Drug-target interaction prediction method based on multi-modal feature fusion mixed architecture

Abstract

The embodiment of the invention discloses a drug-target interaction prediction method based on a multi-modal feature fusion mixed architecture, relates to the technical field of biomedical prediction, and aims to improve the prediction efficiency of drug-target interaction of the multi-modal feature fusion mixed architecture. The drug-target interaction prediction method based on the multi-modal feature fusion mixed architecture comprises the following steps: inputting a target drug molecule SMILES and a target protein sequence into an MFF-DTI model, and sequentially performing feature extraction and feature fusion; after a position code is added, the position code is sent to a Transform, and target output is carried out based on a task type; when the task type is a predicted drug-target interaction probability, outputting a probability value; and when the target task is the predicted drug-target interaction type, outputting the type.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biomedical prediction, in particular to a drug-target interaction prediction method based on a multi-modal feature fusion hybrid architecture. BACKGROUND

[0002] Current deep learning-based drug-target interaction (DTI) prediction research relies heavily on the feature representation of basic information such as SMILES sequence of drug molecules, drug molecule graph, and protein sequence. However, it fails to fully integrate multi-level features of drugs and target proteins, such as chemical properties, spatial configuration, biological activity of drugs, and functional domain distribution of target proteins, and other key information, which has the limitation of insufficient feature coverage.

[0003] At the same time, most existing DTI prediction models can only predict the probability or affinity of drug-target interaction, and it is difficult to further realize accurate determination of the interaction type. In fact, if the recognition of drug-target interaction type can be completed simultaneously in the prediction process, it will effectively screen out unexpected interaction types in the early stage of drug development, thereby avoiding the waste of research resources. SUMMARY

[0004] The main purpose of the present application is to provide a drug-target interaction prediction method based on a multi-modal feature fusion hybrid architecture, aiming to improve the prediction efficiency of the multi-modal feature fusion hybrid architecture of drug-target interaction.

[0005] To achieve the above purpose, the technical solutions adopted by the embodiments of the present application are as follows: In a first aspect, the embodiments of the present application provide a drug-target interaction prediction method based on a multi-modal feature fusion hybrid architecture, comprising the following steps: Input the target drug molecule SMILES and the target protein sequence into the MFF-DTI model, and sequentially perform feature extraction and feature fusion; and after adding position coding, input into the Transformer, and based on the task type, perform target output; When the task type is to predict the probability of drug-target interaction, output the probability value; When the target task is to predict the type of drug-target interaction, output the type.

[0006] As some optional embodiments of the present application, the input of the target drug molecule SMILES and the target protein sequence into the MFF-DTI model, and the sequential feature extraction and feature fusion, comprise: Input the target drug molecule SMILES and the target protein sequence into the MFF-DTI model to perform feature extraction; After the extraction is completed, the SMILES sequence, protein sequence, molecular graph structure, molecular fingerprint, and physicochemical property characteristics of the protein and the protein PDB file are converted into a unified embedding space through respective embedding layers or processing modules to align them in dimension.

[0007] As some optional embodiments of the present application, the input of the target drug molecule SMILES and the target protein sequence into the MFF-DTI model for feature extraction includes: generating a molecular graph feature based on the target drug molecule SMILES, and obtaining atom features and edge features based on the molecular graph feature; based on the target protein sequence, sequentially performing basic sequence encoding, physicochemical property supplementation, multi-scale feature enhancement, and global dependence modeling, and based on the protein PDB file, extracting geometric features, physicochemical features, interaction features, and dynamic features, and then performing standardization processing, and converting into a vector representation compatible with sequence features through a learnable embedding layer.

[0008] As some optional embodiments of the present application, the geometric features are obtained by calculating the spatial properties of the protein binding site; they include: the convex hull volume for reflecting the size of the binding cavity, the surface area to volume ratio for describing the shape regularity, the radius of gyration Rg for measuring the structural compactness, and the moment of inertia for characterizing the mass distribution.

[0009] As some optional embodiments of the present application, the physicochemical features are obtained by analyzing the amino acid composition and properties; they include: average hydrophobicity based on the Kyte-Doolittle scale, electrostatic potential distribution, isoelectric point prediction, and polarity / non-polarity residue ratio.

[0010] As some optional embodiments of the present application, the interaction features include hydrogen bond networks and active site distributions. The hydrogen bond network refers to the interaction of N / O-H within a distance of 3.5 Å with an included angle <20°; the active site distribution refers to the residues within a range of 5 Å from the ligand.

