A multi-modal feature-based drug target interaction multi-task prediction method

Abstract

The application discloses a drug target interaction multi-task prediction method based on multi-modal features. The method comprehensively utilizes drug multi-dimensional structure features and drug target text semantic features to construct a multi-modal feature embedding framework. By designing a multi-core magnetic relationship convolution layer, introducing a magnetic Laplace matrix and a frequency domain graph filter, a DTI multi-classification prediction model MFFGCN based on a graph convolutional neural network is constructed. Furthermore, a DTI natural language explanation text generation task is proposed, and a DTI natural language explanation text generation model EXDTI is used to formalize the drug target interaction prediction into natural language explanation text generation, and readable mechanism explanation and relationship explanation are generated. The method performs excellently in the DTI multi-task joint prediction, has good cold start adaptability, and provides an efficient and reliable computing tool for new drug research and development and candidate screening.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and bioinformatics, specifically relating to a multi-task prediction method for drug target interactions based on multimodal features. Background Technology

[0002] Drug-target interaction (DTI) prediction is a core task in drug discovery and development. Identifying the interaction between drugs and target proteins can effectively guide new drug design, pharmacodynamic mechanism analysis, and drug retargeting. However, traditional experimental methods for large-scale identification of novel DTIs, such as affinity assays and cell viability assays, are costly and time-consuming, making them unsuitable for the high-throughput and high-precision demands of modern drug development. Therefore, automating the prediction of drug-target interactions using computational methods has become an important research direction at the intersection of pharmaceutical informatics and artificial intelligence.

[0003] Early computational prediction methods were largely based on machine learning algorithms, such as support vector machines, random forests, and matrix factorization. These methods typically rely on manually extracting structural features or molecular descriptors of drugs and targets, and then performing task-specific classification or regression predictions. However, due to the vast medicinal chemistry space and complex feature dimensions, traditional machine learning methods struggle to capture the nonlinear dependencies between drugs and targets, resulting in limited generalization ability for predicting unknown drugs or new targets.

[0004] With the development of deep learning, researchers have proposed various DTI prediction models based on neural networks. Convolutional neural networks and recurrent neural networks can automatically extract latent features in one-dimensional or two-dimensional feature spaces for feature learning of drug compounds and protein sequences; while graph neural networks (GNNs) further model drugs and targets as nodes and edges of graph structure data, enabling them to capture complex nonlinear relationships in drug-target networks, thus achieving more expressive feature modeling. In particular, graph convolutional networks (GCNs) and graph attention networks (GATs) have achieved significant results in drug discovery tasks.

[0005] However, existing deep learning models still have the following limitations: (1) Most methods are based on the two-dimensional structure or protein sequence of drugs, ignoring the important influence of the three-dimensional conformation information of drugs on intermolecular interactions; (2) Most models generally do not utilize the biological semantic information contained in the natural language description of drugs and targets, such as pharmacological mechanisms, biological pathways and clinical efficacy, which can provide richer contextual knowledge for the model; (3) Existing graph neural networks have limitations in feature modeling in the spatial domain and spectral domain—spatial domain methods are difficult to characterize global structural features, while spectral domain methods are not flexible enough when dealing with heterogeneous drug networks; (4) The prediction task type is single, only targeting binary or multi-class prediction tasks of DTI; (5) Most models rely on large-scale labeled data, and the prediction performance drops significantly in data-scarce or cold-start scenarios, with poor generalization ability.

[0006] To address these issues, recent studies have attempted to combine multimodal feature information, such as chemical structures, molecular diagrams, and biological text descriptions, to improve the accuracy of DTI predictions.

[0007] Among numerous deep learning methods, graph neural networks, as a representative framework based on graph structure data that has emerged in recent years, have attracted widespread attention due to their ability to effectively model complex biomedical entities and their interactions. Unlike traditional sequence or vectorized modeling methods, graph neural networks can directly represent the relationship between drugs and targets in the form of graph structures, modeling biomolecules such as drugs and protein targets as nodes in the graph, and modeling drug-target interactions as edges connecting these nodes.

[0008] Through this graph structure modeling approach, graph neural networks can comprehensively utilize the characteristics of nodes themselves and the structural information of their neighborhoods during network propagation, thereby capturing the nonlinear dependency relationship between drugs and targets.

[0009] In summary, these models not only reflect the topological features of drug chemical structures, but also learn complex semantic information from the global and local relationships of molecular graphs, providing more efficient computational support for drug discovery, target identification, and pharmacodynamic mechanism inference, and greatly improving efficiency and success rate in actual drug development.

[0010] In recent years, Natural Language Interpretation (NLE) text generation for DTI has provided another promising direction for further exploration of DTI prediction. NLE text generation aims to generate natural language text explanations for DTI prediction tasks, helping users better understand model behavior and make decisions. Summary of the Invention

[0011] The present invention explores the following issues: (1) Based on graph neural networks, DTI symbol heterogeneous graphs are constructed, and multi-dimensional features such as two-dimensional chemical features and three-dimensional conformational features of drugs and semantic features derived from the large language model of drug targets are integrated to realize the unified embedding representation of multi-source information; (2) A graph convolution mechanism combining spectral domain and spatial domain is introduced to capture global topological features and local reaction event patterns in the drug target network; (3) A DTI natural language interpretation text generation task is further proposed, and the one-dimensional features of drug targets are encoded and then input into the trained DTI natural language interpretation text generation model EXDTI; (4) DTI multi-classification prediction and DTI natural language interpretation text generation multi-task prediction are performed simultaneously; (5) It still has high robustness and generalization ability in the cold start scenario, providing efficient intelligent computing support for drug discovery and target identification.

