Drug target interaction prediction method and system based on multi-modal feature fusion

Through multimodal feature fusion and bidirectional collaborative attention mechanism, the problem of insufficient utilization of single modal features in drug-target interaction prediction is solved, and high-precision prediction of drug-target binding affinity and improvement of the model's generalization ability are achieved.

CN120783848AActive Publication Date: 2025-10-14TAISHAN UNIV

Patent Information

Application Number
CN202510965283.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-14
Publication Date
2025-10-14
Estimated Expiration
2045-07-14

AI Technical Summary

Technical Problem

Existing drug-target interaction prediction methods lack the utilization of single-modal features, have simple multimodal fusion methods, and lack modeling of inter-modal interactions, resulting in limited accuracy of prediction results and model generalization ability.

Method used

A multimodal feature fusion method is adopted to combine the multimodal complementary information of drugs and targets. Through the cross-attention mechanism and bidirectional collaborative attention mechanism, the subsequence features and graph structure features of drugs and targets are integrated to achieve deep interaction modeling.

Benefits of technology

It significantly improves the accuracy of drug-target binding affinity prediction and the generalization performance of the model, enhances the understanding of spatial configuration and functional regions, and accurately captures key atom-residue pairs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120783848A_ABST
    Figure CN120783848A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of bioinformatics and artificial intelligence, and provides a drug-target interaction prediction method and system based on multi-modal feature fusion, and the method comprises the steps: carrying out the word segmentation coding of an obtained to-be-recognized drug sequence and a target sequence, and respectively extracting the subsequence features of a drug and a target; constructing a two-dimensional molecular diagram of the drug and a three-dimensional structure diagram of the target, and respectively extracting diagram structure characteristics of the drug and the target; fusing the subsequence features and graph structure features of the drug and the target through a cross attention mechanism; and carrying out interactive fusion by adopting a bidirectional collaborative attention mechanism to obtain a prediction result of the binding affinity of the drug and the target. According to the invention, by combining the multi-modal complementary information of the drug and the target, deep interaction between different modal features is deeply mined; meanwhile, a two-way collaborative attention mechanism is introduced into interaction modeling of the drug and the target, and the accuracy of drug-target binding affinity prediction is effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics and artificial intelligence, in particular, to a drug target interaction prediction method and system based on multi-modal feature fusion. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] In the field of drug research and development and precision medicine, accurately predicting the interaction between drugs and targets (Drug-Target Interaction, DTI) is of great significance for new drug discovery and drug repositioning. Traditional DTI prediction methods mostly rely on single modal information such as sequences or structures of drugs and protein targets for modeling, which is difficult to fully capture the complex interaction mechanism between the two, limiting the accuracy of the prediction results and the generalization ability of the model. With the development of deep learning technology, integrating multi-modal information and mining potential interaction rules from multiple perspectives has become an important research direction to improve the performance of DTI prediction.

[0004] Existing deep learning-based DTI prediction methods mainly include self-supervised sequence modeling based on Transformer and representation methods for structural information based on graph neural networks. The self-supervised learning strategy based on Transformer structure can effectively learn context-dependent feature representations from sequence data. For example, some research proposes to use the Frequent Consecutive Subsequence (FCS) algorithm to effectively encode drug and target sequences, effectively capturing the potential context-dependent relationships in the sequence and obtaining more expressive embedding features. However, this method mainly focuses on the sequence modality, and the problem is that the model's expressive ability is limited. In addition, some research attempts to combine graph neural networks and Transformer structures to extract molecular graph and protein structure graph features and perform multi-modal fusion, but generally uses a simple concatenation method that fails to deeply mine the interaction mechanism between modalities. At the same time, in the modeling process of drug-target interactions, most methods use a one-way attention mechanism, which makes it difficult to accurately and effectively capture the key atom-residue pairs of drugs and targets, limiting the improvement of prediction performance. SUMMARY

[0005] In order to solve the above problems, the present invention proposes a drug-target interaction prediction method and system based on multimodal feature fusion to address the problems of insufficient utilization of single-modal features, simple multimodal fusion methods, and insufficient modeling of inter-modal interaction in existing drug-target interaction prediction methods. By combining the multimodal complementary information of drugs and targets, the deep interaction between different modal features is deeply explored; at the same time, a bidirectional collaborative attention mechanism is introduced in the interaction modeling of drugs and targets, which effectively improves the accuracy of drug-target binding affinity prediction.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: One or more embodiments provide a drug-target interaction prediction method based on multimodal feature fusion, comprising the following steps: The obtained drug sequence to be identified and target sequence are segmented and encoded, and the subsequence features of the drug and target are extracted respectively; Construct a two-dimensional molecular graph of the drug and a three-dimensional structural graph of the target, and extract the graph structural features of the drug and target respectively through the graph-structure-based Transformer model; Through the cross-attention mechanism, the subsequence features and graph structure features of drugs and targets are integrated to obtain multimodal fusion feature representations at the drug atomic level and target structure level; A bidirectional collaborative attention mechanism is adopted to calculate the attention weights of the drug-target and target-drug directions respectively. The obtained multimodal fusion feature representations at the drug atomic level and target structure level are interactively fused to obtain the binding affinity prediction results between the drug and the target.

