Biomolecule interaction prediction method based on multi-modal attention fusion

Abstract

The invention discloses a biomolecular interaction prediction method based on multi-modal attention fusion, and belongs to the technical field of artificial intelligence drug discovery. The method comprises the following steps: acquiring multi-modal characteristics of drugs, targets, diseases and genes: sequence structure characteristics, 3D structure characteristics, similarity network characteristics and biological relation network embedding characteristics; constructing a feature fusion prediction model, and performing training; inputting the multi-modal features of the two biological entities into the trained feature fusion prediction model, and outputting the probability of interaction of the two biological entities; the feature fusion prediction model comprises a Transform encoder and an MLP (Markup Language Protocol) network; the multi-modal features are input into a feature fusion prediction model for stacking and then are input into a Transform encoder, the features are processed by using a multi-head self-attention mechanism, and an output result is flattened and subjected to dimension reduction processing to obtain embedded vector representation; and finally, performing element corresponding multiplication on the embedded vectors of the two biological entities, inputting the embedded vectors into an MLP network, and outputting an interaction probability between the two biological entities.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence drug discovery technology, specifically relating to a method for predicting biomolecular interactions based on multimodal attention fusion. Background Technology

[0002] Drug development and repositioning are core aspects of the biopharmaceutical field. Accurate identification of drug-target interactions (DTI), drug-disease interactions (DrDiI), and drug-drug interactions (DDI) is crucial for candidate drug screening, target validation, indication expansion, and drug safety assessment. Traditional biomolecular interaction studies primarily rely on experimental methods, such as in vitro experiments and high-throughput screening. While these methods can yield relatively reliable experimental data, they suffer from long experimental cycles, high costs, and low efficiency, making them unsuitable for the large-scale screening needs of new drug development.

[0003] With the development of computational methods, researchers have begun to utilize machine learning and deep learning methods to predict biomolecular interactions (including drug-target interactions, drug-drug interactions, and drug-disease interactions). These methods can infer potential interaction relationships based on existing biological association data, thereby reducing experimental screening costs. However, existing methods still have the following shortcomings: First, they lack unified modeling capabilities for multiple tasks, severing the intrinsic connections between drugs, targets, and diseases; second, multimodal feature fusion is insufficient, often neglecting the complementarity of information such as disease semantics (e.g., MeSH) and the topological structure of heterogeneous biological networks; third, they lack robustness to sparse data and the "cold start" problem, making it difficult to generate high-quality node embeddings when processing sparse graph structures. Therefore, this invention proposes a biomolecular interaction prediction method based on multimodal attention fusion to address the above problems. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a method for predicting biomolecular interactions based on multimodal attention fusion, thereby solving the problems in existing technologies.

[0005] The objective of this invention can be achieved through the following technical solutions: A method for predicting biomolecular interactions based on multimodal attention fusion includes the following steps: Acquire multimodal features of biological entities: drugs, targets, diseases, or genes; the multimodal features include: sequence structure features, 3D structure features, similarity network features, and biological relationship network embedding features; Construct and train a feature fusion prediction model based on a multi-head self-attention mechanism; The multimodal features of two biological entities are input into a trained feature fusion prediction model, which outputs the probability of interaction between the two biological entities. The interaction includes: drug-target interaction, drug-disease interaction, drug-drug interaction, protein-gene interaction, and gene-disease relationship. The feature fusion prediction model includes a Transformer encoder and an MLP network. The multimodal features are input into the feature fusion prediction model and stacked to generate a three-dimensional feature array. The array is then input into the Transformer encoder, where a multi-head self-attention mechanism is used to process the features, learn interaction information, and add it to the original input through residual connections. The output is flattened and dimensionality reduced to obtain the embedding vector representation of the drug, target protein, or disease that integrates multimodal information. Finally, the embedding vector representations of the two biological entities are multiplied element-wise and input into the MLP network. After nonlinear transformation, the interaction probability between the two biological entities is output.

[0006] Furthermore, the sequence structure features include: molecular sequence features of the drug, sequence features of the target, 3D structural features of the drug, 3D structural features of the target, and semantic features of the disease; The molecular sequence features of drugs and the sequence features of targets are extracted through a pre-trained language model; the semantic features of diseases are constructed based on a medical subject term database.

[0007] Furthermore, the extraction process of the similarity network features is as follows: calculate the similarity between drugs, target proteins, diseases, and genes to construct a similarity matrix, construct a similarity network graph based on the similarity matrix, and then use a graph neural network to learn the representation of the nodes to obtain the similarity network features.

[0008] Furthermore, the construction process of the similarity network graph is as follows: based on the similarity matrix, select the top N nearest neighbor nodes for each entity node and establish connections to construct a directed similarity graph G; input the similarity graph G into the graph neural network for feature learning; wherein, for any node, its updated feature vector is calculated by aggregating the node information in the neighborhood.

[0009] Furthermore, the process of obtaining the biological relationship network embedding features is as follows: integrating multiple biological databases to construct a heterogeneous network containing drugs, proteins, and disease entities, and using a graph embedding algorithm to learn the topological structure representation of each node in the heterogeneous network to obtain the biological relationship network embedding features; the graph embedding algorithm includes a graph embedding algorithm based on random walk or a graph autoencoder algorithm based on contrastive learning; for sparse heterogeneous networks, a graph autoencoder algorithm based on contrastive learning is used to generate an enhanced graph through an edge masking strategy for feature learning.