[0011] As some optional embodiments of the present application, the dynamic features are obtained by deriving parameters through molecular dynamics simulation; they include structural deviation-RMSD, residue fluctuation-RMSF, and gyration radius change-Rg.

[0012] As some optional embodiments of the present application, the structural deviation, the residue fluctuation, and the gyration radius change satisfy the following relationships, respectively: where RMSD is the root mean square deviation, N is the number of atoms compared, and are the coordinates of the ith point in the two sets of data, respectively; RMSF is the root mean square fluctuation, where N is the number of sampling points; Rg is the gyration radius, where N is the total number of particles, is the position of the molecule at time N is the number of sampling points; Rg is the gyration radius, where N is the total number of particles, is the position vector of the ith particle, is the center of mass position vector of the system, is the square of the distance of the particle to the center of mass; As some optional embodiments of the present application, the SMILES sequence, protein sequence, molecular graph structure, molecular fingerprint, and physicochemical property characteristics of the protein PDB file are converted into a unified embedding space through respective embedding layers or processing modules, including: the SMILES and protein sequence are processed through an embedding layer and an LSTM module; the molecular graph structure is extracted through GAT to obtain a global graph embedding; the molecular fingerprint and the physicochemical property characteristics of the protein are mapped to the same embedding dimension through a linear layer; and it is determined whether to add the protein PDB file as a feature according to the task type. When the task is to predict the interaction type, the protein PDB file is added as a feature. The protein PDB file first extracts 15 basic features per residue, and then supplements 8 feature interaction items to obtain a 23-dimensional vector. Subsequently, it is linearly mapped to the same embedding dimension as the SMILES and protein sequence.

[0013] As some optional embodiments of the present application, when the task type is to predict the drug-target interaction probability, a probability value is output, including: When the task type is to predict the drug-target interaction probability, a multi-head attention mechanism is used to calculate the cross-attention score between the drug features and the protein features, so that the model can automatically focus on the most relevant feature combination for the prediction task; If there is a relevant molecular dynamics trajectory file, molecular dynamics features are extracted as auxiliary features, which are processed through an independent embedding layer and then interact with the main feature stream through attention to enhance the physical and chemical perception ability of the model; If there is no relevant molecular dynamics trajectory file, the default value 0 is returned.

[0014] Compared with the prior art, the present application provides a multimodal feature fusion hybrid architecture, a drug-target interaction prediction network (MFF-DTI) for multimodal feature fusion, which can effectively integrate various modal data related to drug-target interactions. The model described in the present application is based on Transformer, uses SMILES and protein sequence as input, and through automatic calculation of drug and protein features, can finally predict and output the interaction probability and type of drug-target, so as to improve the accuracy and efficiency of drug-target interaction (DTI) prediction. In addition, based on the existing drug-target data, more comprehensive drug-target interaction features are proposed to further enhance the prediction ability of the model for unknown DTI, and the effectiveness of MFF-DTI for new DTI discovery is verified. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 Integration diagram of disclosed data sources related to embodiments of the present application; Figure 2 Multimodal feature fusion hybrid architecture related to embodiments of the present application; Figure 3 Drug-target interaction prediction method flowchart of the multimodal feature fusion hybrid architecture related to embodiments of the present application. DETAILED DESCRIPTION

[0016] It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0017] Drug-target interaction (DTI) refers to the interaction between small molecule drugs and biological macromolecular targets, which is the basis for drugs to produce biological effects. As a core link of drug research and development, DTI is of great significance for adverse reaction prediction, drug mechanism analysis, and protein function exploration. Traditional DTI verification methods mainly rely on wet experiments (such as biochemical experiments, cell experiments, etc.), which not only consume a lot of time, but also have high cost.

[0018] In recent years, DTI prediction methods based on deep learning have made significant progress. The sequence feature-oriented method DeepDTA proposed by Öztürk et al. extracts the sequence features of compounds and proteins through a CNN module, respectively, and finally outputs the prediction result of the binding affinity. However, in DTI research, the expression of a single feature is difficult to meet the demand for the richness of compound and protein features. Therefore, DeepConv-DTI proposed by Lee et al. takes ECFPs fingerprint as the feature vector of the compound, and at the same time extracts the feature vector of the protein sequence through CNN, so as to introduce more rich compound and protein feature representation in modeling.