[0007] One or more embodiments provide a drug-target interaction prediction system based on multimodal feature fusion, including: The subsequence feature extraction module is configured to perform word segmentation encoding on the obtained drug sequence to be identified and the target sequence, and extract subsequence features of the drug and the target respectively; The graph structure feature extraction module is configured to construct a two-dimensional molecular graph of the drug and a three-dimensional structural graph of the target, and extract the graph structure features of the drug and target respectively through the graph structure-based Transformer model; The cross-attention module is configured to fuse the subsequence features and graph structure features of drugs and targets through a cross-attention mechanism to obtain a multimodal fusion feature representation of drug atomic level and target structure level; The bidirectional collaborative attention module is configured to adopt a bidirectional collaborative attention mechanism to calculate the attention weights of the drug to the target and the target to the drug respectively, and interactively fuse the obtained multimodal fusion feature representations of the drug atomic level and the target structure level to obtain the binding affinity prediction results between the drug and the target.

[0008] Compared with the prior art, the present invention has the following beneficial effects: The method of the present invention integrates the structural features and subsequence features of drugs and targets for the first time, fully explores the complementary information between structural modalities and sequence semantics, enhances the model's comprehensive understanding of spatial configurations and functional regions, and significantly improves the feature expression ability; fully explores the deep interactive relationship between different modalities including structural features and subsequence features, and realizes information interaction between sequences and graph structures by designing a cross-attention mechanism, breaking the limitations of modal isolation; at the same time, a bidirectional collaborative attention mechanism is used to establish a bidirectional matching mechanism for atom-residue key pairs, enhancing the model's ability to capture effective binding sites, effectively improving the prediction accuracy of drug-target binding affinity and the generalization performance of the model.

[0009] The advantages of the present invention and its additional aspects will be described in detail in the following specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their description are used to explain the present invention but do not constitute a limitation of the present invention.

[0011] Figure 1 is a flow chart of the drug-target interaction prediction method of Example 1 of the present invention; Figure 2 Schematic diagram of performing a cross-attention operation on drug subsequence features and drug graph structure features according to Example 1 of the present invention; Figure 3 Schematic diagram of performing a cross-attention operation on target subsequence features and target graph structure features according to Example 1 of the present invention; FIG4( a ) is a curve showing the variation of mean square error with training rounds in an experimental example of Example 1 of the present invention; FIG4( b ) is a curve showing the change of the root mean square error with the number of training rounds in the experimental example of Example 1 of the present invention; FIG4( c ) is a curve showing the change of consistency index with training rounds in the experimental example of Example 1 of the present invention; FIG4( d ) is a curve showing the variation of the modified determination coefficient with the number of training rounds in the experimental example of embodiment 1 of the present invention. DETAILED DESCRIPTION

[0012] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0013] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0014] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof. It should be noted that, in the absence of conflict, the various embodiments of the present invention and the features in the embodiments can be combined with each other. The embodiments will be described in detail below with reference to the accompanying drawings.

[0015] Example 1 In the technical solutions disclosed in one or more embodiments, Figure 1 As shown in FIG4 , a drug-target interaction prediction method based on multimodal feature fusion includes the following steps: Step 1: perform word segmentation encoding on the obtained drug sequence to be identified and the target sequence to extract the drug target subsequence features; Step 2: Construct a two-dimensional molecular graph of the drug and a three-dimensional structural graph of the target, and extract the graph structural features of the drug and target respectively through the graph-structure-based Transformer model; Step 3: Through the cross-attention mechanism, the subsequence features and graph structure features of drugs and targets are fused to obtain a multimodal fusion feature representation of drug atomic level and target structure level; Step 4: Use a bidirectional collaborative attention mechanism to calculate the attention weights of the drug to the target and the target to the drug respectively, and interactively fuse the obtained multimodal fusion feature representations of the drug atomic level and the target structure level to obtain the binding affinity prediction results between the drug and the target.

[0016] Furthermore, the bidirectional collaborative attention mechanism specifically calculates the attention weights of the drug-target and target-drug directions respectively, and performs element-wise product and normalization on the attention weights calculated in the two directions to obtain bidirectionally consistent key atom-residue interaction pairs. This enables the screening of key interaction pairs from the perspective of bidirectional collaboration between drugs and targets, accurately captures key atom-residue pairs, and improves prediction performance. In this embodiment, the method first performs standardized encoding on the sequence information of the drug and the target. Subsequently, the drug is converted into its corresponding two-dimensional molecular graph and a three-dimensional structural graph of the target protein is constructed, and input into the respective graph structure Transformer modules to extract representation features at the graph structure level. Next, a cross-attention mechanism is used to establish a fusion path between the sequence features and graph features of the drug and target, respectively, to achieve dynamic weighted interaction between the modalities and obtain a unified feature representation at the drug atomic level and the target structural level. To further enhance the discriminability of drug-target interaction pairs, a bidirectional collaborative attention mechanism is introduced. By constructing two attention directions (drug to target and target to drug), two types of attention distribution weights are learned respectively and mutual information is fused, thereby more accurately screening key atom-residue pairing relationships, and finally outputting the binding affinity value through a regression network.