[0010] A biomolecular interaction prediction system based on multimodal attention fusion includes: Feature extraction module: Acquires multimodal features of biological entities: drugs, targets, diseases, or genes; the multimodal features include: sequence structure features, 3D structure features, similarity network features, and biological relationship network embedding features; Model training module: Construct and train a feature fusion prediction model based on a multi-head self-attention mechanism; Prediction module: Inputs the multimodal features of two biological entities into the trained feature fusion prediction model and outputs the probability of interaction between the two biological entities; the interaction includes: drug-target interaction, drug-disease interaction, drug-drug interaction, protein-gene interaction, and gene-disease relationship; The feature fusion prediction model includes a Transformer encoder and an MLP network. The multimodal features are input into the feature fusion prediction model and stacked to generate a three-dimensional feature array. The array is then input into the Transformer encoder, where a multi-head self-attention mechanism is used to process the features, learn interaction information, and add it to the original input through residual connections. The output is flattened and dimensionality reduced to obtain the embedding vector representation of the drug, target protein, or disease that integrates multimodal information. Finally, the embedding vector representations of the two biological entities are multiplied element-wise and input into the MLP network. After nonlinear transformation, the interaction probability between the two biological entities is output.

[0011] The above-mentioned biomolecular interaction prediction method based on multimodal attention fusion is applied to the screening of targeted drug molecules.

[0012] An electronic device includes: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction that causes the processor to perform operations corresponding to the above-described biomolecular interaction prediction method based on multimodal attention fusion.

[0013] A computer program product includes computer instructions that instruct a computing device to perform operations corresponding to the above-described method for predicting biomolecular interactions based on multimodal attention fusion.

[0014] The beneficial effects of this invention are: 1. By integrating three heterogeneous data sources—sequence structure information, similarity network topology, and biological relationship network embedding—in parallel, and utilizing a Transformer encoder based on an attention mechanism for deep fusion, this method effectively overcomes the limitations of traditional single-modal methods, which suffer from limited information dimensions and an inability to comprehensively depict complex relationships between entities. Through multi-perspective and multi-level feature extraction, a comprehensive feature module for drugs, targets, diseases, and genes is constructed. The attention fusion mechanism of this invention can dynamically weight features from different sources, capturing deep interaction relationships between modalities rather than simple feature splicing. This significantly improves the accuracy and generalization ability of predicting drug-target, drug-disease, and drug-drug interactions, especially demonstrating strong robustness in handling the "cold start" problem of new drugs or targets. Furthermore, this method not only exhibits excellent predictive performance, but its framework also possesses high versatility and scalability, providing a more accurate and reliable computational tool for key application scenarios such as new drug discovery and drug repositioning.

[0015] 2. By incorporating drug molecule sequences, 3D structural features of the drug-target pair, target sequences, and disease semantic features into sequence structure features, and employing large-scale pre-trained language models (such as ChemBERTa-2 and ESM-2) for deep representation, we can fully mine the contextual information of molecular structure and protein function, overcoming the limitations of traditional molecular fingerprinting or homology alignment methods in feature expression. Simultaneously, the disease semantic features constructed based on the Medical Subject Headings (MeSH) database effectively integrate disease hierarchy relationships and biological semantic information, enabling the drug-target-disease triad to be modeled in a unified feature space, thereby significantly enhancing the prediction accuracy and generalization of complex interactions.

[0016] 3. By incorporating similarity network information of drugs, targets, diseases, and genes during feature construction, and utilizing molecular fingerprinting, sequence alignment, and phenotypic methods to calculate the similarity matrix between entities, a similarity network graph is further constructed. Feature learning is then performed using a graph neural network. This approach not only captures the potential topological relationships and implicit patterns between similar entities but also effectively compensates for the shortcomings of single sequence or semantic features in structural representation. Unlike traditional methods that rely solely on local descriptions, similarity network features can reveal the intrinsic connections between entities from a global perspective, thereby enhancing the richness and discriminative power of node representations and improving the model's stability and adaptability in data-sparse or cold-start scenarios.

[0017] 4. By selecting the top N nearest neighbors for each entity node based on the similarity matrix and constructing a directed similarity graph, not only is redundancy and noise caused by fully connected graphs reduced, but the most representative local similarity relationships are also highlighted, thereby improving graph sparsity and learning efficiency. Under the neighborhood aggregation mechanism, node features can dynamically integrate local and global information, enabling the full expression of similarity patterns between drugs, proteins, diseases, and genes, further enhancing the model's ability to model complex biomolecular relationships and its robustness.

[0018] 5. By integrating multiple biological databases to construct a heterogeneous network containing entities such as drugs, proteins, diseases, and genes, and utilizing graph embedding algorithms to learn the topological features of nodes, this invention can characterize the potential connections between biomolecules at a more macroscopic level. Specifically, for highly sparse biological networks, this invention introduces a contrastive learning-based graph autoencoder (such as IrGAE) and uses an edge masking strategy to generate an enhanced graph, effectively alleviating the representation learning difficulties caused by sparse data. This design can not only capture cross-type relationships (such as drug-protein, protein-disease, drug-disease, gene-disease, etc.), but also reveal the potential functional coupling between nodes of the same type. Compared with single-modal input, this design can more comprehensively utilize information from existing biological knowledge bases, effectively alleviate the data sparsity problem, improve the model's reasoning and interpretation capabilities for unknown interaction relationships, thereby enhancing the accuracy and biological rationality of prediction results.

