A Drug Interaction Prediction Method and System Based on Multi-Source Heterogeneous Features and Sample-Driven Learning
By employing a multi-source heterogeneous feature and sample-driven learning approach, this method addresses the issues of low efficiency and class imbalance in existing DDI prediction methods, achieving more accurate drug interaction prediction and rare type identification, and providing high-quality drug feature representation and interpretable drug interaction mechanisms.
Patent Information
- Application Number
- CN202611259987.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-19
- Publication Date
- 2026-11-17
AI Technical Summary
Existing DDI prediction methods suffer from high cost, long processing time, and low efficiency, making it difficult to meet the needs of the rapid development of modern medicine. Furthermore, they face the problem of class imbalance, which leads to insufficient ability to identify rare interaction types.
We employ a method based on multi-source heterogeneous features and sample-driven learning. By using SMILES-driven Transformer and interaction information-driven graph information bottleneck dual-modal collaborative extraction, combined with KL divergence regularization and dynamic optimization of sample centers, we achieve deep fusion of global sequence semantics and key subgraph structure of drug molecules. Furthermore, we alleviate the class imbalance problem through the sample-driven learning mechanism.
It improves the accuracy of drug interaction prediction and the ability to identify rare types, provides high-quality drug feature representation and interpretable drug interaction mechanisms, and enhances the accuracy and practicality of prediction.
Smart Images

Figure CN123120108A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer-aided drug design technology, specifically to a method and system for predicting drug interactions based on multi-source heterogeneous features and sample-driven learning. Background Technology
[0002] Combination therapy is a common clinical treatment strategy, but physicochemical incompatibility can easily occur during combined drug use, leading to adverse drug interactions (DDIs). Therefore, systematic prediction and monitoring of DDIs are crucial for ensuring the safety of clinical medication use.
[0003] Traditional drug discovery indices (DDI) analysis methods primarily rely on in vivo and in vitro experiments, which suffer from high costs, long processing times, and low efficiency, making it difficult to meet the demands of high-throughput drug screening in modern medicine. In recent years, computational methods based on data mining and deep learning technologies have become the mainstream direction in DDI prediction research. Deep learning-based DDI prediction methods have significantly improved prediction efficiency and accuracy, promoting the development of precision medicine and personalized treatment. Existing DDI prediction methods can be mainly divided into the following three categories: The first category comprises molecular modeling methods based on SMILES sequences. These methods treat SMILES molecular representations as text sequences and employ natural language processing (NLP) techniques to extract semantic features of drug molecules. For example, ExDDI uses NLP generation to convert SMILES-encoded drug pairs into interpretable DDI mechanism descriptions; LLM-DDI combines large language models with biomedical knowledge graphs to enhance molecular understanding. However, as a linear representation, SMILES struggles to accurately represent complex chemical features such as ring structures and molecular symmetries, easily leading to structural misinterpretations and deviations from chemical principles, resulting in inherent information loss.
[0004] The second category comprises molecular graph-based modeling methods. To overcome the limitations of SMILES representation, some studies have turned to molecular graph methods, representing drugs as graph structures with atoms as nodes and chemical bonds as edges, and using graph neural networks to extract topological features. For example, GGI-DDI abstracts key molecular substructures as "particles" and captures their interactions; SRR-DDI integrates multi-scale substructure extraction and uses graph Transformers to dynamically model global dependencies between substructures. However, these methods still rely on local topological adjacency information, are sensitive to structural noise and missing information, and have significant limitations in modeling long-distance dependencies and global semantic information.
[0005] The third category is multimodal fusion modeling methods. Some studies attempt to integrate molecular structure and sequence information to achieve more comprehensive drug representation, but existing models still struggle to effectively utilize the complementary advantages of different modalities. For example, MSFF-MA-DDI mainly fuses heterogeneous features through simple splicing, lacking an effective information alignment mechanism, easily introducing redundant or conflicting information, and limiting intermodal complementarity.