[0019] However, the above methods still have limitations: the existing methods mostly rely on the feature representation of molecular structure SMILES sequence, molecular graph, molecular fingerprint, etc., and the coverage of the features is insufficient, because they fail to fully utilize the multi-level information of drugs and target proteins (such as chemical properties, protein functional domain distribution, etc.); in addition, most of the current DTI prediction models can only predict the probability or affinity of interaction, and it is difficult to accurately predict the specific interaction type (such as agonist, inhibitor, blocker, etc.). Knowing only the interaction probability is far from enough - different action types have a significant impact on the efficacy and safety of drugs, and if the drug-target action type that may cause serious side effects can be identified in the early screening stage, the related drugs can be eliminated in advance, so as to avoid the waste of resources in subsequent expensive experiments.

[0020] Therefore, it is urgent to develop more effective feature analysis methods and more efficient models to further improve the performance of DTI prediction.

[0021] Based on this, the present application constructs a DTI prediction model as follows Figure 2The multi-modal feature fusion hybrid architecture is shown. That is, for the current majority DTI prediction model, it relies on single-scale features such as SMILES sequence, molecular graph, molecular fingerprint, etc., which leads to the problem of insufficient feature coverage. The model adopts a design scheme of cooperation between the Transformer framework and multi-functional modules, integrates the Graph Attention Network (GAT) module, the Long Short-Term Memory (LSTM) module, the molecular feature extraction module, the protein feature extraction module, the protein PDB file extraction module, the MultiscaleConv1D module, the Positional module, and the Multilayer Perceptron (MLP) module, etc., extracts features from known drug-target interaction data, realizes the prediction of drug-target potential interaction probability and type, and names the model as Multimodal Features Fusion for Drug-Target Interactions (MFF-DTI).

[0022] The MFF-DTI model can effectively solve the problems of insufficient feature extraction and difficulty in accurately predicting the interaction type of the current DTI prediction model. The core is to integrate the multi-modal features of drug-target, including SMILES sequence, protein sequence, molecular graph structure, molecular fingerprint feature, protein physicochemical feature, and protein spatial structure information, etc.: the GAT extracts the molecular graph feature of the drug, the LSTM extracts the SMILES and protein sequence features, the RDkit extracts the molecular fingerprint information from the SMILES structural formula, and the hydrophobicity and isoelectric point physicochemical features are extracted from the protein sequence. Compared with the single-scale features relied on by most existing DTI prediction models, the feature extraction of MFF-DTI is more comprehensive; in addition, the model additionally introduces spatial structure information features extracted from the protein PDB file, which has a significant auxiliary role in predicting the type of drug-target interaction, can provide more refined insights into the nature of the interaction, and constitutes a significant breakthrough compared with existing DTI prediction models.

[0023] In the feature processing process, feature fusion is realized through feature splicing and MLP, the fused features are input into the Transformer encoder for global modeling to capture long-distance dependencies, and finally the prediction results of interaction probability and type are output.

[0024] Two public datasets and multiple public databases were selected as data sources to ensure the generality and reliability of the source data, including the Human dataset, the C. elegans dataset, the DrugBank database, the Uniprot database, the chEMBL database, and the AlphaFold Protein Structure Database. The integration of the public data sources is shown in FIG. 1. Figure 1

[0025] Among the selected public data sources, the Human dataset and the C. elegans dataset are commonly used standard datasets, widely used to evaluate the generalization ability and cross-species applicability of the model. The Human dataset is derived from known interactions between human protein targets and drugs, containing a large amount of high-quality human target information verified by experiments, and the target types are diverse (such as enzymes, ion channels, nuclear receptors, etc.), and the drug categories are rich (covering small molecule drugs and some biological macromolecules). The C. elegans dataset is constructed based on the drug-target interaction data of the model organism Caenorhabditis elegans. Compared with the Human dataset, this dataset can better test the model's understanding of non-human targets, and is an important basis for evaluating the cross-species generalization ability of the model.

[0026] DrugBank database is a comprehensive bioinformatics and cheminformatics database focusing on collecting detailed information of drugs and targets, maintained by the Wishart research group of the University of Alberta, Canada, covering multi-dimensional data of drugs, pharmacology, interaction, metabolism, and targets. The database integrates experimental data and computational prediction results, and is widely used in drug research and development, precision medicine, and bioinformatics research fields. Its online version (5.1.12, released on March 14, 2024) contains 16,618 drug entries, including 2,780 approved small molecule drugs, 1,626 approved biological agents, 135 health foods, and more than 6,723 experimental drugs; In addition, there are 5,294 non-redundant protein sequences associated with these drug entries, each entry contains more than 200 data fields, covering drug / chemical data and drug targets or protein data.