[0019] 6. By employing the cross-entropy loss function as the training objective, this invention effectively distinguishes between positive and negative samples in binary classification tasks, ensuring that the difference between the predicted probability distribution and the true label distribution is minimized. This loss function exhibits convergence stability and good generalization, maintaining robustness even in scenarios with imbalanced data and high-dimensional inputs. This results in higher reliability and accuracy for predicting interactions such as drug-target, drug-disease, drug-drug, and gene-disease interactions. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the biomolecular interaction prediction framework of the present invention.

[0022] Figure 2 This is a schematic diagram showing the molecular structure, molecular docking simulation, and bioactivity verification results of potential active compounds screened for MR targets in this embodiment of the invention.

[0023] Figure 3 This is a schematic diagram illustrating the prediction results of potential associated diseases and indications for the MR inhibitor fenelazolone and the NET inhibitor maprotiline in an embodiment of the present invention.

[0024] Figure 4 This is a schematic diagram of the workflow and functional interface of the drug discovery visualization and prediction platform provided in this embodiment of the invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1 like Figure 1 As shown, a method for predicting biomolecular interactions based on multimodal attention fusion includes the following steps: S1, acquire multimodal features of drugs, target proteins, or diseases; the multimodal features include: sequence structure features, similarity network features, and biological relationship network embedding features; 1) Sequence structure characteristics Sequence structure features include: molecular sequence features of drugs, protein sequence features of target proteins, and semantic features of diseases; The extraction method involves using pre-trained language models (such as ChemBERTa-2 and ESM-2) to extract the molecular sequence features of the drug and the protein sequence features of the target protein, respectively, and constructing the semantic features of the disease based on the Medical Subject Headings (MeSH) database; specifically: In this embodiment, for drugs, their SMILES string representation is obtained and input into a pre-trained chemical language model ChemBERTa-2 to extract high-dimensional feature vectors that characterize their chemical structure and properties. For target proteins, their FASTA amino acid sequences are obtained and input into a pre-trained protein language model ESM-2 to extract high-dimensional feature vectors that characterize their structure and function. For diseases, a directed acyclic graph (DAG) is constructed based on the MeSH database, and a graph embedding algorithm is used to learn the semantic vectors of MeSH titles as feature representations of the disease.

[0027] 2) Similarity network features; A similarity matrix is ​​constructed by calculating the similarities between drugs, target proteins, diseases, and genes. Based on this matrix, a similarity network graph is built. Then, a graph neural network is used to learn the representations of the nodes to obtain the similarity network features. The similarity calculation methods are as follows: (1) Similarity between drugs A weighted fusion method was employed, combining Tanimoto coefficients based on molecular fingerprinting with Gaussian Interactive Profile (GIP) kernel similarity. First, the drug structure similarity matrix was calculated based on the molecular fingerprint. And calculate the GIP kernel similarity matrix based on drug-target association. The final comprehensive drug similarity matrix The weighted average of the two is calculated using the following formula: in, , For the quantity of drugs, Indicates drug and Gaussian Interactive Profile (GIP) kernel similarity between them , It is a drug entity. Indicates drug Interactive contour vectors, To normalize the adjustment parameter, It is usually set to 1. The Tanimoto similarity matrix based on molecular fingerprints is calculated using the RDKit tool; This is the final fused drug similarity matrix.

[0028] (2) Similarity between target proteins: A weighted fusion of BLAST-based sequence alignment similarity and Gaussian Interaction Profile (GIP) kernel similarity was employed. First, a protein sequence similarity matrix was constructed based on the BLAST alignment scores. And calculate the GIP kernel similarity matrix based on drug-target association. The final protein comprehensive similarity matrix The weighted average of the two is calculated using the following formula: in, N is the amount of protein. Indicates target protein and Gaussian Interactive Profile (GIP) kernel similarity between them , The target protein entity, Protein Interactive contour vectors, To normalize the adjustment parameter, It is usually set to 1. This is a normalized sequence similarity matrix constructed based on BLAST sequence alignment scores; This is the protein similarity matrix after final fusion.

[0029] (3) Similarities between diseases: Semantic analysis of medical texts in the OMIM database was performed using the MimMiner method to obtain a disease similarity matrix: in, For the number of diseases, Represents the disease similarity matrix. Represents the real number field. They each correspond to the same set of diseases. This represents the similarity score between diseases in the matrix.

[0030] The process of constructing a similarity network graph is as follows: Based on the similarity matrix, the top N nearest neighbors are selected for each entity node and connections are established, thereby constructing a directed similarity graph G. Graph G is then input into an improved Graph Transformer model (graph neural network) for feature learning, where, for any node... Its updated feature vector By aggregating neighborhoods The node information within is calculated, and the calculation process is as follows: First, for the nodes Current features and its neighboring nodes Features Through the weight matrix respectively Perform a linear transformation to obtain the query vector. Key vector Sum value vector : Then, the attention weights are calculated: Finally, the neighborhood information is aggregated and the node features are updated: in, Let be the dimension of the key vector. This indicates that bullish traders should pay attention to the splicing operation. This is the projection weight matrix.

