, Indicates batch size, This indicates the number of nodes (for drugs, each atom is considered a node; when the total number of atoms is less than [a certain number], the node is considered a separate node). When the total number of amino acid residues is insufficient, zero-filling is used for processing; for the target, each amino acid residue is treated as a node, and when the total number of amino acid residues is insufficient... When this happens, fill the space with zeros. The size of the drug mapping dimension, dimension The specific value can be defined manually;
[0157] make Regarding the structural characteristics of drugs Perform dimensional transformation to obtain the transformed features. , , , This indicates the actual number of atoms in the drug, for the target. , This represents the number of amino acid residues in the target. Since the number of nodes is aligned by padding with zeros for both drugs and targets, the structural features of the drug or target are converted into actual dimensions here, and the information corresponding to the zero-padding is removed.
[0158] Step S42: Transform the features along the height and width directions. Decomposition yields spatial substructures along the width direction. and spatial substructures in the height direction ;in, , ;
[0159] and features Divided into Each spatial substructure will feature Divided into Each spatial substructure contains [number] spatial substructures. One channel;
[0160]
[0161]
[0162] in, Indicates along The first axis assignment obtained A spatial substructure, express correspond The first in The first channel to the first One channel, Indicates along The first axis assignment obtained A spatial substructure, express correspond The first in The first channel to the first One channel; =1,2,…, ;
[0163] Step S43: Extract the spatial feature matrix of each spatial substructure obtained in step S42 by depthwise separable convolution;
[0164]
[0165]
[0166] in, express The spatial feature matrix, express The spatial feature matrix, This represents a depthwise separable convolutional layer, where the kernel size of the depthwise separable convolutional layer is... ;
[0167] The spatial attention matrix is obtained by aggregating and group normalizing (GN) the spatial feature matrices of the spatial substructures.
[0168]
[0169]
[0170] in, Indicates will Add;
[0171] Indicates will Add;
[0172] This indicates that group normalization is performed along the width direction;
[0173] This represents performing group normalization along the height direction;
[0174] Indicates along The spatial attention matrix obtained from the axis;
[0175] Indicates along The spatial attention matrix obtained from the axis;
[0176] Step S44, will , and Multiplying yields the result of multiplication. The result of multiplication The drug's channel features are extracted by sequentially passing through a global average pooling layer and a depthwise separable convolutional layer.
[0177]
[0178] in, This represents a global average pooling operation with a kernel size of . The feature matrix can be changed by applying global pooling operations. Size, This indicates the extracted drug pathway characteristics;
[0179] Step S45: Process the extracted drug channel features using a single-head self-attention mechanism to obtain enhanced drug channel features;
[0180] The specific steps are as follows:
[0181]
[0182]
[0183]
[0184] in, , and These represent queries generated through linear mapping. ,key Sum , yes The dimension, where the superscript T indicates transpose. This represents a global average pooling operation with a kernel size of . , It is an enhanced drug pathway feature;
[0185] Similarly, the enhanced channel features of the target amino acid sequence are obtained using the methods in steps S41 to S45;
[0186] It should be noted that: the process of steps S42 to S43 is called the calculation process of the multi-scale spatial attention mechanism (MSSA), and the process of steps S44 to S45 is called the calculation process of the channel-enhanced attention mechanism (CESA). The discriminative features of the substructure are captured through the synergy of the multi-scale spatial attention mechanism and the channel-enhanced attention mechanism.
[0187] Step S5: Sparsify the enhanced drug channel features using a sparse self-attention mechanism to obtain sparse drug features. The enhanced target amino acid sequence channel features were sparsified using a sparse self-attention mechanism to obtain sparse target amino acid sequence features. ;
[0188] Taking the sparsification of enhanced drug channel features as an example:
[0189]
[0190]
[0191] in, It is the weight of the i-th attention head;
[0192] , and This represents the query, key, and value of D after the i-th head mapping;
[0193] This indicates the percentage of element values retained in each row of D. , This indicates that a binary mask operation is performed based on the index corresponding to the retained element value. TOP- in each row The maximum value is in The corresponding value in the middle remains unchanged, while the other values in each row are... The corresponding value is negative infinity;
[0194] For example, for For any row [0.1, 0.5, 0.2, 0.8], when When the value of is 1 / 2, the number of element values retained in each row is . , This indicates the number of elements in each row. This indicates rounding down. After masking, the line becomes [-inf, 0.5, -inf, 0.8], where inf represents infinity; similarly, for... Each row is processed separately. In this invention, The possible values are 1 / 2, 2 / 3, 3 / 4, and 4 / 5.