[0017] The method of this embodiment integrates the structural features and subsequence features of drugs and targets for the first time, fully exploits the complementary information between structural modalities and sequence semantics, enhances the model's comprehensive understanding of spatial configurations and functional regions, and significantly improves the feature expression capability. It fully exploits the deep interactive relationship between different modalities, including structural features and subsequence features, and realizes information interaction between sequences and graph structures by designing a cross-attention mechanism, thus breaking the limitation of modal isolation. At the same time, a bidirectional collaborative attention mechanism is used to establish a bidirectional matching mechanism for atom-residue key pairs, thereby enhancing the model's ability to capture effective binding sites and effectively improving the prediction accuracy of drug-target binding affinity and the generalization performance of the model.

[0018] In step 1, the method for extracting subsequence features of drugs and targets includes the following steps: Step 11: For the obtained drug sequence to be identified and target sequence, the Frequent Consecutive Subsequence (FCS) algorithm is used to perform word segmentation encoding and then splicing to obtain the drug target subsequence encoding representation; In this step, after encoding, the word segmentation results obtained from the drug sequence and the target sequence are concatenated into a subsequence and input into the trained BERT model for feature extraction; Alternatively, the drug sequence can use the drug's SMILES sequence. SMILES is a standardized method for representing molecular structural information as ASCII strings. It uses a series of rules to convert the atoms, bonds, rings, and branches in chemical molecules into a combination of characters and symbols, making complex chemical structures easy to store, process, and analyze in a text environment.

[0019] Optionally, in the step 11, the length of the frequently continuous subsequence of the drug sequence can be set as an integer in the interval [1, 8]; and the length of the frequently continuous subsequence of the target sequence can be set as an integer in the interval [1, 11] to balance the representation ability and the model complexity.

[0020] Optionally, the target can be a biological macromolecule with sequence information or structure information, such as a protein. In a specific implementation, the word segmentation coding adopts the FCS algorithm; a unified and universal subsequence dictionary (FCS dictionary) is constructed in advance, which contains 23532 drug subsequences and 16693 target subsequences, a total of 40230 subsequences (including 5 special markers), and the FCS algorithm is used based on the dictionary to perform word segmentation coding on the drug sequence and the target sequence respectively; The FCS algorithm (Frequent Consecutive Subsequence) is a method for automatically extracting frequently occurring continuous subsequences from sequence data, which is commonly used in bioinformatics and chemical informatics for feature coding of protein sequences or drug molecular structure sequences. Step 12, the drug target subsequence coding representation after word segmentation coding is used to extract the subsequence features of the drug and the target by using the trained BERT model; Specifically, the BERT model used in the embodiment contains multiple Transformer encoder layers (denoted as N), each layer is provided with multiple attention heads (denoted as h), and the intermediate hidden dimension of each layer is dm; and the final generated subsequence feature dimension is de. Optionally, N = 6, h = 12, dm = 1536, and de = 384.

[0021] In some embodiments, the BERT model is pre-trained in a self-supervised manner. The drug sequence coding representation and the target sequence coding representation in the training set are used as inputs, the FCS algorithm is used for word segmentation coding to obtain the corresponding subsequence representation, then the original subsequence content of the masked position is predicted by randomly performing a mask operation on part of the subsequence position, the pre-training model of the subsequence feature with migration ability is obtained, and then the pre-training model is used for formal training and fine-tuning to obtain the trained BERT model; wherein the pre-training and the formal training process are the same, and the description is as follows. Further, the training process of the BERT model includes the following steps: Step S121, data of the drug target database is obtained, and the data is divided to obtain a training set and a test set; During pre-training, the training process can obtain drug-target interaction sample data from the public BindingDB database. This database contains a total of 263,584 unlabeled drug-target sequence pairs, which can be used to learn contextual semantic representations within sequences during the BERT model pre-training phase.

[0022] The drug-target interaction sample data were randomly divided into a training set and a test set in a ratio of 8:2. The training set was used for model training, and the test set was used for model performance evaluation. It should be noted that in this embodiment, drug-target: connected by a dash, represents a drug-target data with a label, the label is interaction relationship or no interaction relationship; the data in the dataset appears in pieces, one drug data corresponds to one target data, and a dash is used to represent the associated data; Step S122: For the data sequences in the training set, the Frequent Consecutive Subsequence (FCS) dictionary constructed by the FCS algorithm in step 1 is used to perform word segmentation encoding on each drug-target sequence data to be predicted, and convert it into a corresponding digital index to obtain the encoded drug-target subsequence encoding representation after word segmentation encoding to meet the input requirements of the BERT model.

[0023] Step S123: In each training batch, based on the maximum length of the data in the corresponding training batch, length alignment processing is performed on the drug-target subsequence encoding representation after word segmentation encoding; During the training process, there are many training data, and the data needs to be divided into different batches. For example, batch=60 means that the length of 60 drug target sequence data is aligned. This alignment process is performed within the batch.