[0031] The improved Graph Transformer model takes a directed similarity graph constructed based on a similarity matrix as input and only selects the top N nearest neighbor nodes for each entity node to establish connections. During the node feature update process, a multi-head attention mechanism that combines node features and edge similarity strength is used to weight and aggregate neighborhood information. Graph topology encoding, residual connections, and normalization mechanisms are also introduced to distinguish it from the conventional Graph Transformer model, which relies solely on static adjacency relationships and node features. This allows for feature learning that better conforms to the similarity patterns of biological entities.

[0032] In this embodiment, the process of obtaining similarity network features is as follows: First, a drug similarity matrix and a protein similarity matrix are constructed. The drug similarity matrix is ​​obtained by combining the Tanimoto coefficients (based on Morgan fingerprints) calculated using RDKit and Gaussian interaction spectrum (GIP) kernel similarity. The protein similarity matrix is ​​obtained by combining BLAST sequence alignment scores and GIP kernel similarity. Then, based on the similarity matrices, a nearest neighbor graph is constructed for each entity (drug or protein). These graphs are then fed into a graph attention network (Graph Transformer) model, where the embedding representation of each node in its local network context, i.e., the similarity network features, is learned through a message passing mechanism.

[0033] 3) Embedding features of biological relationship networks; Validated drug-target, drug-disease, protein-protein, and drug-drug association information is extracted from multiple biological databases (such as DrugBank, OMIM, etc.) and integrated to construct a large-scale heterogeneous information network graph. In a preferred embodiment (for cases with sufficient data), the DeepWalk algorithm is applied to this heterogeneous information network graph to generate node sequences through numerous random walks. Then, a Skip-gram model is used to learn the low-dimensional vector representation of each node (drug or protein), i.e., the biological relationship network embedding features. In another preferred embodiment (for sparse data or cold-start scenarios, such as B-dataset), a graph autoencoder based on contrastive learning (such as IrGAE) is used instead of the DeepWalk algorithm. The specific steps are: edge masking is performed on the heterogeneous information network graph to generate an enhanced graph; the sequence / semantic features of drugs, proteins, and diseases are used as node features, and the associations between nodes are used as edges; the enhanced graph is input into a graph neural network (GNN) for encoding, and the model is optimized through a contrastive loss function to learn robust node embedding vectors.

[0034] S2, Construct and train a feature fusion prediction model based on a multi-head self-attention mechanism; Feature fusion prediction models based on multi-head self-attention mechanisms include: Transformer encoders and multilayer perceptron (MLP) networks; The multimodal features acquired by S1 are stacked into the feature fusion prediction model to generate a three-dimensional feature array. The three-dimensional feature array is then input into the Transformer encoder, where the features are processed using a multi-head self-attention mechanism to learn interaction information. The array is then added to the original input through residual connections. The output is flattened and dimensionality reduced to obtain the embedding vector representations of drugs, target proteins, or diseases that fuse multimodal information. Finally, the embedding vector representations of the two biological entities (one being a drug and the other a target protein or disease) are multiplied element-wise and then input into the MLP network. After nonlinear transformation, a probability value between 0 and 1 is output, which represents the interaction probability between the two biological entities. The multimodal features are stacked as follows: the molecular sequence features of the drug, the sequence features of the target protein, the semantic features of the disease, the similarity network features, and the biological relationship network embedding features are aligned according to the modal dimension and stacked sequentially in the modal dimension to form a three-dimensional feature array of "modality × feature dimension × number of samples"; the feature vectors of different modalities are input into the fusion prediction model as different channels, so that the model can achieve deep feature interaction and dynamic weighting among multimodalities, rather than simple vector concatenation; Multimodal features of drugs, proteins, and diseases are input into the fusion model. Known interaction data are used as supervision labels. The model parameters are iteratively updated through forward and backward propagation, and the cross-entropy loss function is used as the optimization objective until the model converges and training is completed.

[0035] The loss function used in the feature fusion prediction model is the cross-entropy loss function, expressed as follows: in, This represents the total number of samples involved in the calculation. Indicates the first The sample at the th Real labels in each category The model represents the first The sample belongs to the first The predicted probability of each category.

[0036] S3 takes the multimodal features of two biological entities, inputs them into the trained feature fusion prediction model, and outputs the probability of the interaction between the two biological entities. This method can be used simultaneously for drug-target interaction prediction, drug-disease interaction prediction, and drug-drug interaction prediction, that is, one of the two biological entities is a drug, and the other is a target protein, disease, or other drug.

[0037] In this embodiment, the multimodal features of the drug and target protein are input into the model and processed to obtain an embedding vector representation of the drug and the target protein. These two vectors are then element-wise multiplied, and the result is input into an MLP network. After nonlinear transformation, the MLP network outputs a probability value between 0 and 1, which is the interaction probability of the drug-target pair.

[0038] In other embodiments, the prediction process for drug-disease interaction and drug-drug interaction is similar to that of drug-target interaction described above, except that the two input entity features are replaced with their corresponding types.

[0039] Based on a similar inventive concept, embodiments of the present invention also provide a computer storage medium storing a readable program that, when run by a processor, can execute the above-described method for predicting biomolecular interactions based on multimodal attention fusion.

[0040] Based on a similar inventive concept, this invention provides an electronic device, including: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described biomolecular interaction prediction method based on multimodal attention fusion.