[0027] ​Uniprot (Universal Protein) is a comprehensive protein database that integrates protein sequences, functional information, and research paper indexes, bringing together resources from three databases: the European Bioinformatics Institute (EBI), the Swiss Bioinformatics Institute (SIB), and the Protein Information Resource (PIR). EBI is affiliated with the European Molecular Biology Laboratory (EMBL), a flagship research institution in the field of European life sciences, located in the Wellcome Genome Campus in Hinxton, Cambridge, UK, which is one of the core areas for genomic research worldwide. SIB is located in Geneva, Switzerland, and is responsible for maintaining the ExPASy (Expert Protein Analysis System) server, which is a core resource platform for proteomics tools and databases. PIR was established by the National Biomedical Research Foundation (NBRF) in 1984 to help researchers identify and analyze protein sequence information.

[0028] The ChEMBL database is a manually managed database maintained by the European Bioinformatics Institute (EBI), containing over 2.5 million compounds and over 21 million biological activity data, covering over 16,000 drug targets and over 92,000 scientific papers. As a large open-access drug discovery database, it focuses on collecting and organizing chemical, biological activity, and genomic data of small molecule compounds with drug activity, and is committed to promoting the conversion of genomic information into effective new drugs. The database integrates multiple sources of data such as scientific literature, clinical trial data, and drug approval information, providing reliable data support for drug research and development.

[0029] The AlphaFold Protein Structure Database was released by the DeepMind team of Google and the European Bioinformatics Institute (EMBL-EBI) in July 2021, based on the protein structure database predicted by the artificial intelligence system AlphaFold. The database contains about 350,000 protein structures predicted by AlphaFold, covering humans and 20 commonly used model organisms in biological research (such as E. coli, fruit flies, zebrafish, mice, etc.). In terms of human proteome, artificial intelligence has predicted the structure of 98.5% of human proteins, while in the past decades, scientists have only analyzed 17% of the amino acids in human protein sequences. The release of this database is considered to have the potential to revolutionize the field of life sciences, and Dr. Ewan Birney, director of the European Bioinformatics Institute, calls it one of the most important databases since the release of the human genome map.

[0030] Specifically, the prediction method described in the present application is as shown in Figure 3 , that is: To predict new DTI, MFF-DTI first extracts drug features from its SMILES sequences. The model then converts these features into a molecular graph using the RDKit library and inputs it into a GNN (Graph Neural Network) to parse the SMILES sequences and extract atomic and edge features.

[0031] Atomic characteristics include: atomic type (e.g., carbon, nitrogen, oxygen), atomic degree, hybridization type, aromaticity, number of hydrogen atoms, and chirality. Among these: One-hot encoding is used for atom types, mapping them to 10 classes: Carbon (C), Nitrogen (N), Oxygen (O), Fluorine (F), Phosphorus (P), Sulfur (S), Chlorine (Cl), Bromine (Br), Iodine (I), and Other; Atomic degree refers to the number of direct neighbors of an atom in a molecular diagram. 0 means that the atom is not bonded to any other atom, 1 means that the atom is bonded to only one other atom, and so on, with a maximum of 6. Numbers greater than 6 are treated as 6, because degrees greater than 6 are extremely rare in organic molecules, and this is done to maintain a reasonable range of values. Hybrid types are encoded in six categories: SP, SP2, SP3, SP3D, SP3D2, and Other; Chiral features are encoded in three categories: non-chiral, clockwise (CW), and counterclockwise (CCW). The edge features include: the type of bond and the potential for hydrogen bond formation; that is, single bond, double bond, triple bond, aromatic bond, and others are encoded as bond type, and whether hydrogen bonds are likely to form is encoded as hydrogen bond probability.

[0032] The general formula for using GATConv is as follows: (1) In the formula, It refers to the first l Layer nodes i The input feature vector; It refers to the first l The output feature vector of node i in layer +1; It refers to a learnable linear mapping weight matrix that transfers features from... Dimension mapping to dimension; It is the set of first-order neighbors of node i; It is the attention coefficient.

[0033] It should be noted that the GAT is a kind of GNN, which analyzes the molecular graph structure. First, the smiles structural formula is converted into a molecular graph by using the RDKit tool. This process includes extracting atomic features and edge features. After the GAT receives the molecular graph constructed by RDKit, it is deeply analyzed through a four-layer graph attention mechanism, so that each atom can simultaneously focus on multiple interaction modes of its neighbor atoms, and the molecular embedding vector containing rich chemical information is output through global average pooling.

[0034] where the attention coefficient Calculation: (2) (3) wherein, is the unnormalized attention score between node i and neighbor node j; k is an index variable traversing all neighbors of node i (including i itself) when calculating the denominator of normalization; is the unnormalized attention score between node i and neighbor node k; is the LeakyReLU activation function; is the learnable attention vector; T is the vector transpose symbol; is the linear mapping weight matrix; is the input feature vector of node i at the current layer; is the input feature vector of neighbor node j at the current layer.