[0024] Specifically, for a sequence whose length exceeds the preset maximum length Lmax, only the first Lmax subsequences are retained; a sequence whose length is less than Lmax is padded with a special padding symbol [PAD].

[0025] Optionally, set Lmax=512.

[0026] After the above encoding and padding alignment, the dimension of each batch of sequence data can be expressed as (batch_size, Lmax). Optionally, set batch_size=60.

[0027] Step S124: using the masked subsequence prediction task as the self-supervised learning objective, randomly performing masking operations on some subsequences in the input sequence; Optionally, some subsequences can be taken as 15% of the total sequence data; The specific mask operation process is as follows: (1) Mask the first set proportion of subsequences in the selected subsequences and replace them with the set special mark [MSK]; The first set ratio may be 70%-80%, that is, 70%-80% of the 15% subsequences are selected from the total sequence data; (2) replacing the subsequences of the second set ratio in the selected subsequences with the subsequences in the FCS dictionary; For example, 10% of the subsequences can be replaced with subsequences in the FCS dictionary to achieve masking; wherein the sum of the first set ratio and the second set ratio is less than 1; (3) The remaining input sequence remains unchanged; The above masking strategy can provide effective input for the subsequent BERT model to learn context-related features. The masked input sequence The structure can be expressed as:

[0028] Among them, [CLS] is the classification marker, and [SEP] is the sequence separation marker, which is used to distinguish drug sequences from target sequences. is the input sequence, is the masked sequence; Step S125: Input the masked subsequence into the BERT model, extract the subsequence feature representation, adjust the parameters of the BERT model by the size of the loss function, and iterate the training until the convergence condition is met to obtain the trained BERT model; During the BERT model training process, the Adam optimizer can be used to optimize the model parameters. For example, the learning rate can be set to 1×10 −5 , the Dropout probability can be set to 0.1, the training batch size to 32, and the total number of training rounds to 100.

[0029] Optionally, the BERT model is trained using a masked subsequence prediction task, and its loss function can be expressed as: ; in, represents the masked ith subsequence, s≠i represents all other subsequences except the ith subsequence, p(s i |s≠i) is the masked subsequence predicted by the BERT model probability.

[0030] Unlike the traditional BERT method that only uses the features corresponding to the [CLS] tag, this implementation retains all subsequence feature representations through masking during the above training process to obtain richer sequence information.

[0031] It should be noted that the BERT model pre-training process only needs to be performed once. Once a pre-trained model with good transferability is obtained, it can be used for feature extraction in subsequent tasks. This pre-trained model can be used as a general semantic feature extractor, suitable for representation learning and downstream task modeling of various drug target datasets.

[0032] In some embodiments, a method for encoding a drug target subsequence after word segmentation encoding and extracting drug target subsequence features using a pre-trained BERT model includes the following steps: Step 121: Input the encoded representation of the drug-target subsequence after word segmentation encoding into the trained BERT model, extract the subsequence embedding vectors of the drug and target respectively, and obtain the feature representation at the word segmentation level; Unlike the traditional BERT approach, which only uses features corresponding to the [CLS] tag, this implementation retains all subsequence feature representations to obtain richer sequence information. The feature dimensions after the BERT model output can be expressed as (batch_size, Lmax, de), where de is the subsequence feature dimension. Optionally, de is set to 128.

[0033] Step 122: embed the extracted drug and target subsequences into vectors, divide them into drug parts and target parts, and perform padding and alignment on them to obtain a feature matrix with unified dimensions, that is, obtain the extracted drug and target subsequence features; Specifically, the padding alignment method can identify the maximum length of the drug's subsequence embedding vector and the target's subsequence embedding vector, and fill all the obtained subsequence embedding vectors with the set padding characters.

[0034] In step 123, each drug and target subsequence embedding vector is segmented into a drug part and a target part according to the position of the separator [SEP] explicitly inserted when inputting the BERT model.

[0035] Since the lengths of drug sequences and target sequences may be different, in order to meet the alignment requirements of the subsequent cross-attention module, the respective feature matrices are padded and aligned based on the maximum subsequence length of the drug to be identified and the target.

[0036] After padding and alignment, the dimensions of the drug subsequence feature matrix can be expressed as (batch_size, Ld, de), and the dimensions of the target subsequence feature matrix can be expressed as (batch_size, Lp, de), where Ld and Lp represent the maximum lengths of drug and target subsequences within the corresponding batch, respectively. It should be noted that the maximum subsequence lengths of the drugs and targets to be identified can vary, resulting in different dimensions in the output feature matrix.

[0037] Furthermore, in step 2, a two-dimensional molecular graph of the drug and a three-dimensional structural graph of the target are constructed, and the graph structure features of the drug and target are extracted respectively through a graph-structure-based Transformer model, including the following steps: Step 21: construct a two-dimensional molecular graph for the SMILES sequence of the drug molecule, where the nodes in the two-dimensional molecular graph represent the atoms in the drug molecule, and the edges represent the chemical bond connections between the atoms.

[0038] Alternatively, the SMILES sequence data of the drug can be directly obtained from public drug-target interaction datasets such as Davis, KIBA, BIOSNAP, etc., or retrieved from public drug databases such as DrugBank, ChEMBL, etc. based on the drug identification information therein.