[0041] Based on a similar inventive concept, embodiments of the present invention also provide a computer program product, including computer instructions, which instruct a computing device to perform the operations corresponding to the above-described biomolecular interaction prediction method based on multimodal attention fusion.

[0042] Example 2 To verify the effectiveness of the prediction method of this invention, this embodiment conducted DTI prediction experiments on the publicly available DrugBank, KIBA, and Davis datasets; the datasets were divided into training, validation, and test sets in an 8:1:1 ratio. The model was implemented using the PyTorch framework and trained on an NVIDIA A100 GPU. The hyperparameters were set as follows: learning rate 1e-4, batch size 32, and Adam optimizer used.

[0043] The experimental results are shown in Table 1: Table 1. Comparison of experimental data between the DrugGPS model and other existing methods on the DTI dataset. As can be seen from the results in Table 1, the prediction method of this invention significantly outperforms various existing baseline models, such as TransformerCPI, GIFDTI, MSI-DTI, and the latest RRGDTA and DACMF-DTI, in terms of AUC and AUPR on all three datasets, demonstrating the superiority of this invention in terms of prediction accuracy and robustness.

[0044] Example 3 In this embodiment, to further verify the universality and scalability of the method of the present invention, the present invention also applies the method to the prediction task of drug-disease interaction.

[0045] In this embodiment, the target protein entity is replaced with a disease entity, and correspondingly, multimodal features of the disease are extracted, including features based on MeSH (Medical Subject Headings). The experiment uses three publicly available DDI benchmark datasets for evaluation: B-dataset, C-dataset, and F-dataset. To ensure fairness in the comparison, when comparing with existing state-of-the-art models (such as HINGRL, DRHGCN, and AMDGT), the baseline models all use the optimal parameter configurations reported in their original literature. The experimental results are shown in Table 2. Table 2 Comparison of experimental data of the DrugGPS model and other existing methods on the DDI dataset. As shown in Table 2, the proposed method significantly outperforms other baseline methods in all six evaluation metrics across all three DDI benchmark datasets. Particularly noteworthy are the significant advantages demonstrated by the proposed method on the C-dataset and F-dataset, particularly achieving leading AUC and AUPR scores of 0.9703 and 0.9577 respectively, reflecting the model's superior classification ability when handling imbalanced datasets. For the relatively sparse B-dataset, this embodiment further employs a contrastive learning-based graph autoencoder (IrGAE) to replace the random walk algorithm for network embedding learning. Experimental results show that after introducing this improved strategy, the model's AUC on the B-dataset improves by 1.4%, and its accuracy improves by 1.9%, significantly outperforming the baseline model AMDGT. This verifies that the proposed optimization strategy for sparse data (i.e., generating enhanced graphs using edge masks) effectively overcomes the feature learning difficulties caused by data sparsity, greatly enhancing the model's robustness. The superior performance of this invention can be attributed to the following factors: First, it employs a multi-source information fusion strategy, which can capture the intrinsic information of drugs and diseases from multiple dimensions; second, through a multi-head attention mechanism, the model can effectively focus on key features, enhancing its ability to model complex drug-disease associations; and third, it introduces potential associations of homogeneous entities (such as drug-drug) into the constructed heterogeneous network, further ensuring the model's generalization ability.

[0046] Example 4 In this embodiment, to verify the applicability of the model to a wider range of biological association tasks, the method of the present invention is further applied to the prediction task of drug-drug interaction (DDI). Experimental data comes from the DrugBank database, containing 191,808 DDI triples, involving 1,706 drugs and 86 interaction types. In this embodiment, the input is constructed using the multimodal features of the drugs (sequence, similarity, and heterogeneous network embeddings). Unlike DTI prediction, DDI prediction is a multi-class classification task. Therefore, in the model's output layer (MLP), the output dimension is adjusted to the number of corresponding interaction types (86 classes), and the Softmax function is used to calculate the probability of belonging to each interaction type. Experimental results show that the model of the present invention achieves excellent performance on the DDI prediction task, with an AUC of 0.963, an AUPR of 0.921, and an accuracy of 0.922. This result demonstrates that the multimodal attention fusion framework proposed in this invention is not only applicable to the prediction of binary relationships (such as drug-target), but also effectively handles multivariate prediction tasks with complex relationship types, proving the model's great application potential in the field of drug interaction safety assessment.

[0047] Example 5 To fully demonstrate the predictive performance and generalization ability of the method in this invention when handling large-scale, unseen data, rigorous external validation experiments were conducted on the BIOSNAP standard dataset. System experiments were performed under three test scenarios constructed from the BIOSNAP standard dataset: full-data (regular random partitioning), unseen-drug, and unseen-protein. The unseen-drug and unseen-protein scenarios were used to simulate the model's predictive ability for drugs or proteins not encountered during the training phase, respectively, which presents a more realistic challenge. The experimental results are shown in Table 3, which compares the performance of the DrugGPS and HyperAttention-DTI models under the three partitioning modes of the BIOSNAP dataset. Table 3. Comparison of experimental data between the DrugGPS model and the HyperAttention-DTI model; Table 3 shows that DrugGPS consistently outperforms the comparative model HyperAttention-DTI on most evaluation metrics. Particularly in the unseen-protein scenario, the model exhibits superior cross-protein generalization performance, with an AUC of 0.929 ± 0.001 and an AUPR of 0.935 ± 0.001, significantly higher than the baseline model. Furthermore, DrugGPS also performs stably in the unseen-drug segmentation, with an AUC of 0.926 ± 0.002, indicating its excellent inference ability for novel drugs. In summary, DrugGPS maintains leading and robust performance across various BIOSNAP test settings, demonstrating the significant application value of the model in handling imbalanced samples and predicting unseen entities in real-world scenarios.