[0195] Indicates the total number of heads of attention;
[0196] Indicates the dimension of D;
[0197] This indicates a sparse drug characteristic.
[0198] Similarly, sparse target amino acid sequence characteristics were obtained. Then, the sparse drug signature and the sparse target amino acid sequence signature are cross-fused to obtain the cross-fused drug signature. and the target amino acid sequence characteristics after cross-fusion Specifically:
[0199] Sparse drug features are transformed by linear transformation Mapped to query s and keys By using linear transformation, the sparse target amino acid sequence features are transformed. Mapped to query s and keys ,
[0200]
[0201]
[0202]
[0203]
[0204] in, express Dimensions express The dimension;
[0205] This step, through cross-fusion, can obtain a refined feature representation and reduce the impact of redundant information.
[0206] Step S6: Based on the enhanced drug channel feature D, the enhanced target amino acid sequence channel feature P, and the cross-fused drug feature... and the target amino acid sequence characteristics after cross-fusion The joint representation of the sample pair consisting of drug and target is obtained. ;
[0207]
[0208] in, and They are Obtained through average pooling operation Indicates multiplication. Indicates feature splicing;
[0209] Step S7: Combine the representations Input the MLP, predict the interaction probability of the drug-target sample pair using the MLP, and calculate the loss function based on the predicted drug-target interaction probability and the label of the drug-target sample pair:
[0210] ;
[0211] in, It is the predicted interaction probability;
[0212] These are the true labels of the sample pairs; positive sample pairs are labeled with 1, and negative sample pairs are labeled with 0.
[0213] It represents the total number of drug-target pairs.
[0214] Adjust the parameters of each layer in reverse according to the loss function until the loss function converges, and then continue to execute step S8.
[0215] Step S8: Process the drug-target pair to be predicted using steps S2 to S6 to obtain the joint representation of the drug-target pair to be predicted. Input the joint representation of the drug-target pair to be predicted into the MLP to obtain the interaction probability of the drug-target pair to be predicted.
[0216] If the probability of drug-target interaction is greater than the threshold, then the drug-target pair to be predicted has an interaction; otherwise, the drug-target pair to be predicted does not have an interaction.
[0217] Experimental Section
[0218] As shown in Tables 1, 2 and 3, the performance of the method of this invention (DSSG-DTI) is compared with other baselines on the BindingDB dataset, BioSNAP dataset and Human dataset, respectively.
[0219] Table 1 Performance Comparison of BindingDB Dataset
[0220]
[0221] Table 2 Performance comparison of the BioSNAP dataset
[0222]
[0223] Table 3 Performance Comparison of the Human Dataset
[0224]
[0225] As can be seen, the method of this invention achieved the best prediction performance on all three datasets.
[0226] To visually highlight the advantages of the method of this invention, we have also plotted the ROC and PRC curves of the proposed method and other baselines on the BioSNAP dataset. For example... Figure 2a The image shows an ROC curve, with the horizontal axis representing the "False Positive Rate" (from 0 to 1) and the vertical axis representing the "True Positive Rate" (also from 0 to 1). Figure 2b The image shows a PRC curve, with the horizontal axis representing "Recall" (range 0 to 1) and the vertical axis representing "Precision" (range 0 to 1). Figure 2a and Figure 2b The performance of multiple models (GraphDTA, Moltrans, MGNDTI, FMCA-DTI, MCANet, DrugBAN, DSSG-DTI) under different performance metrics (AUROC and AUPR) was compared. As can be seen from the figures, different models have their own strengths and weaknesses under different metrics, but the method of this invention shows superior performance in both AUC and AUPR metrics.
[0227] We used the Area Under the Receiver Operating Characteristic (AUROC) and the Area Under the Precision-Recall (AUPR) curve as evaluation metrics. AUROC, ranging from 0.5 to 1, quantifies the difference between positive and negative classes to assess the performance of the binary classifier. The ROC curve integrals for the True Positive Rate (TPR) and False Positive Rate (FPR) are calculated as follows:
[0228]
[0229] AUPR (Area Under Precision and Recall) effectively evaluates a model's performance on imbalanced datasets, particularly those with a limited number of positive samples. Increased AUPR scores are directly correlated with enhanced discrimination capabilities.