[0035] Normalized attention weight: (4) wherein, is the normalized attention weight; is the original attention score between target node i and neighbor node j in the lth layer on the kth attention head; is the loop variable; is the original attention score between target node i and another neighbor node m in the lth layer on the kth attention head.

[0036] Then, the interaction between atoms in the molecule is captured through the multi-head attention mechanism. The formulas of the attention mechanism and the multi-head attention mechanism are as follows: (5) (6) (7) where, denotes the "subject" or "question" that needs attention at the moment; denotes the "identifier" or "summary" that is being queried; denotes the "actual content" or "complete information" that is being queried; denotes the dimension of Q and K vectors; denotes the input word embedding matrix; denotes the output of the 1st, 2nd, …, hth parallel attention head of multi-head attention; h denotes the total number of heads of multi-head attention; denotes the output projection matrix; denotes the output of the i-th attention head; , , denotes the learnable linear mapping matrix of the i-th head.

[0037] Finally, the global representation of the molecule is generated by global average pooling as the overall feature of the drug. In addition, the LSTM module is used to encode the embedded SMILES sequence for the SMILES sequence. At the same time, the model also extracts the Morgan fingerprint and MACCS fingerprint from the SMILES sequence, respectively mapped to a fixed-dimensional vector through the embedding layer, to further enrich the feature representation of the drug.

[0038] For protein sequences, the features of the protein are extracted from its amino acid sequence. First, the model maps each amino acid in the amino acid sequence to a fixed-dimensional vector through the embedding layer. Subsequently, the embedded protein sequence is encoded using the LSTM module, which can capture long-term dependencies in the sequence to generate a contextual representation of the protein sequence. To enhance the robustness of the model, in addition, the model also extracts physicochemical properties (such as amino acid composition, dipeptide composition, hydrophobicity, isoelectric point, etc.) from the protein sequence and maps them to a fixed-dimensional vector through the embedding layer to form a feature representation of the amino acid, further enriching the feature representation of the protein. Next, to further extract the multi-scale features of the protein sequence, the model uses a multi-scale convolution module that combines convolution kernels of different scales and deformable convolution to capture features of different scales and enhance the expression ability of the model. In this study, the convolution kernel size is specified as [3, 5, 7], with smaller convolution kernels such as 3 used to capture local details and larger convolution kernels such as 7 used to capture more global features. The formula for deformable convolution is as follows: Offset generation: (8) where, is a learnable offset vector of the nth convolution position; n refers to the serial number of the offset vector; is a learnable weight matrix; is a matrix multiplication; is an input feature map.

[0039] Feature extraction: (9) In the formula, is the final feature value at position p on the output feature map; is the total number of sampling points of the convolution kernel; is the weight coefficient corresponding to the kth sampling point; is an input feature map; is a two-dimensional (or one-dimensional) coordinate position on the output feature map; is the standard deviation of the kth sampling point of the convolution kernel relative to the center; is a learnable offset of the kth sampling point, which causes the sampling position to deform.

[0040] At the same time, the model uses a multi-head self-attention mechanism to model the global dependency of the protein sequence embedding. The attention mechanism can capture the interaction between different positions in the sequence and generate a global context representation (i.e., the embedding vector of each position + information about surrounding positions, for example, the protein sequence is MKVLPA, and before the multi-head attention, it only contains its own information. After multi-head attention, such as V, it not only contains its own information, but also contains information about surrounding MKL, etc.). In order to ensure that the model can capture the sequential relationship in the sequence, the model also introduces position encoding to add position information to the sequence features, so that the model can know which position features are more important. The position encoding formula is as follows: (10) (11) In the formula, is the position encoding value at position pos and embedding dimension index 2i; is the position serial number of token in the sequence; is the even dimension index of the position encoding vector; is the model (embedding) dimension, i.e., the total dimension of the position encoding vector; is the index variable of the dimension grouping.

[0041] The purpose of the above formula (10) and formula (11) is to enable the model to easily learn the relative position information. That means, once the model learns the encoding of a certain position, it can derive the encoding of any other position through a simple linear operation (for example, multiply a rotation matrix), thus easily learning and understanding the concept of "relative position".