[0048] Example 6 To verify the necessity and contribution of each core component in the method of this invention, this embodiment designed and conducted a series of ablation experiments. The experiments were performed on the KIBA and DrugBank datasets, and the impact on overall performance was evaluated by systematically removing or replacing key modules of the model. Specifically, four variant models were set up: Variant 1: Replace the Transformer encoder with a simple feature concatenation method.

[0049] Variant 2: Remove the sequence structure feature extraction module.

[0050] Variant 3: Remove the similarity network feature extraction module.

[0051] Variant 4: Remove the biological relationship network feature extraction module (DeepWalk).

[0052] The experimental results are shown in Table 4, which presents a performance comparison of DrugGPS and four variants in ablation experiments.

[0053] Table 4 Comparison of ablation experimental data between the DrugGPS model and four variants As shown in Table 4, compared with the complete model of this invention, the performance of all four variant models decreased to varying degrees. Variant 1 showed the most significant performance decline, indicating that the deep fusion strategy based on the attention mechanism is far superior to simple feature concatenation and is key to ensuring high model performance. Variant 2, after removing sequence structure features, experienced a 4.2% decrease in accuracy (using the KIBA dataset as an example), revealing the crucial role of sequence features based on pre-trained language models in capturing the intrinsic biological attributes of entities. The performance declines of Variant 3 and Variant 4 also confirm the indispensable role of similarity networks and biological relationship networks in providing topological and relational information between entities.

[0054] Furthermore, to address the issue of data sparsity, this embodiment specifically conducted comparative experiments on a highly sparse B-dataset. The experiments involved replacing the DeepWalk algorithm in the "Biological Relationship Network Feature Extraction Module" with an improved model based on "Contrastive Learning-Based Graph Autoencoder (IrGAE / MaskGAE)". Experimental results show that after introducing IrGAE for edge masking enhancement, the model's AUC improved by 1.4% and accuracy by 1.9% on the B-dataset. This further demonstrates that when processing sparse biological networks, simple random walks (DeepWalk) may not be sufficient to capture structural information, while the contrastive learning-based graph enhancement strategy proposed in this invention can effectively solve this problem and improve the model's robustness in cold-start or sparse data scenarios.

[0055] In summary, the results of the ablation experiments strongly demonstrate that each module in the method of this invention (sequence structure features, similarity network features, biological relationship network features, and attention fusion mechanism) is an important and complementary component of the complete system. These modules work together to comprehensively characterize the feature attributes of biological entities from multiple perspectives, and through effective interaction and synergy, they jointly improve the predictive performance and generalization ability of the overall model.

[0056] Example 7 In this embodiment, a biomolecular interaction prediction system based on multimodal attention fusion is proposed, including: Feature extraction module: Acquires multimodal features of biological entities: drugs, target proteins, or diseases; the multimodal features include: sequence structure features, 3D structural feature similarity network features, and biological relationship network embedding features; Model training module: Construct and train a feature fusion prediction model based on a multi-head self-attention mechanism; Prediction module: Input the multimodal features of two biological entities into the trained feature fusion prediction model, and output the probability of the interaction between the two biological entities; The two biological entities include: drug and target protein, drug and disease, and drug and drug; the visualization interaction module is used to provide a graphical user interface, allowing users to input biological entity data, display the predicted probability results, and visualize the association between drugs, proteins and diseases in the heterogeneous network.

[0057] The feature fusion prediction model includes a Transformer encoder and an MLP network. The multimodal features are input into the feature fusion prediction model and stacked to generate a three-dimensional feature array. The array is then input into the Transformer encoder, where a multi-head self-attention mechanism is used to process the features, learn interaction information, and add it to the original input through residual connections. The output is flattened and dimensionality reduced to obtain the embedding vector representation of the drug, target protein, or disease that integrates multimodal information. Finally, the embedding vector representations of the two biological entities are multiplied element-wise and input into the MLP network. After nonlinear transformation, the interaction probability between the two biological entities is output.

[0058] Example 8 This embodiment verifies the effectiveness of the method of the present invention in practical application scenarios through specific drug discovery cases, including screening of active compounds for specific targets (wet experimental verification) and explanation of the side effect mechanisms of marketed drugs (reverse verification).

[0059] 1. Screening and validation of active compounds targeting the mineralocorticoid receptor (MR): Mineralocorticoid receptor (MR) is a key target for the treatment of hypertension, heart failure, and chronic kidney disease. To verify the ability of the model of this invention to discover new lead compounds, this embodiment removed all known MR modulators from the DrugBank dataset, retrained the model using the remaining data, and performed virtual screening on a self-built small molecule compound library.

[0060] Prediction results (see appendix) Figure 2 The diagram illustrates the structures and model predictions of potentially active small molecules. The model predicts several potentially highly active compounds, including five representative small molecules numbered C391, C640, C641, C974, and C977.