[0012] Based on this, this invention proposes a multi-task prediction method for drug target interactions based on multimodal features. By introducing a DTI multi-classification prediction model MFFGCN based on a multi-kernel magnetic relation graph convolutional structure, a DTI natural language interpretation text generation model EXDTI, and a multimodal feature extraction model, it achieves deep fusion of multimodal features of drug targets and completes DTI multi-task prediction, significantly improving the model's generalization and application value. Figure 1 As shown.

[0013] The present invention solves the above-mentioned technical problems through the following technical solution, which includes the following steps: Figure 2 As shown:

[0014] S1: DTI Data Preprocessing. To implement a cold start scenario, the two collected DTI datasets are divided into training and testing sets according to a set ratio; and feature and textual semantic data information are obtained from databases such as DrugBank, PubChem, and UniPort3 based on the drug target node information in the DTI datasets.

[0015] S2: Model Building. This involves building the initial DTI multi-class prediction model MFFGCN, the DTI natural language interpretation text generation model EXDTI, and models for extracting multimodal features of drugs and targets, including VideoMol and BGE.

[0016] S3: Model Training. The two initial DTI prediction models are trained using the training set to obtain the trained DTI prediction model.

[0017] S4: DTI Multi-Task Prediction. The trained DTI prediction model is tested to obtain the DTI prediction results for the test set and output the corresponding evaluation metrics.

[0018] Furthermore, in step S1, the specific processing procedure for DTI data preprocessing is as follows:

[0019] S5: Construct a cold start scenario by dividing the two DTI datasets into training and testing sets in a ratio of 0.8:0.2. The DTI datasets include information such as the number of drug target nodes, DrugBankIDs, and the number of their respective DTIs.

[0020] S6: Based on the drug DrugBankID, obtain complete drug information from DrugBank, including drug name, SMILES structural sequence and semantic text description, etc. The semantic text description includes information such as therapeutic use, pharmacological mechanism, pharmacokinetic properties and physicochemical properties.

[0021] S7: Based on the target DrugBankID, obtain complete target information from DrugBank, including the target SwissID, name, and SEQUENCE amino acid sequence, and obtain the corresponding semantic text description from UniProt3.

[0022] S8: Based on the drug DrugBankID, complete 2D features of the drug, including information such as chemical structure and biological characteristics, were extracted from databases such as Drugbank and Puchem.

[0023] S9: Map the target SwissID obtained from DrugBank to the String database, and obtain the target-target interactions based on the mapping results. These interactions are adjacency matrix features and serve as an embedding feature of the target. .

[0024] Furthermore, in step S2, the specific processing steps for extracting the drug and target multimodal features from the models VideoMol and BGE are as follows:

[0025] S10: Employs a pre-trained video model, VideoMol, to extract effective 3D structural features from the three-dimensional conformation of the drug.

[0026] S11: Relevant textual descriptions of drugs and targets extracted from Drugbank and UniProt3 are input into the pre-trained language model BGE for high-dimensional semantic space mapping of natural text language. By directly applying this model, LLM-derived semantic features can be extracted from the natural textual descriptions of drugs and proteins.

[0027] Furthermore, in step S10, the specific processing procedure for acquiring the 3D features of the drug is as follows:

[0028] S12: Specifically, firstly, a three-dimensional conformation of the drug molecule is generated, and then it is rotated around the coordinate axes by a certain angle. Rotation angle ,in This indicates the number of samples per rotation axis. Render 20 frames per rotation axis. Molecular video, each frame of which is represented as ( Subsequently, a frame of image is randomly sampled from the molecular video corresponding to each rotation axis as the corresponding view. For each view, VideoMol encodes it into a view representation vector, as follows: .

[0029] S13: The three-view representation vector matrix of all drugs is represented as follows Its splicing serves as a 3D feature of the drug. Finally, the drug node partial embedding will be initialized as a concatenation of 2D and 3D features, as follows: .

[0030] S14: The pre-trained language model BGE is used to extract effective derived semantic initialization embedding features, i.e., LLM features, from the textual description of drugs and targets.

[0031] Furthermore, in step S11, the specific processing procedure for obtaining the LLM features of the drug and target is as follows:

[0032] S15: Specifically, for the drug component, a JSON file is first generated based on the text description obtained from DrugBank. This JSON file is then input into the BGE language model for high-dimensional semantic space mapping of the natural text language. By directly applying this model, embeddings can be extracted from the natural text language description of the drug without any additional parameter updates.

[0033] S16: For the target part, firstly, a JSON file is generated based on the text description obtained from UniPort3. Then, the JSON file is input into the language model BGE to perform high-dimensional semantic space mapping of natural text language and extract the target LLM embedding features.