[0039] Specifically, in this embodiment, the chemical information processing tool RDKit may be used to parse the SMILES sequence to generate a standard molecular graph structure.

[0040] Optionally, the node features of the two-dimensional drug molecular graph may include chemical properties such as atom type, atomic valence state, hybrid orbital type, aromaticity mark, number of hydrogen atoms, etc.; the edge features include information such as chemical bond type (such as single bond, double bond, triple bond), whether it is an aromatic bond, conjugation status, and whether it is located in a ring structure.

[0041] Step 22: Input the two-dimensional molecular graph obtained in step 21 into the graph-based Transformer model for feature extraction. By aggregating and transforming the features of the nodes and their neighborhoods, a high-dimensional graph structure feature representation of the drug, i.e., the graph structure feature of the drug, is obtained. Furthermore, the graph-based Transformer model includes multiple layers of stacked Transformer encoders, each layer of which is connected to multiple attention heads; each layer of output is connected to a Dropout regularization module, and the regularization probability is set to 0.2.

[0042] The graph-based Transformer model stacks Lg layers of Transformer encoders, each containing hg attention heads, and the node feature dimension is set to dg. In this embodiment, Lg = 6, hg = 8, and dg = 128.

[0043] In the Transformer model based on graph structure The feature update method of the i-th node in the layer can be expressed as: ; in, represents the feature representation of the i-th node in the l-th layer, and N(i) represents the set of neighbor nodes of the i-th node.

[0044] After processing by the graph-based Transformer model, the final dimension of the drug molecule graph feature for each batch can be expressed as (batch_size, nd, dg), where nd represents the maximum number of atomic nodes in the drug molecules in the corresponding batch.

[0045] Step 23: constructing a target graph structure based on the three-dimensional structure data of the target, wherein the nodes in the target graph structure represent amino acid residues, and the edges represent the spatial interaction relationship between the residues; Alternatively, the three-dimensional structural data of the target can be obtained from a public database such as the Protein Data Bank (PDB) or the AlphaFold database, and a residue-level graph structure representation can be obtained using existing graph construction methods.

[0046] During target graph construction, node features can be selected based on residue type, secondary structure type, physicochemical properties (such as hydrophobicity and charge), and the coordinates of the residue's central atom. Edge features can be defined based on the spatial distance between residues. Specifically, an edge connection is established when the minimum distance between the central atoms of two residues is less than 10 Å.

[0047] Step 24: Input the target graph structure obtained in step 23 into the graph-structure-based Transformer model, aggregate and transform the node features and edge features in the graph, and obtain a high-dimensional graph structure feature representation of the target, that is, the graph structure feature of the target.

[0048] After processing by the graph-based Transformer model, the final dimension of the target graph structure features of each batch can be expressed as (batch_size, np, dg), where np represents the maximum number of residue nodes contained in the target structure in the corresponding batch.

[0049] Furthermore, in step 3, the subsequence features and graph structure features of drugs and targets are fused through the cross-attention mechanism to obtain a multimodal fusion feature representation of drug atomic level and target structure level, including the following steps: Step 31: For the drug subsequence features and drug graph structural features, the drug graph structural features are mapped as the query vector, and the drug subsequence features are mapped as the key vector and value vector, and a cross-attention operation is performed to obtain the drug atomic-level multimodal fusion feature representation; Specifically, the drug subsequence features extracted in step 1 and the drug graph structure features extracted in step 2 are input into the cross attention module, and the scaled dot product attention mechanism is used for feature fusion.

[0050] like Figure 2 As shown in the figure, in this embodiment, the drug graph structural features are linearly mapped as the query vector, and the drug subsequence features are linearly mapped as the key vector and value vector respectively. The three types of vectors are all mapped to the same feature dimension dg to meet the attention calculation requirements. The calculation formula for feature fusion based on the attention mechanism is: ; Where Q represents the query vector, K represents the key vector, the superscript T represents the transpose, V represents the value vector, and dg represents the feature dimension. The matrix dimension of the calculated atomic-level multimodal fusion feature representation of drugs is (batch_size, nd, dg).

[0051] Step 32: For the target subsequence features and target graph structure features, the target graph structure features are mapped as the query vector, and the target subsequence features are mapped as the key vector and value vector, and a cross-attention operation is performed to obtain a target structure-level multimodal fusion feature representation; Specifically, the target subsequence features extracted in step 1 and the target graph structure features extracted in step 2 are input into the cross attention module, and the scaled dot product attention mechanism is used for feature fusion.

[0052] like Figure 3 As shown, in this embodiment, the target graph structure features are linearly mapped as the query vector, and the target subsequence features are linearly mapped as the key vector and value vector. The fusion calculation process is similar to that in step 31. The calculated multimodal fusion feature representation matrix of the target residue level within each batch has a dimension of (batch_size, np, dg).

