A Drug Target Interaction Prediction Method and System Based on Multimodal Feature Fusion

Abstract

This invention relates to the fields of bioinformatics and artificial intelligence, and proposes a method and system for predicting drug-target interactions based on multimodal feature fusion. The method includes: segmenting and encoding the acquired drug and target sequences to be identified; extracting sub-sequence features of the drug and target respectively; constructing a two-dimensional molecular graph of the drug and a three-dimensional structural graph of the target, and extracting graph structural features of the drug and target respectively; fusing the sub-sequence features and graph structural features of the drug and target through a cross-attention mechanism; and using a bidirectional collaborative attention mechanism for interactive fusion to obtain the predicted binding affinity of the drug and target. This disclosure, by combining multimodal complementary information of the drug and target, deeply explores the deep interactions between different modal features; simultaneously, by introducing a bidirectional collaborative attention mechanism in the interaction modeling of the drug and target, it effectively improves the accuracy of drug-target binding affinity prediction.

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] The present application aims to solve the above problems, aiming at the problems of insufficient utilization of single modal features, simple multi-modal fusion method and insufficient interaction modeling between modes in the existing drug-target interaction prediction method, a drug-target interaction prediction method and system based on multi-modal feature fusion are proposed, by combining the multi-modal complementary information of drugs and targets, the deep interaction between different modal features is deeply mined, and at the same time, the 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 purpose, the present application adopts the following technical scheme:

[0007] One or more embodiments provide a drug-target interaction prediction method based on multi-modal feature fusion, comprising the following steps:

[0008] The obtained to-be-recognized drug sequence and target sequence are segmented and encoded, and the subsequence features of the drug and the target are extracted respectively;

[0009] A two-dimensional molecular graph of the drug and a three-dimensional structure graph of the target are constructed, and the graph structure features of the drug and the target are extracted respectively through a graph structure based Transformer model;

[0010] Through the cross attention mechanism, the subsequence features and the graph structure features of the drug and the target are fused to obtain multi-modal fusion feature representations of the drug at the atomic level and the target at the structure level;

[0011] A bidirectional collaborative attention mechanism is adopted to calculate the attention weights of the drug to the target and the target to the drug in two directions respectively, and the obtained multi-modal fusion feature representations of the drug at the atomic level and the target at the structure level are interactively fused to obtain the binding affinity prediction result of the drug and the target.

[0012] One or more embodiments provide a drug-target interaction prediction system based on multi-modal feature fusion, comprising:

[0013] The subsequence feature extraction module is configured to segment and encode the obtained to-be-recognized drug sequence and target sequence, and extract the subsequence features of the drug and the target respectively;

[0014] The graph structure feature extraction module is configured to construct a two-dimensional molecular graph of the drug and a three-dimensional structure graph of the target, and extract the graph structure features of the drug and the target through a graph structure based Transformer model respectively;

[0015] The cross attention module is configured to fuse the subsequence features and the graph structure features of the drug and the target through the cross attention mechanism to obtain multi-modal fusion feature representations of the drug at the atomic level and the target at the structure level;

[0016] The bidirectional collaborative attention module is configured to adopt a bidirectional collaborative attention mechanism to calculate attention weights of two directions of a drug to a target and a target to the drug respectively, and to obtain a drug-target binding affinity prediction result by interacting and fusing obtained multi-modal fusion feature representations of drug atom levels and target structure levels.

[0017] Compared with the prior art, the present application has the following beneficial effects:

[0018] The method of the present application firstly fuses the structural features and subsequence features of the drug and the target, fully mines the complementary information between the structure modal and the sequence semantics, enhances the comprehensive understanding of the model to the spatial configuration and the functional region, and significantly improves the feature expression capability; fully mines the deep interaction relationship between different modes including the structural features and the subsequence features, breaks the limitation of the modal isolation by designing the cross-attention mechanism to realize the information interaction between the sequence and the graph structure; at the same time, the bidirectional matching mechanism of the atom-residue key pair is established by using the bidirectional collaborative attention mechanism, the capture ability of the model to the effective binding site is enhanced, and the prediction accuracy of the drug-target binding affinity and the generalization performance of the model are effectively improved.

[0019] The advantages of the present application and the advantages of the additional aspects will be described in detail in the following specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0020] The drawings accompanying the specification of the present application form a part thereof and serve to provide further understanding of the present application, the illustrative embodiments of the present application and their descriptions serve to explain the present application, and do not constitute limitations thereof.