[0034] S17: For a given text description of a drug and target and pre-trained language models The corresponding semantic calculation formula is: .in and These represent the LLM characteristics of the drug and the target, respectively.

[0035] S18: This allows us to obtain the LLM-derived semantic initialization feature embeddings for the drug and target. .

[0036] S19: The final embedding of drug and target nodes is obtained through concatenation operations: and .

[0037] Furthermore, in step S2, the construction process of the DTI multi-class prediction model MFFGCN is as follows:

[0038] S20: Initialize the DTI graph neural convolutional network with the drug target node features and DTI input from the training set.

[0039] S21: The input DTI is used as an edge in the graph convolutional neural network, thus modeling the DTI network as a DTI symbolic heterogeneous graph. Construct the DTI multi-class prediction model MFFGCN. For example... Figure 3 As shown.

[0040] S22: where It is by The nodes composed of various drugs Indicates drug nodes; It is by Nodes composed of various targets Represents the target node. Edge set Indicates drug With target Edge links. Indicates from the first The drug to the first The edge of a target, i.e., the drug For the target The impact.

[0041] S23: In addition, all drug node feature matrices Each line represents Drug properties include characteristics such as chemical structure, target protein, and enzyme protein.

[0042] S24: Divide the DTI symbolic heterogeneous graph into two symbolic heterogeneous subgraphs, one positive and one negative.

[0043] S25: Further, in step S24, the DTI symbol heterogeneous graph is divided into positive and negative symbol heterogeneous subgraphs to better represent the enhancing and inhibiting mechanisms of DTI. The positive and negative symbol heterogeneous subgraphs are respectively represented as follows: .

[0044] S26: Define the magnetic map convolution kernel MKConv for the DTI multi-class prediction model MFFGCN. The process is as follows:

[0045] S27: Further, in step S25, let and These represent the adjacency matrices corresponding to the heteroprotic subgraphs of the positive and negative DTI symbols, respectively. ,otherwise .

[0046] S28: Further, in step S26, the design of MKConv allows the model to adaptively capture heterogeneous structure information in a local range, while retaining the advantages of continuity and boundary modeling in frequency domain filtering.

[0047] S29: Normalize the adjacency matrix and then perform graph filtering to obtain the normalized graph convolution matrix on the heteroprotic graph with positive and negative signs.

[0048] S30: Furthermore, in step S29, by using a given low-to-high frequency response function : and This allows us to obtain the low-to-high frequency graph filtering matrix on the positive and negative sign heteroproton graph. .

[0049] S31: The embedding representation of the drug target node in the... Layer is denoted as It is made of and It consists of two parts, respectively, through the symbolic heteroproton diagram. and The spectrogram is obtained by performing a convolution process on it.

[0050] S32: Furthermore, the node form of the graph convolution process is represented as follows: , , , .

[0051] S33: Among them There are two learnable parameters, corresponding to the node itself and its neighboring nodes, respectively. A concatenation operation is used to integrate these four parts, and a hierarchical nonlinear activation function is applied. In the experiment, the specific activation function was: The final node is represented as .

[0052] S34: From a spectral analysis perspective, the low-frequency and high-frequency response functions achieve feature decomposition for different frequency components in the frequency domain; from a node message passing perspective, the dual-frequency channels, combined with a relationship-aware aggregation mechanism, enable nodes to obtain rich information from both local consistency and local differences during feature updates. Overall, MKConv achieves modeling of the DTI's operational mode.

[0053] S35: Embed the drug target nodes in the test set into the trained DTI multi-class prediction model MFFGCN, and obtain the DTI multi-class prediction results.

[0054] Furthermore, in step S2, the EXDTI (Extreme Derivative Language Interpretation) text generation model construction process is as follows:

[0055] S36: Collect paradigm data and one-dimensional features describing drug-target interaction mechanisms from the training and testing sets in DrugBank, namely drug SMILES and target SEQUENCE.

[0056] S37: Constructing the DTI Natural Language Interpretation Text Generation Model EXDTI: Using the molecular-text pre-trained model MolT5 and the protein language model ProtBERT to encode one-dimensional features, namely drug SMILES and target SEQUENCE, to achieve the model's mapping from "SMILES + SEQUENCE" to "natural language interpretation". For example... Figure 4 As shown.

[0057] S38: Use the molecular-text pre-trained model MolT5 and the protein language pre-trained model ProtBERT to encode drug SMILES and target SEQUENCE.

[0058] S39: The maximum target sequence length, max_target_length, is explicitly specified in the model configuration.

[0059] S40: Call the pre-trained molecular-text model MolT5 to translate the chemical structures SMILES of the drug into a natural language description Caption.

[0060] S41: Similarly, the pre-trained protein language model ProtBERT is invoked to translate the target sequence features (SEQUENCE) into a natural language description (Caption).

[0061] S42: Assemble the various information of the above drugs and targets into a source sequence according to a specific template: drug[SMILES]([drug_Caption]); target[SEQUENCE]([target_Caption]).

[0062] S43: In step S42, the source sequence clearly distinguishes two entities, including the drug and the target.

[0063] S44: The target sequence of the training data is formatted as follows: <label>Explanation <exp>.