[0006] Furthermore, existing DDI prediction models generally face the challenge of class imbalance. In the real world, the frequencies of different DDI types vary significantly, resulting in a severe long-tailed distribution in the training data. Existing methods primarily employ reweighting and resampling strategies for data-level adjustments, rather than addressing the representational bias caused by imbalance. This easily leads to overfitting due to repeated samples or instability in optimization due to manual parameter tuning. Even current state-of-the-art deep DDI models rarely introduce explicit mechanisms to preserve the compactness of minority class features, causing the decision boundary to be dominated by the majority class, resulting in bias and severely limiting the ability to identify clinically important rare interaction types. Summary of the Invention The purpose of this invention is to propose a drug interaction prediction method and system based on multi-source heterogeneous features and sample-driven learning. Through dual-modal collaborative extraction, redundancy suppression and dynamic optimization of class centers, it achieves more comprehensive and robust drug characterization and improves prediction accuracy and minority class recognition capabilities.
[0007] According to a first aspect of the embodiments of this disclosure, a drug interaction prediction method based on multi-source heterogeneous features and sample-driven learning is provided, comprising the following steps: Obtain drug pairs, convert them into standardized SMILES sequences and drug molecule maps, and remove duplicates and self-acting pairs; The SMILES sequence is segmented into semantic units using SMILES pairing encoding. After embedding and positional encoding, the units are input into a Transformer, which captures global semantics and long-distance dependencies through multi-head attention and outputs sequence features. Based on the node features extracted from the drug molecule graph, an interaction mapping matrix is constructed to fuse drug pair association information; key subgraphs are extracted through gating compression and graph information bottleneck, KL divergence regularization is added to retain core information and suppress redundancy, and subgraph features are output. The sequence features and the subgraph features are concatenated by dimension to obtain a comprehensive representation; Initialize sample centers for each DDI category, convert the distance between the comprehensive representation and each sample center into similarity; classify the similarity and the comprehensive representation, dynamically update the sample centers and model parameters through backpropagation based on the overall loss function, and output the DDI type.
[0008] In one embodiment, the Transformer is a stack of multiple Transformer encoders, each layer containing a multi-head self-attention sublayer and a feedforward network sublayer, and each layer is followed by a normalization layer and a residual connection; after outputting the sequence features, the sequence dimension is averaged and pooled to obtain a fixed-length sequence feature vector.
[0009] In one embodiment, the interaction mapping matrix is constructed as follows: for drug pairs (A, B), their drug molecule graph node feature matrices H are extracted respectively. A and H B The interaction matrix M = H is obtained. A *H B T The interaction matrix is used as a quantitative representation of information interaction between drug pairs, where H A Let H be the molecular graph node feature matrix of drug A. B H represents the molecular graph node feature matrix of drug B. B T For H B The transpose of .
[0010] In one embodiment, the gated compression is based on the interaction mapping matrix M, through a gate matrix. The original node features are weighted to obtain the gated compressed node features. ,in, It is the sigmoid activation function. and These are learnable parameters.
[0011] In one embodiment, the optimization objective of the graph information bottleneck is: Where G represents the drug molecule diagram. The subgraph for extracting information bottlenecks in the graph. For DDI tags, Indicates mutual information, To balance the hyperparameters, KL divergence regularization is used to constrain the subgraph. The dependence on drug input graph G retains the core substructure related to DDI prediction.
[0012] In one embodiment, the distance is an Euclidean distance: Similarity is: in, As a comprehensive representation vector, Let j be the sample center vector of the j-th DDI category. represents the dimension of the vector.
[0013] In one embodiment, the similarity and the comprehensive representation are concatenated and input into a fully connected classification layer, and after softmax activation, the probability distribution of each category is output; the overall loss function is the sum of cross-entropy loss and sample center regularization loss.
[0014] According to a second aspect of the embodiments of this disclosure, a drug interaction prediction system based on multi-source heterogeneous features and sample-driven learning is provided, comprising: The data preprocessing module acquires drug pairs, converts them into standardized SMILES sequences and drug molecule maps, and removes repetitions and self-acting pairs. The SMILES-driven Transformer module uses SMILES pairing encoding to segment the SMILES sequence into semantic units, which are then embedded and positionally encoded before being input into the Transformer. Multi-head attention is used to capture global semantics and long-distance dependencies, and the sequence features are output. The interactive information-driven graph information bottleneck module extracts node features based on the drug molecule graph and constructs an interactive mapping matrix to fuse drug pair association information. It extracts key subgraphs through gating compression and graph information bottleneck, adds KL divergence regularization to retain core information and suppress redundancy, and outputs subgraph features. The multi-source feature fusion module concatenates the sequence features and the sub-graph features according to dimensions to obtain a comprehensive representation. The sample-driven learning classification module initializes sample centers for each DDI category, converts the distance between the comprehensive representation and each sample center into a similarity, fuses the similarity with the comprehensive representation, and then classifies the DDI category. Based on the overall loss function, it dynamically updates the sample centers and model parameters through backpropagation and outputs the DDI type.