[0061] Molecular docking verification (see appendix) Figure 2Molecular docking analysis using AutoDock software revealed that all the aforementioned compounds could bind to the ligand-binding domain of MR with high affinity. In particular, C391, C640, and C641 formed stable hydrogen bonds or hydrophobic interactions with key residues of MR (such as Met810, Ser873, and Arg817), exhibiting binding modes similar to those of classic MR antagonists. C974 and C977, however, primarily interacted with Arg817 and Gln776, displaying different binding conformations.

[0062] In vitro wet test verification (see appendix) Figure 2 A HEK293T cell line expressing hMR protein was constructed, and the bioactivity of the compounds was detected using a luciferase reporter assay. The results showed that C391, C640, and C641 exhibited significant concentration-dependent inhibitory activity, confirming them as effective MR antagonists. C974 and C977, on the other hand, exhibited significant agonist activity, suggesting their potential as alternative therapies for Addison's disease. This result is highly consistent with the differential binding patterns observed in molecular docking, demonstrating that DrugGPS can not only recognize "binding" but also differentiate functional types to some extent.

[0063] 2. Explainability analysis of drug side effect mechanisms (using Finerenone as an example): To verify the model's capabilities in drug safety assessment and mechanism explanation, this embodiment conducted an in-depth association prediction analysis on the third-generation MR antagonist Finerenone.

[0064] Side effect prediction (drug-disease association task): Predicting potential disease phenotypes associated with Finerenone using the model of this invention (see appendix). Figure 3 The results showed that, in addition to its approved indications (such as chronic kidney disease and diabetic nephropathy), the model also gave a high predictive score for hyperkalemia. Clinical data confirm that hyperkalemia is indeed one of the most common side effects of Finerenone.

[0065] Mechanism reverse validation (drug-target task): To explain the source of the above side effects, the model of this invention was used to predict the full target of Finerenone (see appendix). Figure 3The results showed that the model predicted a very high association score (0.8887) between Finerenone and the potassium channel regulation-related protein Q9NZV8. This finding provides a reasonable mechanistic explanation for the clinical phenotype of "hyperkalemia" at the molecular target level.

[0066] 3. Drug repositioning potential analysis (taking Maprotiline as an example): This embodiment also examines the potential for new indications for the tetracyclic antidepressant maprotiline.

[0067] Drug-disease association prediction (see appendix) Figure 3 The model accurately predicted its original antidepressant indication (score 0.935). Furthermore, the model predicted its association with "Cancer" with scores between 0.68 and 0.87, suggesting its potential for anti-tumor retargeting.

[0068] Target validation: The model predicted that Maprotiline could potently antagonize M1 / M3 receptors (score >0.999, explaining its anticholinergic side effects), H1 receptors (explaining its sedative effect), and α1 receptors (explaining orthostatic hypotension). These high-confidence target predictions are in perfect agreement with clinically known pharmacological characteristics, further confirming the model's reliability in exploring the potential for repurposing existing drugs.

[0069] Example 9 To further promote the practical application of artificial intelligence prediction models, lower the barrier to entry for biomedical researchers, and accelerate the digitalization process of drug development, this embodiment constructs a visualized intelligent drug discovery platform system (DrugGPS Platform) based on the aforementioned prediction methods.

[0070] This platform is developed using Python, utilizes the Grado library to build the front-end interactive interface, and deploys a pre-trained multimodal attention fusion model (i.e., the model described in Example 1) on the back-end through a Python interface. Users do not need a deep background in artificial intelligence; they can complete complex data analysis and prediction tasks simply through a web interface.

[0071] This visualization platform mainly includes the following functional modules (see appendix) Figure 4 ): 1. Multi-mode prediction module: The platform integrates four core prediction modes to meet different R&D needs: Protein-to-Drug Prediction: Users input the amino acid sequence of a target protein, and the model performs reverse screening in a drug library, outputting a list of potential candidate drugs that may interact with the target and their prediction confidence scores. This is of great significance for ligand discovery of new targets.

[0072] Drug-to-Protein Prediction: Users input the string "SMILES" of a drug molecule, and the model predicts the set of target proteins it may bind to. This helps in elucidating the mechanism of action (MOA) of drugs and identifying potential off-target effects.

[0073] Disease-to-Drug Prediction: Screen potential therapeutic drugs for specific diseases (enter MeSH description or ID) to assist in drug repurposing research.

[0074] Drug-to-Disease Prediction: Predicting the disease phenotype that a particular drug may be associated with, used to assess the drug’s potential for expanding indications or the risk of potential side effects (such as the prediction of fenestrone side effects in Example 7).

[0075] 2. Visualization module for biological entity relationships: In addition to numerical prediction, the platform also features built-in network visualization capabilities. It can visually represent drugs, proteins, and disease entities in a dataset, along with their interrelationships, in the form of a graphical network. Users can clearly view the neighboring nodes of a specific drug in a heterogeneous network, intuitively understanding the topological relationships between "drug-target-disease," thus assisting researchers in interpreting prediction results from a systems biology perspective.

[0076] 3. Automated workflow: Users simply need to select a prediction dataset (such as DrugBank, Davis, etc.) on the interface, input or upload the feature data of the target biological entity (such as sequences or SMILES), and click the prediction button. The system can then call the trained Transformer encoder and MLP network in the background to perform inference and return the AI-generated prediction report within minutes.