[0064] S45: Among them, <label>This is a category tag. "Explanation" is a key trigger word, namely "TriggerToken".

[0065] S46: Convert the above text into corresponding Token IDs.

[0066] S47: During the inference phase, the DTI natural language interpretation text generation model EXDTI generates interpretation text autoregressively based on the input drug-target pair.

[0067] S48: Using the text generation loss function .

[0068] S49: Where N represents the size of the training set, T represents the length of the target sequence, and Θ represents the parameters of the encoder and decoder.

[0069] S50: Further, in step S47, the BLEU (Bilingual Evaluation Understudy) index is introduced to quantify the accuracy of the generated explanation.

[0070] S51: Train the initial model using the source sequences of drug targets in the training set to obtain the trained EXDTI model.

[0071] S52: Input the drug-target pairs from the test set into the trained DTI natural language interpretation text generation model ExDTI to generate natural language interpretation text.

[0072] S53: Further, in step S52, an end-to-end ExDTI model is used. As follows: Input: DRUG <smiles>,TARGET <sequence>Output: <label>Explanation: <exp>.

[0073] S54: The final prediction results of drug target interaction are obtained based on the DTI multi-class prediction model MFFGCN and the DTI natural language interpretation text generation model EXDTI.

[0074] S55: Further, in step S54, the DTI multi-class prediction results are matched and mapped with the DTI natural language interpretation text generation prediction results to finally output the result.

[0075] S56: If the matches are the same, output the prediction results of both tasks directly; otherwise, output the prediction result of the task with higher accuracy.

[0076] The present invention has the following advantages over the prior art:

[0077] 1. In the drug target feature extraction part, different modal features are integrated to improve the rich feature information of drugs and targets in text description, chemical structure and molecular diagram, make up for the limitations of traditional two-dimensional characterization, and finally obtain more comprehensive drug target feature information.

[0078] 2. A multi-kernel magnetic relation convolutional layer was designed to model various DTI semantic information. This convolutional kernel introduces multiple magnetic Laplacian matrices and frequency domain graph filters to globally model the interaction forms of DTI relations, thereby constructing the DTI multi-classification prediction model MFFGCN;

[0079] 3. DTI Natural Language Interpretation Text Generation: Through a novel task—natural language interpretation of drug target interactions—EXDTI, an innovative DTI natural language interpretation text generation model is constructed, formalizing DTI prediction into text generation. Multi-task prediction enhances the predictive performance of DTI. Attached Figure Description

[0080] 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, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0081] Figure 1 This is a flowchart illustrating a multi-task prediction method for drug target interaction based on multimodal features in an embodiment of the present invention.

[0082] Figure 2 This is a schematic diagram illustrating the steps of a multi-task prediction method for drug target interaction based on multimodal features in an embodiment of the present invention.

[0083] Figure 3 This is a schematic diagram of the structure of the DTI multi-class prediction model MFFGCN in an embodiment of the present invention;

[0084] Figure 4 This is a schematic diagram of the structure of the EXDTI (Explained Natural Language) text generation model in this embodiment of the invention. Detailed Implementation

[0085] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. However, the scope of protection of this invention is not limited to the following embodiments.

[0086] like Figure 2 As shown, this invention provides a multi-task prediction method for drug target interactions based on multimodal features, comprising the following steps:

[0087] S1: DTI Data Preprocessing. To implement a cold start scenario, the two collected DTI datasets are divided into training and testing sets according to a set ratio; and feature and textual semantic data information are obtained from databases such as DrugBank, PubChem, and UniPort3 based on the drug target node information in the DTI datasets.

[0088] S2: Model Building. This involves building the initial DTI multi-class prediction model MFFGCN, the DTI natural language interpretation text generation model EXDTI, and models for extracting multimodal features of drugs and targets, including VideoMol and BGE.

[0089] S3: Model Training. The two initial DTI prediction models are trained using the training set to obtain the trained DTI prediction model.

[0090] S4: DTI Multi-Task Prediction. The trained DTI prediction model is tested to obtain the DTI prediction results for the test set and output the corresponding evaluation metrics.

[0091] In this embodiment, step S1 includes the following sub-steps:

[0092] S5: Construct a cold start scenario by dividing the two DTI datasets into training and testing sets in a ratio of 0.8:0.2. The DTI datasets include information such as the number of drug target nodes, DrugBankIDs, and the number of their respective DTIs. Dataset 1 contains 2183 drugs, 1296 targets, and 11976 DTIs. Dataset 2 contains 6050 drugs, 3675 targets, and 25580 DTIs.

[0093] Table 1 shows the number of data of each type in the dataset.

[0094] .

[0095] S6: Based on the drug DrugBankID, obtain complete drug information from DrugBank, including drug name, SMILES structural sequence and semantic text description, etc. The semantic text description includes information such as therapeutic use, pharmacological mechanism, pharmacokinetic properties and physicochemical properties.

[0096] S7: Based on the target DrugBankID, obtain complete target information from DrugBank, including the target SwissID, name, and SEQUENCE amino acid sequence, and obtain the corresponding semantic text description from UniProt3.