[0015] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the drug interaction prediction method based on multi-source heterogeneous features and sample-driven learning.
[0016] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the drug interaction prediction method based on multi-source heterogeneous features and sample-driven learning.
[0017] The advantages of the above technical solutions adopted in this invention compared with the prior art are as follows: 1. This invention employs a dual-modal collaborative extraction method, driven by SMILES-driven Transformer and interactive information-driven graph information bottleneck, to capture the global sequence semantics and key subgraph structure information of drug molecules. By deeply integrating the two, it overcomes the shortcomings of incomplete information in existing single-modal methods, achieves complementary representation of molecular sequence and spatial structure, and provides more accurate and complete drug feature representation, thus providing high-quality input for subsequent classification.
[0018] 2. This invention uses a sample-driven learning mechanism to maintain dynamically optimizable category sample centers for each DDI category. Classification decisions are based on the similarity between samples and category centers rather than the number of category samples. This effectively alleviates the decision boundary tilt problem caused by category imbalance at the feature representation level, and significantly improves the accuracy and recall of identifying clinically important rare drug interaction types.
[0019] 3. This invention focuses on the key sub-graphs extracted from the graph information bottleneck, which are highly correlated with the DDI prediction of pharmacological molecules. This can clearly reveal the mechanism of action of drug-drug interactions and provide intuitive and interpretable evidence for clinical decision-making.
[0020] 4. The dual-modal feature extraction and classification mechanism of this invention achieves a good balance between model expressive power and computational overhead, and has excellent practicality and promotion value. Attached Figure Description
[0021] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.
[0022] Figure 1 This is a schematic diagram of a drug interaction prediction method based on multi-source heterogeneous features and sample-driven learning. Detailed Implementation
[0023] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.
[0024] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0025] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0026] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.
[0027] Example 1: This embodiment provides a drug interaction prediction method based on multi-source heterogeneous features and sample-driven learning, including the following steps: S1. Obtain drug pairs, convert them into standardized SMILES sequences and drug molecule maps, and remove duplicates and self-acting pairs; Specifically, for each drug's SMILES sequence, RDKit is invoked to attempt to parse it into an RDKit molecular object. If parsing is successful and the returned molecular object is not empty, the SMILES sequence is considered valid. If parsing fails (e.g., the SMILES string contains syntax errors, abnormal atomic valence states, or chemical symbols that RDKit cannot recognize), the drug is removed from the dataset, and the corresponding error log is recorded for subsequent investigation. Furthermore, for drug pairs containing two drugs, if either drug is deemed invalid, the entire drug pair is discarded.
[0028] After passing the above legality checks, the valid SMILES sequences are further processed using RDKit to generate corresponding drug molecule graph structures: each atom in the molecule is extracted as a graph node, and each chemical bond is extracted as a graph edge. Initial feature vectors at both the atomic and chemical bond levels are extracted to serve as node and edge inputs for subsequent encoding in the graph isomorphic network, respectively. Furthermore, for valid drug pairs that pass the checks, cases where two drugs are identical (i.e., self-interacting pairs where A=A) are removed, and duplicate combinations of the same drug pair are eliminated from the dataset to ensure data purity for subsequent model training and evaluation.
[0029] S2. The SMILES sequence is segmented into semantic units using SMILES pairing encoding, and after embedding and positional encoding, it is input into the Transformer. Multi-head attention is used to capture global semantics and long-distance dependencies, and the sequence features are output. In this embodiment, a SMILES pairing encoding strategy is used to segment the two SMILES sequences of the drug pair into semantic units. These semantic units mainly include functional groups and ring structures. Specifically, based on a predefined chemical semantic dictionary, regular expression matching is used to identify ring structures (such as benzene rings, furan rings, pyridine rings, etc.) and functional groups (such as carboxyl-COOH, hydroxyl-OH, amino-NH2, carbonyl C=O, etc.) in the SMILES sequence. The complete SMILES sequence is then segmented into multiple sequentially arranged semantic unit fragments according to the boundaries of these characteristic groups. Each semantic unit is treated as a token to preserve the semantic integrity of the chemical groups.