[0077] Through this visualization platform, the present invention encapsulates complex deep learning algorithms into an intuitive "end-to-end" tool, realizing full automation of the process from data input, feature extraction, model inference to result display, greatly improving the efficiency and convenience of drug discovery research.

[0078] The methods of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code that, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses the code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for performing the methods shown herein.

[0079] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A biomolecular interaction prediction method based on multi-modal attention fusion, characterized in that, The method comprises the following steps: Obtaining multi-modal features of biological entities: drugs, targets, diseases or genes; the multi-modal features comprise sequence structure features, 3D structure features, similarity network features and biological relationship network embedding features; Constructing a feature fusion prediction model based on a multi-head self-attention mechanism and training the model; Inputting the multi-modal features of two biological entities into the trained feature fusion prediction model to output the probability of interaction between the two biological entities; the interaction comprises drug-target interaction, drug-disease interaction, drug-drug interaction, protein-gene interaction and the connection between genes and diseases; The feature fusion prediction model comprises a Transformer encoder and an MLP network; the multi-modal features are input into the feature fusion prediction model for stacking to generate a three-dimensional feature array; the features are processed by the multi-head self-attention mechanism in the Transformer encoder to learn interaction information, and the output result is added to the original input through a residual connection, flattened and dimensionally reduced to obtain an embedding vector representation of the drug, target protein or disease fused with multi-modal information; finally, the embedding vector representations of the two biological entities are multiplied element by element, and the MLP network is input to output the probability of interaction between the two biological entities after nonlinear transformation.

2. The method of claim 1, wherein, The sequence structure features comprise molecular sequence features of drugs, sequence features of targets, 3D structure features of drugs, 3D structure features of targets and semantic features of diseases. The molecular sequence features of drugs and the sequence features of targets are extracted by a pre-trained language model; the semantic features of diseases are constructed based on a medical subject heading database.

3. The method of claim 1, wherein the method is based on multi-modal attention fusion for biomolecular interaction prediction. The extraction process of the similarity network features is as follows: similarity matrices of drugs, target proteins, diseases and genes are calculated, a similarity network graph is constructed based on the similarity matrices, and a graph neural network is used to learn the representation of nodes to obtain the similarity network features.

4. The method of claim 3, wherein the method is characterized by, The construction process of the similarity network graph is as follows: based on the similarity matrix, the top N neighbor nodes of each entity node are selected and connected to construct a directed similarity graph G; the similarity graph G is input into a graph neural network for feature learning; wherein, for any node, the updated feature vector is calculated by aggregating the node information in the neighborhood.

5. The method of claim 1, wherein, The acquisition process of the biological relationship network embedding features is as follows: multiple biological databases are integrated to construct a heterogeneous network containing drug, protein, disease and gene entities, and a graph embedding algorithm is used to learn the topological structure representation of each node in the heterogeneous network to obtain the biological relationship network embedding features. The graph embedding algorithm comprises a graph embedding algorithm based on random walk or a graph autoencoder algorithm based on contrastive learning.

6. The method of claim 5, wherein the method is based on multi-modal attention fusion for biomolecular interaction prediction. When the heterogeneous network is a sparse data structure, the graph autoencoder algorithm based on contrastive learning is used; the graph autoencoder algorithm generates an enhanced graph by performing edge masking on the heterogeneous network and learns the node representation by using a graph neural network.

7. A biomolecular interaction prediction system based on multi-modal attention fusion, performing the prediction method of any one of claims 1-6, characterized in that, The method comprises the following steps: The feature extraction module: obtains multi-modal features of biological entities: drugs, targets or diseases; the multi-modal features include: sequence structure features, 3D structure features, similarity network features and biological relationship network embedding features; The model training module: constructs a feature fusion prediction model based on a multi-head self-attention mechanism and trains it; The prediction module: inputs the multi-modal features of two biological entities into the trained feature fusion prediction model, and outputs the probability of interaction between the two biological entities; the interaction includes: drug-target interaction, drug-disease interaction, drug-drug interaction, protein-gene interaction and the connection between genes and diseases; The feature fusion prediction model includes: a Transformer encoder and an MLP network; the multi-modal features are input into the feature fusion prediction model for stacking to generate a three-dimensional feature array; then the features are input into the Transformer encoder for processing using the multi-head self-attention mechanism to learn interaction information, and the output result is added to the original input through a residual connection, flattened and dimensionally reduced to obtain an embedding vector representation of the drug, target protein or disease that fuses multi-modal information; finally, the embedding vector representations of the two biological entities are multiplied element by element, and input into the MLP network, and after nonlinear transformation, the interaction probability between the two biological entities is output.

8. Use of the biomolecular interaction prediction method based on multi-modal attention fusion according to any one of claims 1-6 in screening targeted drug molecules.

9. An electronic device, comprising: It includes: a processor, a memory, a communication interface and a communication bus, which complete communication with each other through the communication bus; The memory is used to store at least one executable instruction, which makes the processor execute the corresponding operation of the biomolecular interaction prediction method based on multi-modal attention fusion according to any one of claims 1-6.

10. A computer program product comprising computer instructions, characterized in that, The computer instructions instruct the computing device to perform the corresponding operation of the biomolecular interaction prediction method based on multi-modal attention fusion according to any one of claims 1-6.

Images