[0097] S8: Based on the drug DrugBankID, complete 2D features of the drug, namely chemical structure, enzyme protein, etc., were extracted from databases such as Drugbank and Puchem, which cover information such as chemical structure and biological characteristics.

[0098] S9: Map the target SwissID obtained from DrugBank to the String database, and obtain the target-target interactions based on the mapping results. These interactions are adjacency matrix features and serve as an embedding feature of the target. .

[0099] In this embodiment, the specific processing steps for the models VideoMol and BGE, which extract multimodal features of the drug and target, are as follows:

[0100] S10: Employs a pre-trained video model, VideoMol, to extract effective 3D structural features from the three-dimensional conformation of the drug.

[0101] S11: Relevant textual descriptions of drugs and targets extracted from Drugbank and UniProt3 are input into the pre-trained language model BGE for high-dimensional semantic space mapping of natural text language. By directly applying this model, LLM-derived semantic features can be extracted from the natural textual descriptions of drugs and proteins.

[0102] like Figure 3 As shown, further, in step S10, the specific processing procedure for acquiring the 3D features of the drug is as follows:

[0103] S12: Specifically, firstly, a three-dimensional conformation of the drug molecule is generated, and then it is rotated around the coordinate axes by a certain angle. Rotation angle ,in This indicates the number of samples per rotation axis. Render 20 frames per rotation axis. Molecular video, each frame of which is represented as ( Subsequently, a frame of image is randomly sampled from the molecular video corresponding to each rotation axis as the corresponding view. For each view, VideoMol encodes it into a view representation vector, as follows: .

[0104] S13: The three-view representation vector matrix of all drugs is represented as follows Its splicing serves as a 3D feature of the drug. Finally, the drug node partial embedding will be initialized as a concatenation of 2D and 3D features, as follows: .

[0105] S14: Using the language model BGE, effective derived semantic initialization embedding features, i.e. LLM features, are extracted from the textual descriptions of drugs and targets.

[0106] like Figure 3 As shown, further, in step S12, the specific processing procedure for obtaining the LLM features of the drug and target is as follows:

[0107] S15: Specifically, for the drug component, a JSON file is first generated based on the text description obtained from DrugBank. This JSON file is then input into the BGE language model for high-dimensional semantic space mapping of the natural text language. By directly applying this model, embeddings can be extracted from the natural text language description of the drug without any additional parameter updates.

[0108] S16: For the target part, firstly, a JSON file is generated based on the text description obtained from UniPort3. Then, the JSON file is input into the language model BGE to perform high-dimensional semantic space mapping of natural text language and extract the target LLM embedding features.

[0109] S17: For a given text description of a drug and target And the pre-trained language model BGE, the corresponding semantic calculation formula is: .in and These represent the LLM characteristics of the drug and the target, respectively.

[0110] S18: This allows us to obtain the LLM-derived semantic initialization feature embeddings for the drug and target. .

[0111] S19: Obtain the final embedding of drug and target nodes through concatenation operation. The final embedding of drug and target node features: and .

[0112] In this embodiment, the construction process of the DTI multi-class prediction model MFFGCN is as follows:

[0113] S20: Initialize the DTI graph neural convolutional network with the drug target node features and DTI input from the training set.

[0114] S21: The input DTI is used as an edge in the graph convolutional neural network, thus modeling the DTI network as a DTI symbolic heterogeneous graph. We constructed the DTI multi-class prediction model MFFGCN.

[0115] S22: where It is by The nodes composed of various drugs Represents drug nodes; It is by Nodes composed of various targets Represents the target node. Edge set Indicates drug With target Edge links. Indicates from the first The drug to the first The edge of a target, i.e., the drug For the target The impact.

[0116] S23: In addition, all drug node feature matrices Each line represents Drug properties include characteristics such as chemical structure, target protein, and enzyme protein.

[0117] S24: Divide the DTI symbolic heterogeneous graph into two symbolic heterogeneous subgraphs, one positive and one negative.

[0118] S25: Further, in step S24, the DTI symbol heterogeneous graph is innovatively divided into two symbol heterogeneous subgraphs, thereby better representing the enhancing and inhibiting effects of DTI, i.e., the interaction between positive and negative DTI. The positive and negative symbol heterogeneous subgraphs are respectively represented as follows: .

[0119] S26: Define the magnetic map convolution kernel MKConv for the DTI multi-class prediction model MFFGCN. The process is as follows:

[0120] S27: Further, in step S25, let and These represent the adjacency matrices corresponding to the heteroprotic subgraphs of the positive and negative DTI symbols, respectively. ,otherwise .

[0121] S28: Further, in step S27, the design of MKConv allows the model to adaptively capture heterogeneous structure information in a local range, while retaining the advantages of continuity and boundary modeling in frequency domain filtering.

[0122] S29: Normalize the adjacency matrix and then perform graph filtering to obtain the normalized graph convolution matrix on the heteroprotic graph with positive and negative signs.

[0123] S30: Furthermore, in step S28, by using a given low-to-high frequency response function : and This allows us to obtain the low-to-high frequency graph filtering matrix on the positive and negative sign heteroproton graph. .

[0124] S31: The embedding representation of the drug target node in the... Layer is denoted as It is made of and It consists of two parts, respectively, through the symbolic heteroproton diagram. and The spectrogram is obtained by performing a convolution process on it.