[0030] Subsequently, each semantic unit is mapped to an index via one-hot encoding and then transformed into a dense vector through a learnable embedding layer. Positional encoding is also superimposed to preserve sequence order information, ultimately forming an embedded sequence input Transformer encoder. The Transformer encoder models the dependencies between any two positions in the sequence using a multi-head self-attention mechanism, thereby capturing the global semantics and long-range dependencies of the molecular sequence. After outputting sequence features, the Transformer encoder performs average pooling on the sequence dimension to obtain a fixed-length sequence feature vector.
[0031] S3. Extract node features based on drug molecule graph, construct an interaction mapping matrix to fuse drug pair association information; extract key subgraphs through gating compression and graph information bottleneck, add KL divergence regularization to retain core information and suppress redundancy, and output subgraph features; In this embodiment, the drug molecule is represented as a molecular graph with atoms as nodes and chemical bonds as edges. A graph isomorphic network is used as the encoder, and the features of the node itself and its neighboring nodes are aggregated through multiple rounds of message passing to update the final feature representation of each node, thereby obtaining the node feature matrix of the drug molecule.
[0032] For a drug pair (A, B), after extracting the node feature matrices of both, the interaction mapping matrix is calculated by matrix multiplication. Each element of this matrix represents the interaction strength between a certain atom in drug A and a certain atom in drug B, thus explicitly modeling the association information between drug pairs as an atomic-level interaction strength map.
[0033] Based on the interaction mapping matrix, an important interaction information is selected and redundant interactions are filtered out through a gating compression mechanism. That is, the gating weight is calculated using the interaction mapping matrix to weight the original node features, thereby retaining key interaction information and suppressing irrelevant information.
[0034] Subsequently, based on the graph information bottleneck theory, while retaining information highly relevant to the DDI prediction task, the redundant information in the input graph is compressed to the maximum extent, and the key subgraphs that contribute most to the DDI prediction are extracted. In this process, KL divergence regularization is introduced to constrain the information dependency between the extracted subgraphs and the original input graph, further suppressing redundant information irrelevant to the prediction task.
[0035] Finally, graph-level readout operations are performed on the node features of the extracted key subgraphs to obtain a fixed-length subgraph feature vector.
[0036] S4. Concatenate the sequence features and the subgraph features according to dimensions to obtain a comprehensive representation; In this embodiment, after extracting sequence features and subgraph features, the aforementioned bimodal features are fused. Specifically, the sequence feature vector obtained in step S2 and the subgraph feature vector obtained in step S3 are concatenated along the feature dimension, that is, the two feature vectors are linked end-to-end to form a comprehensive representation vector with a dimension equal to the sum of the two dimensions. This comprehensive representation vector integrates the sequence semantic information of the drug molecule and the key subgraph structural information, and serves as the input to the subsequent classification module.
[0037] S5. Initialize sample centers for each DDI category, convert the distance between the comprehensive representation and each sample center into similarity; classify the similarity and the comprehensive representation, dynamically update the sample centers and model parameters through backpropagation based on the overall loss function, and output the DDI type.
[0038] In this embodiment, a learnable category sample center is initialized for each DDI category. This sample center represents the typical representation of the category in the feature space. For the comprehensive representation obtained in step S4, the distance (e.g., Euclidean distance) between it and each category sample center is calculated. The calculated distance is converted into a similarity score, with a smaller distance indicating a higher similarity.
[0039] The similarity and the comprehensive representation are fused (e.g., by concatenation). The fused features are then input into a fully connected classification layer, activated by softmax, and output as predicted probability distributions for each category. The overall loss function for model training consists of two parts: cross-entropy loss and sample center regularization loss. Cross-entropy loss ensures classification accuracy, while sample center regularization loss constrains features of similar samples to cluster towards their category centers, enhancing the compactness of category features. Based on the overall loss function, the classifier parameters and category sample centers are simultaneously optimized using a gradient descent algorithm, and the sample centers are updated with momentum based on batch sample features during training. Finally, the category with the highest probability in the predicted probability distribution is selected as the DDI prediction result and output.