[0053] Furthermore, in step 4, a bidirectional collaborative attention mechanism is used to interactively fuse the obtained multimodal fusion feature representations at the drug atomic level and the target structure level, including the following steps: Step 41: Input the drug atomic-level multimodal fusion feature representation and target structure-level multimodal fusion feature representation obtained in step 3 into the bidirectional collaborative attention module, and perform bilinear interaction scoring on the drug and target features based on the set learnable matrix to obtain the initial interaction matrix ; The atomic-level multimodal features of drugs are expressed as: ; The target residue-level multimodal features are expressed as: ; Furthermore, bilinear interaction scoring uses a learnable matrix Project and match the drug and target features to obtain the initial interaction matrix: ; in: ; in, represents the eigenvector of the i-th drug atom, and the superscript T indicates transposition; represents the j-th target residue feature vector. represents the interaction matching score between the i-th drug atom and the j-th target residue; In this embodiment, a learnable parameter matrix is ​​used , which realizes the learnable feature transformation capability and can learn the linear mapping relationship between different dimensions; is a parameter matrix, which can be initialized to a normal distribution and updated by gradient descent during training; Step 42: Based on the initial interaction matrix , calculate the first attention weight matrix of drug to target respectively , and the second attention weight matrix of target to drug ; The first attention weight matrix of drug to target The construction method is: fix the drug node i, perform weighted normalization on all target nodes j, and measure which target nodes the drug node pays more attention to. That is, the interaction matrix The i-th row element is calculated: ; Second attention weight matrix of target to drug The construction method is: fix the target node j, perform weighted normalization on all drug nodes i, and measure which drug nodes the target node pays more attention to. That is, the interaction matrix Calculate the j-th column element: ; in, Represents the matrix Perform exponential calculations with base e; Step 43: Multiply and normalize the elements in the first attention weight matrix and the second attention weight matrix element by element to obtain the final bidirectional collaborative attention weight. , the formula is as follows: ; Step 44: Based on bidirectional collaborative attention weight , perform weighted fusion of drug and target node features to obtain the joint feature vector of each drug-target pair ; , ; in, , are all learnable parameters.

[0054] Furthermore, the joint eigenvector Input the multi-layer perceptron model to perform regression prediction of drug-target binding affinity and obtain the prediction result of drug-target binding affinity; In the above scheme of this embodiment, by calculating the attention weights of the drug-target and target-drug directions respectively, and performing element-by-element multiplication and normalization on the two sets of attention weights, a consistent interaction weight matrix that takes into account the perspectives of both parties is obtained, thereby accurately screening the key atom-residue pairs that are bidirectionally recognized between the drug and the target. Compared with the traditional unidirectional attention mechanism that only models the interaction relationship from a single direction, the bidirectional collaborative attention mechanism proposed in this embodiment introduces a cross-screening strategy based on a bidirectional collaborative perspective, effectively filtering out unidirectional bias and low-confidence false pairings, strengthening the key connections in the biological sense, achieving accurate capture of key atom-residue pairs, and improving the model prediction performance; The overall model that implements the above steps includes the BERT model, the graph-based Transformer model, the cross-attention module, the bidirectional collaborative attention module, and the multi-layer perceptron model. During the training of the above overall model, the mean square error between the predicted affinity value and the true affinity value is used as the loss function, which is defined as follows: ; in, is the true affinity value of the i-th sample, is the predicted affinity value of the i-th sample, and n represents the number of samples in the current batch.

[0055] During model training, the parameters of the BERT pre-trained module and the graph-based Transformer model are jointly optimized without freezing or staged fine-tuning. The entire model is trained end-to-end using the aforementioned loss function.

[0056] In order to illustrate the effect of the method of this embodiment, an experimental verification was conducted to evaluate the performance of the method of this embodiment in the drug-target interaction prediction task.

[0057] During model training, in this embodiment, after each round of training, the model performance is evaluated using the test set obtained in step 2. The model performance of each round is continuously recorded during the evaluation process, and the model with the best performance on the test set is ultimately selected as the final output model for subsequent comparison and application analysis.

[0058] The model training uses the Adam optimizer for parameter optimization and uses differentiated learning rate settings. The learning rate of the pre-trained BERT module is set to 1×10−5, and the learning rates of the remaining modules (including the graph-based Transformer model, cross attention module, bidirectional collaborative attention module, and multi-layer perceptron model) are set to .

[0059] To further improve the stability of model training, a learning rate decay strategy is introduced. When the model performance does not improve significantly in 50 consecutive rounds of training, the current learning rate is automatically reduced to 0.8 times the original value until the minimum learning rate is reached. .

[0060] After the model training is completed, the performance change trend of the overall model of this embodiment during the entire training process is visualized based on the test set evaluation results. Figures 4 (a) to 4 (d) show four key indicators: Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Concordance Index (CI), and Corrected Determination Coefficient ( ) changes with the training round (Epoch).

[0061] The experimental results show that with the increase of training rounds, the MSE and RMSE indicators continue to decrease, and the CI and The indicators gradually increased and stabilized, indicating that the method proposed in this example can stably learn the deep interaction relationship between drugs and targets, and has good training effect and generalization ability.