[0021] Figure 1 is a flowchart of the drug-target interaction prediction method of embodiment 1 of the present application;

[0022] Figure 2 is a schematic diagram of the cross-attention operation of the drug subsequence features and the drug graph structure features of embodiment 1 of the present application;

[0023] Figure 3 is a schematic diagram of the cross-attention operation of the target subsequence features and the target graph structure features of embodiment 1 of the present application;

[0024] Figure 4(a) is a curve of the mean square error changing with the training round in the test example of embodiment 1 of the present application;

[0025] Figure 4(b) is a curve of the root mean square error changing with the training round in the test example of embodiment 1 of the present application;

[0026] Figure 4(c) is a curve of the consistency index changing with the training round in the test example of embodiment 1 of the present application;

[0027] Figure 4(d) is the curve of the change of the corrected decision coefficient with the number of training rounds in the experimental example of Embodiment 1 of the present invention. Detailed Implementation

[0028] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0029] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0030] It should be noted that the terminology used herein is for describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features within those embodiments can be combined with each other. The embodiments will now be described in detail with reference to the accompanying drawings.

[0031] Example 1

[0032] In one or more of the technical solutions disclosed in the embodiments, such as Figure 1 As shown in Figure 4, a drug target interaction prediction method based on multimodal feature fusion includes the following steps:

[0033] Step 1: Segment and encode the obtained drug sequence to be identified and target sequence, and extract the features of the drug target sub-sequence;

[0034] Step 2: 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 the target respectively using a graph-based Transformer model;

[0035] Step 3: By using the cross-attention mechanism, the sub-sequence features and graph structure features of the drug and target are fused to obtain multimodal fusion feature representations at the drug atomic level and target structural level;

[0036] Step 4: Using a bidirectional collaborative attention mechanism, the attention weights of the drug to the target and the target to the drug in both directions are calculated. The multimodal fusion feature representations of the drug at the atomic level and the target at the structural level are interactively fused to obtain the prediction results of the binding affinity between the drug and the target.

[0037] Furthermore, the bidirectional synergistic attention mechanism specifically involves calculating the attention weights of the drug on the target and the target on the drug in both directions, and performing element-wise multiplication and normalization on the attention weights calculated in both directions to obtain bidirectional consistent key atom-residue interaction pairs. This enables the screening of key interaction pairs from the perspective of bidirectional synergy between the drug and the target, achieving accurate capture of key atom-residue pairs and improving prediction performance.

[0038] In this embodiment, the method first standardizes and encodes the sequence information of the drug and target. Then, the drug is converted into its corresponding two-dimensional molecular graph, and a three-dimensional structural graph of the target protein is constructed. These are then input into their respective graph structure Transformer modules to extract representational features at the graph structure level. Next, a cross-attention mechanism is used to establish fusion pathways between the sequence features and graph features of the drug and target, respectively, to achieve dynamic weighted interaction between modalities and obtain unified feature representations at the atomic and structural levels of the drug and target. To further enhance the discriminative power 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 and mutual information is fused, thereby more accurately screening key atom-residue pairings. Finally, the binding affinity value is output through a regression network.

[0039] The method in this embodiment, for the first time, integrates the structural features and sub-sequence features of drugs and targets, 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 feature expression capabilities. It fully explores the deep interaction relationships between different modalities, including structural features and sub-sequence features, and achieves information interaction between sequences and graph structures through the design of a cross-attention mechanism, breaking the limitation of modal isolation. At the same time, it uses a bidirectional collaborative attention mechanism to establish a bidirectional matching mechanism for key atomic-residue pairs, enhances the model's ability to capture effective binding sites, and effectively improves the prediction accuracy of drug-target binding affinity and the model's generalization performance.

[0040] Step 1, the method for extracting subsequence features of the drug and target, includes the following steps:

[0041] Step 11: For the obtained drug sequence and target sequence to be identified, the Frequent Consecutive Subsequence (FCS) algorithm is used to perform word segmentation and encoding, and then the sequences are concatenated to obtain the drug target subsequence encoding representation.

[0042] In this step, after encoding, the word segmentation results obtained from the drug sequence and the word segmentation results obtained from the target sequence are concatenated into a subsequence, which is then input into the trained BERT model for feature extraction.

[0043] Optionally, the drug sequence can be the drug's SMILES sequence; SMILES is a standardized method for representing molecular structural information in the form of ASCII strings. It converts the atomic, bond, ring, branch, and other structures in chemical molecules into combinations of characters and symbols through a series of rules, making it easy to store, process, and analyze complex chemical structures in a text environment.

[0044] Optionally, in step 11, the length of the frequent continuous subsequence of the drug sequence can be set to an integer in the range [1, 8]; the length of the frequent continuous subsequence of the target sequence can be set to an integer in the range [1, 11], in order to balance representation capability and model complexity.