[0042] In terms of protein PDB file feature extraction, the model obtains the features of the binding site by parsing the PDB file, and the PDB file contains the three-dimensional structure information of the protein. The features extracted therefrom can be divided into four categories: geometric features, physicochemical features, interaction features, and dynamic features. Geometric features are obtained by calculating the spatial properties of the protein binding site, including convex hull volume (reflecting the size of the binding cavity), surface area to volume ratio (describing shape regularity), radius of gyration Rg (measuring structural compactness), and inertia matrix (characterizing mass distribution). Physicochemical features are achieved by analyzing the composition and properties of amino acids, such as average hydrophobicity based on the Kyte-Doolittle scale, electrostatic potential distribution (extracted from the B-factor field), isoelectric point prediction, and polarity / non-polarity residue ratio. Interaction features focus on detecting hydrogen bond networks (interactions within 3.5 Å distance N / O-H forming an angle <20°) and active site distribution (residues within 5 Å of ligand). Dynamic features derive parameters through molecular dynamics simulation, including RMSD (structure deviation), RMSF (residue volatility), and Rg (change in gyration radius).

[0043] wherein the formulas of RMSD, RMSF and Rg are as follows: wherein, RMSD refers to the root mean square error, N is the number of points (atoms) compared, and are the coordinates of the i-th point in the two groups of data; RMSF refers to the root mean square fluctuation, wherein is the position of the molecule at time , N is the number of sampling points; Rg refers to the change in gyration radius, wherein N is the total number of particles, is the position vector of the i-th particle, is the center of mass position vector of the system, is the square of the distance from the particle to the center of mass; The specially designed 23-dimensional feature vector includes 15 basic features and 8 interaction features, wherein the 15 basic features are average electrostatic potential, average hydrophobicity, number of hydrogen bonds, approximate volume, radius of gyration, number of active site residues, number of hydrophobic patches, maximum moment of inertia, surface area to volume ratio, pocket depth, local electrostatic potential mean, electrostatic gradient, local hydrophobicity mean, average hydrogen bond strength, and hydrogen bond network strength, and the 8 interaction features include polarity-nonpolarity balance, polar environment hydrogen bond, hydrophobic environment hydrogen bond, hydrophobic volume, hydrophobic pocket depth, active site polarity, comprehensive electrostatic-hydrophobic interaction, and polar environment hydrogen bond network. The interaction between features can capture more complex structure-function relationships, such as the interaction between electrostatic and hydrophobicity revealing the synergistic effect of polar and nonpolar regions, and the combination of hydrogen bond and geometric features reflecting the influence of local environment on molecular action. These interaction terms provide higher level of abstract information for the model, thereby improving the accuracy and robustness of predicting drug-target interactions. After standardization, these features are converted into vector representations compatible with sequence features through a learnable embedding layer, providing comprehensive structure-function relationship information for subsequent deep learning models.

[0044] Through the above steps, the MFF-DTI model can extract rich features from SMILES sequences and protein sequences and build a powerful drug-target interaction prediction model.

[0045] After extracting features through the above steps, the following steps are used for feature fusion: In the MFF-DTI model, the feature fusion stage is one of the most important stages in the entire architecture, and its goal is to efficiently integrate molecular and protein features from different modalities to capture cross-domain interaction information. The MFF-DTI model uses a multi-level, multi-dimensional fusion strategy to deeply integrate drug and protein features. First, the model extracts features from multiple input sources, including SMILES sequences, protein sequences, molecular graph structures, molecular fingerprints (Morgan and MACCS), and protein physicochemical property features and protein PDB files. These features are converted to a unified embedding space through their respective embedding layers or processing modules, ensuring that they are aligned in dimension. For example, SMILES and protein sequences are processed through embedding layers and LSTM modules, molecular graph structures are extracted through GAT to obtain global graph embeddings, and molecular fingerprints and protein features are mapped to the same embedding dimension through linear layers. This unified embedding space lays the foundation for subsequent feature fusion. The core step of feature fusion is to concatenate these features from different sources into a comprehensive feature representation. The model concatenates all features along the sequence length dimension to form a complex feature tensor containing multiple types of information. This step ensures that the information from each input source is preserved and integrated into the final feature representation.

[0046] If the task is to predict the type of drug-target interaction, which is a multi-classification task, the model also uses a multi-head attention mechanism to calculate the cross-attention scores between drug features and protein features, allowing the model to automatically focus on the most relevant feature combinations for the prediction task. In this process, if there is a relevant molecular dynamics trajectory file, molecular dynamics features such as RMSD, RMSF, Rg, etc. are extracted as auxiliary features, processed through an independent embedding layer, and then interacted with the main feature stream through attention, further enhancing the model's physical and chemical perception ability. If there is no relevant molecular dynamics trajectory file, the default value is 0.