[0125] S32: Furthermore, the node form of the graph convolution process is represented as follows: , , , .

[0126] S33: Among them There are two learnable parameters, corresponding to the node itself and its neighboring nodes, respectively. A concatenation operation is used to integrate these four parts, and a hierarchical nonlinear activation function is applied. In the experiment, the specific activation function was: The final node is represented as .

[0127] S34: From a spectral analysis perspective, the low-frequency and high-frequency response functions achieve feature decomposition for different frequency components in the frequency domain; from a node message passing perspective, the dual-frequency channels, combined with a relationship-aware aggregation mechanism, enable nodes to obtain rich information from both local consistency and local differences during feature updates. Overall, MKConv achieves modeling of the DTI's operational mode.

[0128] S35: Embed the drug target nodes in the test set into the trained DTI multi-class prediction model MFFGCN, and obtain the DTI multi-class prediction results.

[0129] In this embodiment, both the drug and target multimodal feature models, VideoModel and BGE, are pre-trained models that obtain feature embeddings by inputting different modal information. The VideoModel model takes the 3D structural information of the drug molecule as input and decomposes it into 60-frame-rate images using a video encoder for feature extraction. BGE, on the other hand, is a language model that takes the textual semantic information of the drug target as input and extracts useful information through textual context cues, thereby extracting effective embedding features. X and Y represent the feature embeddings at the drug and target nodes, respectively. The final node feature embeddings are obtained by concatenation, and the concatenation method is as follows:

[0130] and .

[0131] In this embodiment, the DTI prediction model MFFGCN based on graph convolutional neural networks uses the input DTI as edges in the graph convolutional neural network, thereby modeling the DTI network as a DTI symbol heterogeneous graph. This innovative approach divides the DTI symbolic heteroproton graph into two symbolic heteroproton graphs, thereby better representing the enhancing and inhibiting mechanisms of DTI, i.e., the interaction between positive and negative DTI. The positive and negative symbolic heteroproton graphs are represented as follows: The low-frequency and high-frequency response functions in MKConv achieve feature decomposition for different frequency components in the frequency domain. Furthermore, the dual-frequency channels, combined with a relationship-aware aggregation mechanism, enable nodes to obtain rich information from both local consistency and local variability during feature updates. Finally, drug target nodes from the test set are embedded into the pre-trained DTI multi-class prediction model MFFGCN to obtain DTI multi-class prediction results.

[0132] In this embodiment, Table 2 shows the ablation experiment of the DTI multi-class prediction model MFFGCN, aiming to examine the impact of adjusting different variants on the experimental results. All variant models are shown as the average of 5-fold, 5-times cross-validation. In the table, 2D, 2D_3D, 2D_LLM, and 2D_3D_LLM represent variant models with different modal feature inputs, while High and Low represent variant models with high and low frequency filters, respectively.

[0133] Table 2 Comparison of ablation experimental results in the DTI multi-class prediction model MFFGCN.

[0134] .

[0135] Table 2 shows that the variant model with 2D_3D_LLM input modal features achieves the best results in the ablation experiments. Compared to other variant models, the variant model with 2D_3D_LLM input modal features shows significant improvements in AUC, F1, and ACC values ​​on both datasets, indicating the effectiveness of multimodal features in DTI multi-class prediction tasks. Furthermore, the experimental results of the variant model with high- and low-frequency filters demonstrate that different filters have a certain impact on DTI multi-class prediction.

[0136] In this embodiment, Table 3 shows the comparison results of the two datasets on different graph neural network models, including mainstream graph neural network models for the DTI multi-class prediction task, such as SHGNN, MFSGNN, MFSHGCN, SBGNN, and SGRL-DTI. All models are shown as the average of 5-fold, 5-times cross-validation.

[0137] Table 3 Comparison of experimental results for Dataset1 and Dataset2 on different models.

[0138] .

[0139] Table 3 shows that the DTI multi-class prediction model MFFGCN in this invention performs best on both datasets. Compared to the suboptimal model MFSHGCN, MFFGCN improves the AUC by 0.63% and 0.52% on the two datasets, and the F1 score by 0.76% and 1.3%, respectively. This demonstrates the high efficiency of the DTI multi-class prediction model MFFGCN in the DTI multi-class prediction task.

[0140] Furthermore, in step S2, the EXDTI (Extreme Derivative Language Interpretation) text generation model construction process is as follows:

[0141] S36: Collect paradigm data and one-dimensional features describing drug-target interaction mechanisms from the training and testing sets in DrugBank, namely, "drug A as an inhibitor binds to the active site of target B..." and drug SMILES and target SEQUENCE.

[0142] S37: Constructing the DTI Natural Language Interpretation Text Generation Model EXDTI: Using the molecular-text pre-trained model MolT5 and the protein language model ProtBERT to encode one-dimensional features, namely drug SMILES and target SEQUENCE, to achieve the model's mapping from "SMILES + SEQUENCE" to "natural language interpretation". For example... Figure 4 As shown.

[0143] S38: Use the molecular-text pre-trained model MolT5 and the protein language pre-trained model ProtBERT to encode drug SMILES and target SEQUENCE.