[0230] To avoid overestimating performance due to random partitioning and to evaluate the model in more realistic scenarios, we conducted additional tests on the BioSNAP dataset, including three scenarios: warm start, drug cold start, and target cold start. The dataset partitioning was performed using MolTrans. Experimental results are shown below. Figure 3As shown, we observed that in the drug cold start setting, the method of this invention improved AUROC and AUPR by 1.5% and 1.18% respectively compared to the baseline DrugBAN. In the more challenging target cold start scenario, the method of this invention outperformed the best benchmark MCANet by 4.63% and 3.89% in AUROC and AUPR, respectively. Notably, even under conventional warm start conditions, the method of this invention consistently improved AUROC and AUPR by 0.81% and 0.76%, respectively. Furthermore, most state-of-the-art baselines performed worse in the target cold start scenario than in the drug cold start scenario, indicating that their learning of target characterization has not yet reached the level of fundamental features. In contrast, the method of this invention exhibited stable performance in all test scenarios, and its continued advantage in the cold start setting further validates its accuracy in predicting unknown DTIs.
[0231] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A drug-target interaction prediction method that fuses two-stream collaborative attention with sparse features, characterized in that, The method specifically includes the following steps: Step S1: Obtain drug and target data sample pairs, including positive sample pairs and negative sample pairs; Step S2: Process the drug data and target data in each data sample pair separately to obtain the processed drug and target data sample pairs; Step S3: Extract the drug structure features and target amino acid sequence features of each processed drug and target data sample pair; Step S4: Extract the spatial features of the drug based on the drug structure features of the data sample pair, then extract the channel features of the drug based on the spatial features of the drug, and process the extracted drug channel features to obtain enhanced drug channel features. Based on the structural features of the target amino acid sequence of the data sample pair, the spatial features of the target amino acid sequence are extracted. Then, based on the spatial features of the target amino acid sequence, the channel features of the target amino acid sequence are extracted. The extracted channel features of the target amino acid sequence are then processed to obtain the enhanced channel features of the target amino acid sequence. The process involves extracting spatial features of the drug based on the drug structural features of the data sample pairs, then extracting channel features of the drug based on the spatial features, and finally processing the extracted drug channel features to obtain enhanced drug channel features. The specific process is as follows: Step S41: Perform dimensional transformation on the drug structural features: The structural characteristics of the drug are denoted as , ,in, Represent real numbers, Indicates batch size, Indicates the number of nodes. Let be the size of the drug mapping dimension; Regarding the structural characteristics of drugs Perform dimensional transformation to obtain the transformed features. , , , Indicates the actual number of atoms in the drug; Step S42: Transform the features along the height and width directions. Decomposition yields spatial substructures along the width direction. and spatial substructures in the height direction ; and features Divided into Each spatial substructure will feature Divided into Each spatial substructure includes [number] spatial substructures. One channel; in, Indicates along The first axis assignment obtained Each spatial substructure feature Indicates along The first axis assignment obtained Each spatial substructure feature; =1,2,…, ; Step S43: Extract the spatial feature matrix of each spatial substructure obtained in step S42 by depthwise separable convolution; in, express The spatial feature matrix, express The spatial feature matrix, This represents a depthwise separable convolutional layer; The spatial attention matrix is obtained by aggregating and grouping the spatial feature matrices of the spatial substructures: in, Indicates will Add; Indicates will Add; This indicates that group normalization is performed along the width direction; This represents performing group normalization along the height direction; Indicates along The spatial attention matrix obtained from the axis; Indicates along The spatial attention matrix obtained from the axis; Step S44, will , and Multiplying yields the result of multiplication. The result of multiplication The drug's channel features are extracted by sequentially passing through a global average pooling layer and a depthwise separable convolutional layer. in, This indicates a global average pooling operation. This indicates the extracted drug pathway characteristics; Step S45: Process the extracted drug channel features using a single-head self-attention mechanism to obtain enhanced drug channel features. ; Step S5: Sparsify the enhanced drug channel features using a sparse self-attention mechanism to obtain sparse drug features. The enhanced target amino acid sequence channel features were sparsified using a sparse self-attention mechanism to obtain sparse target amino acid sequence features. ; Then, the sparse drug signature and the sparse target amino acid sequence signature are cross-fused to obtain the cross-fused drug signature. and the target amino acid sequence characteristics after cross-fusion ; Step S6: Based on the enhanced drug channel feature D, the enhanced target amino acid sequence channel feature P, and the cross-fused drug feature... and the target amino acid sequence characteristics after cross-fusion The joint representation of the sample pair consisting of drug and target is obtained. ; Step S7: Combine the representations Input the MLP, predict the interaction probability of the sample pair composed of drug and target through the MLP, calculate the loss function based on the predicted drug-target interaction probability and the label of the sample pair composed of drug and target, and adjust the parameters of each layer in reverse according to the loss function until the loss function converges and continue to execute step S8. Step S8: Process the drug-target pair to be predicted using steps S2 to S6 to obtain the joint representation of the drug-target pair to be predicted. Input the joint representation of the drug-target pair to be predicted into the MLP to obtain the interaction probability of the drug-target pair to be predicted. If the probability of drug-target interaction is greater than the threshold, then the drug-target pair to be predicted has an interaction; otherwise, the drug-target pair to be predicted does not have an interaction.