[0047] Subsequently, the model performs nonlinear transformation and dimensionality reduction on the concatenated features through a multi-layer perceptron (MLP), further integrating the interaction information between different features. The structure of the MLP includes multiple fully connected layers, activation functions, and Dropout layers, enhancing the model's nonlinear expression ability and preventing overfitting. In addition, the model also introduces a weighted fusion mechanism to dynamically adjust the contribution of different features in the final representation by learning their importance weights. The weighted fusion mechanism is as follows: Attention weight generation: , (15) where, is a learnable parameter; K refers to the total number of input features that need to be fused, i.e., the number of weights; is the learnable weight corresponding to the jth input; initialized as an all-one vector, and the weight is obtained by Softmax normalization .

[0048] Weighted fusion: (16) where, is the final fused feature vector after weighted fusion; is the feature vector output by the ith input branch (or the ith feature source).

[0049] To further optimize the effect of feature fusion, the model uses multi-scale convolution and self-attention mechanism. Multi-scale convolution captures local and global feature patterns through different sizes of convolution kernels, while self-attention mechanism dynamically learns the dependency between features, highlighting important features and suppressing noise.

[0050] Finally, the fused features are globally modeled by a Transformer encoder to capture long-range dependencies and output prediction results for different tasks, such as binary or multi-classification. This series of fusion steps ensures that the model can fully utilize multi-source heterogeneous data to improve prediction performance.

[0051] The present application studies the performance of the proposed new model by systematically comparing it with the current ten most advanced DTI prediction methods. The experimental results show that the MFF-DTI model proposed in the present application study has a significant advantage in the binary classification task of the Human dataset and the C. elegans dataset: on the Human dataset, the AUROC reaches 98.78%, the AUPRC is 98.75%, and the F1-score is 95.37%; among them, the AUROC index significantly exceeds the existing methods such as MGNDTI (98.55%), FOTF-CPI (98.34%), and MolTrans (97.99%), indicating that the model can more accurately capture the complex patterns in the DTI data. On the C. elegans dataset, MFF-DTI also performs well, with an AUROC of 99.36% and an accuracy of 97.05%, further verifying its robustness.

[0052] In terms of overall performance, the AUROC of the MFF-DTI model on the Human dataset and the C. elegans dataset reaches 98.84% and 99.35%, respectively, significantly better than the single feature model, fully embodying the advantage of multi-modal feature fusion. Further module contribution analysis shows that the introduction of molecular fingerprints improves the F1-score of the Human dataset from 87.50% to 91.32%; the addition of GAT improves the F1-score of the C. elegans dataset from 91.76% to 93.58%, fully demonstrating the important complementary role of chemical structure features and topological information on model performance. Among them, the LSTM module performs particularly outstanding, helping the AUROC to break through 98% in both datasets, showing its unique advantage in modeling long sequence dependencies.

[0053] In the multi-classification task, MFF-DTI still maintains high competitiveness, with AUROC of 93.49% and F1-score of 60.09%, which is obviously superior to the other three main models (AUROC of MGNDTI is 91.83%, F1-score is 53.16%; AUROC of MolTrans is 89.44%, F1-score is 44.73%; AUROC of FOTF-CPI is 91.33%, F1-score is 54.17%). The core competitiveness of the model comes from its multi-modal feature fusion capability: first, the double sequence Transformer can effectively process SMILES and protein sequences and capture long-range dependencies; second, the graph neural network can encode the molecular topological structure; third, the molecular fingerprint provides complementary chemical features. This design avoids the information loss of single feature methods, making the model have strong generalization ability and potential application value.

[0054] In summary, through careful design of feature engineering and algorithm optimization, the project successfully developed a new model that can efficiently and accurately predict the probability and specific type of drug-target interaction, laying a solid foundation for the development of personalized medicine.

[0055] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A drug-target interaction prediction method based on a multimodal feature fusion hybrid architecture, characterized in that, Includes the following steps: The target drug molecule SMILES and the target protein sequence are input into the MFF-DTI model, and feature extraction and feature fusion are performed sequentially. After adding position encoding, the data is fed into the Transformer, and the target output is determined based on the task type. When the task type is to predict the probability of drug-target interaction, the probability value is output. When the target task is to predict the type of drug-target interaction, the output type is as follows.

2. The drug-target interaction prediction method based on a multimodal feature fusion hybrid architecture according to claim 1, characterized in that, The process of inputting the target drug molecule SMILES and the target protein sequence into the MFF-DTI model, and then performing feature extraction and feature fusion sequentially, includes: The target drug molecule SMILES and the target protein sequence are input into the MFF-DTI model for feature extraction; After extraction, the SMILES sequences, protein sequences, molecular graph structures, molecular fingerprints, physicochemical properties of proteins, and protein PDB files are converted into a unified embedding space through their respective embedding layers or processing modules to align them in dimensions.