[0040] Example 2: This embodiment provides a drug interaction prediction system based on multi-source heterogeneous features and sample-driven learning, including: The data preprocessing module acquires drug pairs, converts them into standardized SMILES sequences and drug molecule maps, and removes repetitions and self-acting pairs. The SMILES-driven Transformer module uses SMILES pairing encoding to segment the SMILES sequence into semantic units, which are then embedded and positionally encoded before being input into the Transformer. Multi-head attention is used to capture global semantics and long-distance dependencies, and the sequence features are output. The interactive information-driven graph information bottleneck module extracts node features based on the drug molecule graph and constructs an interactive mapping matrix to fuse drug pair association information. It extracts key subgraphs through gating compression and graph information bottleneck, adds KL divergence regularization to retain core information and suppress redundancy, and outputs subgraph features. The multi-source feature fusion module concatenates the sequence features and the sub-graph features according to dimensions to obtain a comprehensive representation. The sample-driven learning classification module initializes sample centers for each DDI category, converts the distance between the comprehensive representation and each sample center into a similarity, fuses the similarity with the comprehensive representation, and then classifies the DDI category. Based on the overall loss function, it dynamically updates the sample centers and model parameters through backpropagation and outputs the DDI type.
[0041] The above modules can be deployed on the same device or distributed devices; the division of modules is only a functional logic description and does not limit the specific physical boundaries or implementation order.
[0042] Example 3: An electronic device is provided for running the aforementioned "Drug Interaction Prediction Method Based on Multi-Source Heterogeneous Features and Sample-Driven Learning". The electronic device includes a processor, a memory, and optional communication interfaces / display devices / input devices, etc.; the memory stores a computer program that can run on the processor, and when the processor executes the program, it implements steps S1 to S5 of the method described in Embodiment 1, specifically including but not limited to: S1. Obtain drug pairs, convert them into standardized SMILES sequences and drug molecule maps, and remove duplicates and self-acting pairs; S2. The SMILES sequence is segmented into semantic units using SMILES pairing encoding, and after embedding and positional encoding, it is input into the Transformer. Multi-head attention is used to capture global semantics and long-distance dependencies, and the sequence features are output. S3. Extract node features based on drug molecule graph, construct an interaction mapping matrix to fuse drug pair association information; extract key subgraphs through gating compression and graph information bottleneck, add KL divergence regularization to retain core information and suppress redundancy, and output subgraph features; S4. Concatenate the sequence features and the subgraph features according to dimensions to obtain a comprehensive representation; S5. Initialize sample centers for each DDI category, convert the distance between the comprehensive representation and each sample center into similarity; classify the similarity and the comprehensive representation, dynamically update the sample centers and model parameters through backpropagation based on the overall loss function, and output the DDI type.
[0043] The electronic device hardware can be one of a server, personal computer, workstation, industrial controller, edge computing device, or mobile terminal; the processor can be a general-purpose CPU, GPU, NPU, FPGA, or a combination thereof; the memory can be RAM, ROM, flash memory, or disk array. The device can interact with local / remote data storage via a communication interface. The above hardware configuration does not constitute a limitation of the present invention.
[0044] Example 4: A computer-readable storage medium storing a computer program, which, when run on a processor of an electronic device, causes the program to execute the method steps S1 to S5 described in Embodiment 1; the storage medium may be a disk, optical disk, flash memory, solid-state drive, read-only memory, random access memory, or any combination of the above media.
[0045] Those skilled in the art will understand that the modules or steps described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, which can then be stored in a storage device for execution by a computer device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. This disclosure is not limited to any particular combination of hardware and software.
[0046] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
[0047] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.