2. The drug-target interaction prediction method based on the fusion of dual-stream collaborative attention and sparse features according to claim 1, characterized in that, The drug data within the sample pair is drug SMILES string data, and the target data within the sample pair is target amino acid sequence data.
3. The drug-target interaction prediction method based on the fusion of dual-stream collaborative attention and sparse features according to claim 2, characterized in that, The processing of drug data within each data sample pair specifically involves: The drug SMILES string data was converted into an atomic interaction diagram using the open-source dgl tool.
4. The drug-target interaction prediction method based on the fusion of dual-stream collaborative attention and sparse features according to claim 3, characterized in that, The processing of the target data within each data sample pair specifically involves: Step S21: Divide the target amino acid sequence into several fragments of equal length. If there are identical fragments among the divided fragments, only one identical fragment is retained. If there are no identical fragments among the divided fragments, all the divided fragments are retained. Step S22: Convert each retained segment into a digital vector, and use the obtained digital vectors to form an embedding matrix.
5. The drug-target interaction prediction method based on the fusion of dual-stream collaborative attention and sparse features according to claim 4, characterized in that, The specific process of step S3 is as follows: Step S31: For drugs: Extract drug structural features from the atomic interaction diagram using the GCN model; specifically: in: This represents an adjacency matrix with self-connections; In the GCN model, the first... Layer weight matrix; In the GCN model, the first... Layer bias vector; It is the first in the GCN model The output of the layer, , The output of the atomic interaction diagram of the drug after the embedding operation; It is the first in the GCN model Layer output; Represents the nonlinear activation function ReLU; The output of the last layer in the GCN model is used as the drug structure feature; Step S32: For the target: Use the MKCNN model to process the embedding matrix and extract the features of the target amino acid residue sequence; specifically: in: Indicates the batch normalization layer; The MKCNN model represents the first... One convolutional layer; Indicates the first The kernel size of each convolutional layer; The MKCNN model represents the first... The weight matrix of each convolutional layer; The MKCNN model represents the first... The bias vectors of each convolutional layer; In the MKCNN model, the first... The output of each convolutional layer; In the MKCNN model, the first... The output of each convolutional layer; The output of the last ReLU activation function layer is used as a feature of the target amino acid residue sequence.
6. The drug-target interaction prediction method based on the fusion of dual-stream collaborative attention and sparse features according to claim 5, characterized in that, The specific process of step S45 is as follows: in, , and These represent queries generated through linear mapping. ,key Sum , yes The dimension, where the superscript T indicates transpose. This indicates a global average pooling operation. It is an enhanced drug pathway feature.
7. The drug-target interaction prediction method based on the fusion of dual-stream collaborative attention and sparse features according to claim 6, characterized in that, The enhanced drug channel features are sparsified using a sparse self-attention mechanism to obtain sparse drug features. The enhanced target amino acid sequence channel features were sparsified using a sparse self-attention mechanism to obtain sparse target amino acid sequence features. Specifically: Taking the sparsification of enhanced drug channel features as an example: in, It is the weight of the i-th attention head; , and This represents the query, key, and value of D after the i-th head mapping; This indicates the percentage of element values retained in each row of D. , This indicates that a binary mask operation is performed based on the index corresponding to the retained element value; Indicates the total number of heads of attention; Indicates the dimension of D; Indicates a sparse drug characteristic; Similarly, sparse target amino acid sequence characteristics were obtained. .
8. The drug-target interaction prediction method based on the fusion of dual-stream collaborative attention and sparse features according to claim 7, characterized in that, The sparse drug signature and sparse target amino acid sequence signature are cross-fused to obtain the cross-fused drug signature. and the target amino acid sequence characteristics after cross-fusion Specifically: Sparse drug features are transformed by linear transformation Mapped to query s and keys By using linear transformation, the sparse target amino acid sequence features are transformed. Mapped to query s and keys , in, express Dimensions express Dimensions.
9. The drug-target interaction prediction method based on the fusion of dual-stream collaborative attention and sparse features according to claim 8, characterized in that, The specific process of step S6 is as follows: in, and They are Obtained through average pooling operation Indicates multiplication. This indicates feature splicing.