3. The drug-target interaction prediction method based on a multimodal feature fusion hybrid architecture according to claim 2, characterized in that, The step of inputting the target drug molecule SMILES and the target protein sequence into the MFF-DTI model for feature extraction includes: Molecular graph features are generated based on the target drug molecule SMILES, and atomic and edge features are obtained based on the molecular graph features. Based on the target protein sequence, basic sequence encoding, physicochemical property supplementation, multi-scale feature enhancement, and global dependency modeling are performed sequentially. After extracting geometric features, physicochemical features, interaction features, and dynamic features from the protein PDB file, the features are standardized and converted into a vector representation compatible with the sequence features through a learnable embedding layer.

4. The drug-target interaction prediction method based on a multimodal feature fusion hybrid architecture according to claim 3, characterized in that, The geometric features are obtained by calculating the spatial properties of protein binding sites; It includes: the convex hull volume to reflect the size of the cavity, the surface area to volume ratio to describe the regularity of the shape, the radius of gyration Rg to measure the compactness of the structure, and the moment of inertia to characterize the mass distribution.

5. The drug-target interaction prediction method based on a multimodal feature fusion hybrid architecture according to claim 3, characterized in that, The physicochemical characteristics were obtained by analyzing the amino acid composition and properties; It includes: average hydrophobicity, electrostatic potential distribution, isoelectric point prediction, and polar / nonpolar residue ratio based on the Kyte-Doolittle scale.

6. The drug-target interaction prediction method based on a multimodal feature fusion hybrid architecture according to claim 3, characterized in that, The interaction characteristics include hydrogen bond networks and active site distribution; The hydrogen bond network refers to the N / OH interactions with an angle of <20° within a distance of 3.5 Å; the active site distribution refers to the residues within a 5 Å range of the ligand.

7. The drug-target interaction prediction method based on a multimodal feature fusion hybrid architecture according to claim 3, characterized in that, The dynamic characteristics are obtained by deriving parameters through molecular dynamics simulations; these include structural deviation - RMSD, residue volatility - RMSF, and cyclotron radius variation - Rg.

8. The drug-target interaction prediction method based on a multimodal feature fusion hybrid architecture according to claim 7, characterized in that, The structural deviation, the residue volatility, and the cyclotron radius change each satisfy the following relationships: In the formula, RMSD refers to the root mean square error, and N is the number of atomic points being compared. and These are the coordinates of the i-th point in the two sets of data; RMSF refers to the root mean square variability, where... For molecules in time The position is N, where N is the number of sampling points; Rg refers to the change in gyro radius, where N is the total number of particles. It is the position vector of the i-th particle. It is the centroid position vector of the system. It is the square of the distance from the particle to its center of mass.

9. The drug-target interaction prediction method based on a multimodal feature fusion hybrid architecture according to claim 2, characterized in that, The process of converting SMILES sequences, protein sequences, molecular map structures, molecular fingerprints, physicochemical properties of proteins, and protein PDB files into a unified embedding space through their respective embedding layers or processing modules includes: SMILES and protein sequences are processed using an embedding layer and an LSTM module; The molecular graph structure is extracted and embedded using GAT to obtain the global graph. Molecular fingerprints and the physicochemical properties of proteins are mapped to the same embedding dimension through a linear layer; Furthermore, the inclusion of protein PDB files as features is determined based on the task type. When the task is to predict interaction types, protein PDB files are included as features. The protein PDB file first extracts 15 basic features residue by residue, and then adds 8 feature interaction terms to obtain a 23-dimensional vector. Subsequently, it is linearly mapped to the same embedding dimension as SMILES and protein sequences.

10. The drug-target interaction prediction method based on a multimodal feature fusion hybrid architecture according to claim 1, characterized in that, When the task type is predicting the probability of drug-target interaction, the output probability value includes: When the task type is to predict the probability of drug-target interaction, a multi-head attention mechanism is used to calculate the cross-attention score between drug features and protein features, so that the model can automatically focus on the feature combination most relevant to the prediction task. If relevant molecular dynamics trajectory files are available, molecular dynamics features will be extracted as auxiliary features. After being processed through an independent embedding layer, these features will interact with the main feature stream through attention to enhance the model's physicochemical perception capabilities. If no relevant molecular dynamics trajectory file is available, the default value of 0 is returned.

Images