Claims
1. A drug interaction prediction method based on multi-source heterogeneous features and sample-driven learning, characterized in that, Includes the following steps: Obtain drug pairs, convert them into standardized SMILES sequences and drug molecule maps, and remove duplicates and self-acting pairs; The SMILES sequence is segmented into semantic units using SMILES pairing encoding. After embedding and positional encoding, the units are input into a Transformer, which captures global semantics and long-distance dependencies through multi-head attention and outputs sequence features. Based on the node features extracted from the drug molecule graph, an interaction mapping matrix is constructed to fuse drug pair association information; key subgraphs are extracted through gating compression and graph information bottleneck, KL divergence regularization is added to retain core information and suppress redundancy, and subgraph features are output. The sequence features and the subgraph features are concatenated by dimension to obtain a comprehensive representation; Initialize sample centers for each DDI category, convert the distance between the comprehensive representation and each sample center into similarity; classify the similarity and the comprehensive representation, dynamically update the sample centers and model parameters through backpropagation based on the overall loss function, and output the DDI type.
2. The drug interaction prediction method based on multi-source heterogeneous features and sample-driven learning according to claim 1, characterized in that, The Transformer is a stack of multiple Transformer encoders, each layer containing a multi-head self-attention sublayer and a feedforward network sublayer. Each layer is followed by a normalization layer and a residual connection. After outputting the sequence features, the sequence dimension is averaged and pooled to obtain a fixed-length sequence feature vector.
3. The drug interaction prediction method based on multi-source heterogeneous features and sample-driven learning according to claim 1, characterized in that, The interactive mapping matrix is constructed as follows: for drug pairs (A, B), extract their drug molecule graph node feature matrices H. A and H B The interaction matrix M = H is obtained. A *H B T The interaction matrix is used as a quantitative representation of information interaction between drug pairs, where H A Let H be the molecular graph node feature matrix of drug A. B H represents the molecular graph node feature matrix of drug B. B T For H B The transpose of .
4. The drug interaction prediction method based on multi-source heterogeneous features and sample-driven learning according to claim 1, characterized in that, The gated compression is based on the interaction mapping matrix M, through a gated matrix. The original node features are weighted to obtain the gated compressed node features. ,in, It is the sigmoid activation function. and These are learnable parameters.
5. The drug interaction prediction method based on multi-source heterogeneous features and sample-driven learning according to claim 1, characterized in that, The optimization objective of the graph information bottleneck is: Where G represents the drug molecule diagram. The subgraph for extracting information bottlenecks in the graph. For DDI tags, Indicates mutual information, To balance the hyperparameters, KL divergence regularization is used to constrain the subgraph. The dependence on drug input graph G retains the core substructure related to DDI prediction.
6. The drug interaction prediction method based on multi-source heterogeneous features and sample-driven learning according to claim 1, characterized in that, The distance is the Euclidean distance: Similarity is: in, As a comprehensive representation vector, Let j be the sample center vector of the j-th DDI category. represents the dimension of the vector.
7. The drug interaction prediction method based on multi-source heterogeneous features and sample-driven learning according to claim 1, characterized in that, The similarity and the comprehensive representation are concatenated and input into a fully connected classification layer. After softmax activation, the probability distribution of each category is output. The overall loss function is the sum of cross-entropy loss and sample center regularization loss.
8. A drug interaction prediction system based on multi-source heterogeneous features and sample-driven learning, characterized in that, include: The data preprocessing module acquires drug pairs, converts them into standardized SMILES sequences and drug molecule maps, and removes repetitions and self-acting pairs. The SMILES-driven Transformer module uses SMILES pairing encoding to segment the SMILES sequence into semantic units, which are then embedded and positionally encoded before being input into the Transformer. Multi-head attention is used to capture global semantics and long-distance dependencies, and the sequence features are output. The interactive information-driven graph information bottleneck module extracts node features based on the drug molecule graph and constructs an interactive mapping matrix to fuse drug pair association information. It extracts key subgraphs through gating compression and graph information bottleneck, adds KL divergence regularization to retain core information and suppress redundancy, and outputs subgraph features. The multi-source feature fusion module concatenates the sequence features and the sub-graph features according to dimensions to obtain a comprehensive representation. The sample-driven learning classification module initializes sample centers for each DDI category, converts the distance between the comprehensive representation and each sample center into a similarity, fuses the similarity with the comprehensive representation, and then classifies the DDI category. Based on the overall loss function, it dynamically updates the sample centers and model parameters through backpropagation and outputs the DDI type.
9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the drug interaction prediction method based on multi-source heterogeneous features and sample-driven learning as described in any one of claims 1-7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the drug interaction prediction method based on multi-source heterogeneous features and sample-driven learning as described in any one of claims 1-7.