[0144] S39: In the model configuration, the maximum target sequence length `max_target_length` is explicitly specified, which is 128 tokens. This parameter defines the maximum capacity of the generated explanatory text, ensuring that the generated explanation is detailed and non-divergent.

[0145] S40: Invoke the pre-trained molecular-text model MolT5 to translate the chemical structures SMILES of the drug into natural language descriptions (Captions). This description includes the functional group characteristics and physicochemical properties of the molecule, providing "semantic knowledge" rather than just "structural symbols" for interpretation.

[0146] S41: Similarly, the pre-trained protein language model ProtBERT is invoked to translate the target sequence features (SEQUENCE) into a natural language description (Caption).

[0147] S42: Assemble the various information of the above drugs and targets into a source sequence according to a specific template: drug[SMILES]([drug_Caption]); target[SEQUENCE]([target_Caption]).

[0148] S43: In step S42, the source sequence explicitly distinguishes two entities, including the drug and the target. This forces the model to simultaneously consider the semantic information of both entities during the cross-attention calculation phase, serving as a logical premise for generating the interpretation.

[0149] S44: The target sequence of the training data is formatted as follows: <label>Explanation <exp>.

[0150] S45: Here, Label is the classification label, such as "positive" or "negative," and the output is 1 or -1. Explanation: A Trigger Token, as a key trigger word, is explicitly embedded in the target sequence. This acts as a "hint" during model inference, informing the decoder that it must generate a specific mechanism explanation immediately after outputting the interaction state. <exp>Provide a realistic description of the mechanism, such as: "Drug A acts as an inhibitor, binding to the active site of target B."

[0151] S46: Convert the above text into corresponding Token IDs. Replace all IDs at fill positions with -100 to ensure that these positions are ignored when calculating the loss function, allowing the model to focus on generating valid text.

[0152] S47: During the inference phase, the DTI natural language text interpretation generation model EXDTI generates interpretive text autoregressively based on the input drug-target pair.

[0153] S48: Loss function, loss function generated using only text .

[0154] S49: Where N represents the size of the training set, T represents the length of the target sequence, and Θ represents the parameters of the encoder and decoder.

[0155] S50: Further, in step S47, indicators such as BLEU are introduced to quantify the accuracy of the generated explanation.

[0156] S51: Train the initial model using the source sequences of drug targets in the training set to obtain the trained EXDTI model.

[0157] S52: Input the drug-target pairs from the test set into the trained DTI natural language interpretation text generation model EXDTI to generate natural language interpretation text.

[0158] S53: Further, in step S52, an end-to-end EXDTI model is used. As follows: Input: Drug <smiles>Target <sequence>Output: <label>Explanation: <exp>.

[0159] Table 4 shows the results of generating natural language explanation text by inputting the drug-target pairs in the test set into the trained DTI natural language explanation text generation model ExDTI.

[0160] Table 4. DTI Natural Language Interpretation Text Generation Results.

[0161] .

[0162] Table 4 provides a clear view of the detailed text generation results of the DTI natural language interpretation text generation model. This transforms DTI predictions into text form, and the rich and detailed explanations enhance the understanding of DTI, thereby improving the accuracy and generalization of DTI predictions.

[0163] In this embodiment, the natural language explanation text generation model employs the molecular-text pre-trained model MolT5 and the protein language pre-trained model ProtBERT to map drug molecules and target inputs to explanation text. The model uses only cross-entropy as the text generation loss function to train the explanation text. Inputs are drug SMILES and target SEQUENCE, and outputs predicted labels plus natural language explanations.

[0164] In this embodiment, an innovative DTI (Drug-Target Interpretation) natural language interpretation text generation model, EXDTI, is constructed. During the inference phase, the EXDTI model generates interpretation text autoregressively based on the input drug-target pair.

[0165] S54: The final prediction results of drug target interaction are obtained based on the DTI multi-class prediction model MFFGCN and the DTI natural language interpretation text generation model EXDTI.

[0166] S55: Further, in step S54, the DTI multi-class prediction results are matched and mapped with the DTI natural language interpretation text generation prediction results to finally output the result.

[0167] S56: If the matches are the same, output the prediction results of both tasks directly; otherwise, output the prediction result of the task with higher accuracy.

[0168] In this embodiment, the prediction results of drug target interactions are first obtained using the DTI multi-class prediction model MFFGCN and the DTI natural language interpretation text generation model EXDTI, respectively. Then, the prediction results output by the DTI multi-class prediction model are mapped and matched with the corresponding prediction results from the DTI natural language interpretation text generation model EXDTI. When the two results are consistent, the predicted label and its corresponding natural language interpretation text are directly output; if they are inconsistent, the result with higher accuracy is selected as the final conclusion, and the other result is forcibly calibrated to ensure consistency with the final category, thereby ensuring the consistency and reliability of the output DTI multi-class prediction results and DTI natural language interpretation text.