[0062] To further verify the predictive advantages of the method of this embodiment in practical applications, the above four indicators were used to compare its performance with the existing graph neural network-based drug-target interaction prediction method GraphDTA on the same Davis dataset test set. The experimental results are shown in Table 1. GraphDTA method, using the method described in the reference "Nguyen, T., Le, H., Quinn, TP, Nguyen,T., Le, TD,&Venkatesh, S. (2020). GraphDTA: Predicting drug-target binding affinity with graph neural networks. Bioinformatics, 37(8), 1140-1147. http: / / dx.doi.org / 10.1093 / bioinformatics / btaa921." Table 1 Comparison of drug target affinity prediction performance between the method of this embodiment and the GraphDTA method;

[0063] Experimental results demonstrate that this method, based on the proposed multimodal feature fusion and incorporating a bidirectional collaborative attention mechanism for drug-target interaction modeling, outperforms existing GraphDTA methods across all key performance metrics, demonstrating stronger predictive power and generalization. In particular, the mean squared error (MSE) significantly decreases from 0.213 for the GraphDTA method to 0.193 for this method, further validating the effectiveness and advancement of the proposed method in drug-target affinity prediction.

[0064] Example 2 Based on Example 1, this example provides a drug-target interaction prediction system based on multimodal feature fusion, including: The subsequence feature extraction module is configured to perform word segmentation encoding on the obtained drug sequence to be identified and the target sequence, and extract subsequence features of the drug and the target respectively; The graph structure feature extraction module is configured to construct a two-dimensional molecular graph of the drug and a three-dimensional structural graph of the target, and extract the graph structure features of the drug and target respectively through the graph structure-based Transformer model; The cross-attention module is configured to fuse the subsequence features and graph structure features of drugs and targets through a cross-attention mechanism to obtain a multimodal fusion feature representation of drug atomic level and target structure level; The bidirectional collaborative attention module is configured to adopt a bidirectional collaborative attention mechanism to calculate the attention weights of the drug to the target and the target to the drug respectively, and interactively fuse the obtained multimodal fusion feature representations of the drug atomic level and the target structure level to obtain the binding affinity prediction results between the drug and the target.

[0065] Furthermore, the bidirectional collaborative attention module includes: The interactive scoring module is configured to input the drug atomic-level multimodal fusion feature representation and the target structure-level multimodal fusion feature representation into the bidirectional collaborative attention module, and perform bilinear interactive scoring on the drug and target features based on the set learnable matrix to obtain the initial interaction matrix ; A first calculation module is configured to calculate the initial interaction matrix based on the , calculate the first attention weight matrix of drug to target respectively , and the second attention weight matrix of target to drug ; The second calculation module is configured to perform element-by-element multiplication and normalization operations on the elements in the first attention weight matrix and the second attention weight matrix to obtain the final bidirectional collaborative attention weight ; The weighted fusion module is configured based on the bidirectional collaborative attention weights , perform weighted fusion of drug and target node features to obtain the joint feature vector of each drug-target pair .

[0066] It should be noted here that the various modules in this embodiment correspond one-to-one to the various steps in Example 1, and the specific implementation processes are the same, which will not be repeated here.

[0067] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

[0068] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A drug-target interaction prediction method based on multimodal feature fusion, characterized in that: The steps include: The obtained drug sequence to be identified and target sequence are segmented and encoded, and the subsequence features of the drug and target are extracted respectively; Construct a two-dimensional molecular graph of the drug and a three-dimensional structural graph of the target, and extract the graph structural features of the drug and target respectively through the graph-structure-based Transformer model; Through the cross-attention mechanism, the subsequence features and graph structure features of drugs and targets are integrated to obtain multimodal fusion feature representations at the drug atomic level and target structure level; A bidirectional collaborative attention mechanism is adopted to calculate the attention weights of the drug-target and target-drug directions respectively. The obtained multimodal fusion feature representations at the drug atomic level and target structure level are interactively fused to obtain the binding affinity prediction results between the drug and the target.

2. The drug-target interaction prediction method based on multimodal feature fusion according to claim 1, wherein: The method for extracting subsequence features of drugs and targets includes the following steps: For the obtained drug sequences and target sequences to be identified, the frequent retained subsequence algorithm is used to perform word segmentation encoding; The drug subsequence encoding representation and target subsequence encoding representation after word segmentation encoding are used, and the trained BERT model is used to extract the subsequence features of drugs and targets respectively.

3. The drug-target interaction prediction method based on multimodal feature fusion according to claim 2, characterized in that: The method of extracting subsequence features of drugs and targets respectively using the trained BERT model from the drug sequence encoding representation and target sequence encoding representation after word segmentation encoding includes the following steps: The drug-target subsequence encoding representation after word segmentation encoding is input into the trained BERT model to extract the subsequence embedding vectors of the drug and target respectively; The extracted subsequences of drugs and targets are embedded into vectors, divided into drug part and target part, and padded and aligned respectively to obtain a feature matrix with unified dimension, that is, the extracted subsequence features of drugs and targets are obtained.

4. The drug-target interaction prediction method based on multimodal feature fusion according to claim 1, wherein: Construct a two-dimensional molecular graph of the drug and a three-dimensional structural graph of the target. Use the graph-based Transformer model to extract the graph structural features of the drug and target, respectively. This includes the following steps: Construct a two-dimensional molecular graph based on the SMILES sequence of the drug molecule; The obtained two-dimensional molecular graph is input into the graph-based Transformer model for feature extraction. By aggregating and transforming the features of the nodes and their neighborhoods, the graph structure features of the drug are obtained. A target graph structure is constructed based on the three-dimensional structural data of the target. The nodes in the target graph structure represent amino acid residues, and the edges represent the spatial interaction relationship between residues. The obtained target graph structure is input into the graph-structure-based Transformer model, and the node features and edge features in the target graph structure are aggregated and transformed to obtain the graph structure features of the target.

5. The drug-target interaction prediction method based on multimodal feature fusion according to claim 4, characterized in that: The graph-based Transformer model includes multiple layers of stacked Transformer encoders. Each layer of Transformer encoders is connected to multiple attention heads; the output end of each layer is connected to a Dropout regularization module.

6. The drug-target interaction prediction method based on multimodal feature fusion according to claim 1, wherein: Through the cross-attention mechanism, the subsequence features and graph structure features of drugs and targets are fused to obtain a multimodal fusion feature representation of drug atomic level and target structure level, including the following steps: For drug subsequence features and drug graph structural features, the drug graph structural features are mapped as query vectors, and the drug subsequence features are mapped as key vectors and value vectors, and a cross-attention operation is performed to obtain the atomic-level multimodal fusion feature representation of drugs; For the target subsequence features and target graph structure features, the target graph structure features are mapped as the query vector, and the target subsequence features are mapped as the key vector and value vector, and a cross-attention operation is performed to obtain the target structure-level multimodal fusion feature representation.

7. The drug-target interaction prediction method based on multimodal feature fusion according to claim 1, wherein: A bidirectional collaborative attention mechanism is used to interactively fuse the obtained multimodal fusion feature representations of drug atomic level and target structure level, including the following steps: The drug atomic-level multimodal fusion feature representation and the target structure-level multimodal fusion feature representation are bilinearly interactively scored based on the set learnable matrix to obtain the initial interaction matrix ; Based on the initial interaction matrix , calculate the first attention weight matrix of drug to target, and the second attention weight matrix of target to drug respectively; Multiply and normalize the elements in the first attention weight matrix and the second attention weight matrix element by element to obtain the final bidirectional collaborative attention weight; Based on the bidirectional collaborative attention weight, the drug and target node features are weighted fused to obtain the joint feature vector of each drug-target pair .

8. The drug-target interaction prediction method based on multimodal feature fusion according to claim 7, wherein: The joint eigenvector The multi-layer perceptron model is input to perform regression prediction of drug-target binding affinity to obtain the prediction result of drug-target binding affinity.

9. A drug-target interaction prediction system based on multimodal feature fusion, characterized by: include: The subsequence feature extraction module is configured to perform word segmentation encoding on the obtained drug sequence to be identified and the target sequence, and extract subsequence features of the drug and the target respectively; The graph structure feature extraction module is configured to construct a two-dimensional molecular graph of the drug and a three-dimensional structural graph of the target, and extract the graph structure features of the drug and target respectively through the graph structure-based Transformer model; The cross-attention module is configured to fuse the subsequence features and graph structure features of drugs and targets through a cross-attention mechanism to obtain a multimodal fusion feature representation of drug atomic level and target structure level; The bidirectional collaborative attention module is configured to adopt a bidirectional collaborative attention mechanism to calculate the attention weights of the drug to the target and the target to the drug respectively, and interactively fuse the obtained multimodal fusion feature representations of the drug atomic level and the target structure level to obtain the binding affinity prediction results between the drug and the target.

10. The drug-target interaction prediction system based on multimodal feature fusion according to claim 9, characterized in that: Bidirectional collaborative attention module, including: The interactive scoring module is configured to perform bilinear interactive scoring on the drug and target features based on the set learnable matrix by combining the drug atomic-level multimodal fusion feature representation and the target structure-level multimodal fusion feature representation to obtain the initial interaction matrix. ; A first calculation module is configured to calculate the initial interaction matrix based on the , calculate the first attention weight matrix of drug to target, and the second attention weight matrix of target to drug respectively; A second calculation module is configured to perform element-by-element multiplication and normalization operations on the elements in the first attention weight matrix and the second attention weight matrix to obtain a final bidirectional collaborative attention weight; The weighted fusion module is configured to perform weighted fusion of drug and target node features based on bidirectional collaborative attention weights to obtain the joint feature vector of each drug-target pair. .

Citation Information

Patent Citations

  • Drug target affinity prediction method based on multi-mode collaborative generation

    CN119230012A

  • Application of multi-modal feature fusion model in drug target binding affinity prediction

    CN119479783A

  • Drug target combinability prediction method based on variational coding

    CN119580827A

  • Drug target binding affinity prediction method based on multi-modal data fusion enhancement

    CN120199318A

  • Method of determining meteorological information, electronic device and storage medium

    US20250110255A1

Cited By

  • Drug target intelligent prediction method and system

    CN120954490A