[0169] In summary, the multi-modal feature-based multi-task prediction method for drug target interactions described in the above embodiments introduces a graph neural network framework to unify and integrate multi-dimensional features such as the two-dimensional chemical structure features and three-dimensional conformation features of drugs, as well as semantic features derived from the large language model of drug targets. This effectively supplements key structural and semantic information that is difficult to represent with a single modality. A multi-kernel magnetic relation convolutional structure combining spectral and spatial domains is proposed, enabling the model to simultaneously capture global topological dependencies and local interaction patterns in the drug target network. Furthermore, a multi-task prediction mechanism combining DTI multi-classification prediction and DTI natural language interpretation text generation is constructed, allowing the model to not only provide interaction category judgments but also interpretable linguistic descriptions. Simultaneously, this invention maintains high prediction robustness and generalization ability even in cold-start scenarios. Extensive experimental results demonstrate that this invention achieves optimal performance across multiple evaluation metrics, validating the effectiveness and application value of the proposed method in drug discovery and target identification tasks.

[0170] While the foregoing description illustrates specific embodiments of the present invention, it should be understood that these embodiments are merely illustrative and not intended to limit the scope of the invention. Those skilled in the art can make adjustments, improvements, or equivalent substitutions according to actual needs without departing from the core ideas and scope of the invention, and all such adjustments should be considered to fall within the scope of the invention.< / exp> < / label> < / sequence> < / smiles> < / exp> < / exp> < / label> < / exp> < / label> < / sequence> < / smiles> < / label> < / exp> < / label>

Claims

1. A method for multi-task prediction of drug target interaction based on multi-modal features, characterized in that, Includes the following steps: S1: DTI data preprocessing. To implement the cold start scenario, the two collected DTI datasets are divided into training set and test set according to a set ratio. Based on the drug target node information in the DTI dataset, feature and text semantic data information are obtained from databases such as DrugBank, PubChem, and UniPort3. S2: Model building, building the initial DTI multi-class prediction model MFFGCN, the DTI natural language interpretation text generation model EXDTI, and models for extracting multimodal features of drugs and targets, including VideoMol and BGE; The construction process of the DTI multi-class prediction model MFFGCN is as follows: S20: Initialize the DTI graph neural convolutional network with the drug target node features and DTI input from the training set; S21: The input DTI is used as an edge in the graph convolutional neural network, thus modeling the DTI network as a DTI symbolic heterogeneous graph and constructing the DTI multi-class prediction model MFFGCN; S24: Divide the DTI symbol heterogeneous graph into two symbol heterogeneous subgraphs, one positive and one negative; S26: Define the magnetic graph convolution kernel MKConv for the DTI multi-class prediction model MFFGCN; The construction process of the EXDTI (Explained Natural Language Transmission) text generation model is as follows: S36: Collect paradigm data and one-dimensional features describing drug-target interaction mechanisms from the training and test sets in DrugBank, namely drug SMILES and target SEQUENCE. S37: Construct the DTI natural language interpretation text generation model EXDTI, using the molecular-text pre-trained model MolT5 and the protein language model ProtBERT to encode one-dimensional features, namely drug SMILES and target SEQUENCE, to realize the model's mapping from "SMILES + SEQUENCE" to "natural language interpretation"; S3: Model training. The training set is used to train the two initial DTI prediction models to obtain the trained DTI prediction model. S4: DTI multi-task prediction, testing on the trained DTI prediction model, obtaining the DTI prediction results of the test set and outputting the corresponding evaluation metrics.

2. The multi-task prediction method of drug target interactions based on multi-modal features according to claim 1, characterized in that, In S1, the specific processing procedure for DTI data preprocessing is as follows: S5: Construct a cold start scenario by dividing the two DTI datasets into a training set and a test set in a ratio of 0.8:0.

2. The DTI dataset includes information such as the number of drug target nodes, DrugBankID, and the number of its DTIs. S6: Based on the DrugBankID, obtain complete drug information from DrugBank, including drug name, SMILES structure sequence and semantic text description, etc. S7: Based on the target DrugBankID, obtain complete target information from DrugBank, including the target SwissID, name, and SEQUENCE amino acid sequence, and obtain the corresponding semantic text description in UniProt3; S8: Based on the drug DrugBankID, complete 2D features of the drug, including information such as chemical structure and biological characteristics, were extracted from databases such as Drugbank and Puchem. S9: Map the target SwissID obtained from DrugBank to the String database, and obtain the target-target interaction based on the mapping result. This is the adjacency matrix feature, which serves as an embedding feature of the target.

3. The method of claim 2, wherein the method is characterized by, In S2, the specific processing steps for the models VideoMol and BGE, which extract multimodal features of the drug and target, are as follows: S10: Employs a pre-trained video model, VideoMol, to extract effective 3D structural features from the three-dimensional conformation of the drug. S11: Input the relevant text descriptions of drugs and targets extracted from Drugbank and UniProt3 into the pre-trained language model BGE to perform high-dimensional semantic space mapping of natural text language; by directly applying this model, LLM derived semantic features can be extracted from the natural text language descriptions of drugs and proteins.

4. The multi-task prediction method of drug target interactions based on multi-modal features according to claim 1, characterized in that, In S4, the specific processing procedure is as follows: S54: The final prediction results of drug target interaction are obtained based on the DTI multi-class prediction model MFFGCN and the DTI natural language interpretation text generation model EXDTI.

Images