[0045] Optionally, the target can be a biological macromolecule with sequence information or structural information, such as a protein;

[0046] One specific implementation method is to use the FCS algorithm for word segmentation and encoding; a unified and universal subsequence dictionary (FCS dictionary) is pre-constructed, which contains 23,532 drug subsequences and 16,693 target subsequences, totaling 40,230 subsequences (including 5 special markers). Based on this dictionary, the FCS algorithm is used to segment and encode the drug sequence and the target sequence respectively.

[0047] Among them, the FCS algorithm (Frequent Consecutive Subsequence) is a method for automatically extracting frequently occurring contiguous subsequences from sequence data. It is often used in bioinformatics and cheminformatics for feature encoding of protein sequences or drug molecule structure sequences.

[0048] Step 12: Encode the drug target sub-sequence after word segmentation and use the trained BERT model to extract the sub-sequence features of the drug and target.

[0049] Specifically, the BERT model used in this embodiment contains multiple Transformer encoder layers (denoted as N), each layer has multiple attention heads (denoted as h), and the intermediate hidden dimension of each layer is dm; the feature dimension of the finally generated subsequence is de. Optionally, it can be set as: N=6, h=12, dm=1536, de=384.

[0050] In some embodiments, the BERT model is pre-trained using a self-supervised approach. Taking the drug sequence encoding representation and target sequence encoding representation from the training set as input, the FCS algorithm is first used for word segmentation and encoding to obtain the corresponding sub-sequence representations. Then, a partial sub-sequence position is randomly masked, and the model is pre-trained by predicting the original sub-sequence content at the masked positions, resulting in a pre-trained sub-sequence feature model with transferability. Finally, formal training and fine-tuning are performed based on the pre-trained model to obtain the trained BERT model. The pre-training and formal training processes are identical, as explained below.

[0051] Furthermore, the training process of the BERT model includes the following steps:

[0052] Step S121: Obtain data from the drug target database and divide it into training and testing sets;

[0053] During pre-training, drug-target interaction sample data can be obtained from the publicly available BindingDB database. This database contains a total of 263,584 unlabeled drug-target sequence pairs, which can be used to learn the contextual semantic representation within the sequences during the BERT model pre-training phase.

[0054] The drug-target interaction sample data were randomly divided into a training set and a test set in an 8:2 ratio, with the training set used for model training and the test set used for model performance evaluation.

[0055] It should be noted that in this embodiment, drug-target: a short horizontal line connects to represent a tagged drug target data, with the tags indicating whether there is an interaction relationship or not; the data in the dataset appears in rows, with one drug data corresponding to one target data, and the short horizontal line represents the related data;

[0056] Step S122: For the data sequences in the training set, use the FCS dictionary constructed by the Frequent Consecutive Subsequence (FCS) algorithm in Step 1 to segment and encode each drug-target sequence data to be predicted, and convert it into the corresponding numerical index to obtain the segmented and encoded drug-target subsequence encoding representation to meet the input requirements of the BERT model.

[0057] Step S123: Within each training batch, based on the maximum length of the data in the corresponding training batch, perform length alignment processing on the drug-target sub-sequence encoding representation after word segmentation encoding;

[0058] During training, there are many training data sets, which need 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 a batch.

[0059] Specifically, for sequences whose length exceeds the preset maximum length Lmax, only the first Lmax subsequences are retained; sequences whose length is less than Lmax are padded with a special padding symbol [PAD].

[0060] Optionally, set Lmax=512.

[0061] After the above encoding and padding alignment, the sequence data dimension of each batch can be represented as (batch_size, Lmax). Optionally, batch_size is set to 60.

[0062] Step S124: Using the masked subsequence prediction task as the self-supervised learning objective, randomly perform masking operations on a portion of the subsequences in the input sequence;

[0063] Optionally, a subset of the sequence can be selected, such as 15% of the total sequence data.

[0064] The specific masking process is as follows:

[0065] (1) Mask the first set proportion of the selected subsequences and replace it with the set special marker [MSK];

[0066] The first set ratio can be 70%-80%, that is, 70%-80% of the 15% subsequences are selected from the total sequence data;

[0067] (2) Replace the subsequences of the second set proportion in the selected subsequences with the subsequences in the FCS dictionary;

[0068] For example, 10% of the subsequences can be masked by replacing them with subsequences from the FCS dictionary;

[0069] Wherein, the sum of the first set ratio and the second set ratio is less than 1;

[0070] (3) The remaining input sequence remains unchanged;

[0071] The masking strategy described above can provide effective input for subsequent context-sensitive feature learning in the BERT model. The masked input sequence The structure can be represented as:

[0072]

[0073] [CLS] is a classification marker, and [SEP] is a sequence separator used to distinguish between drug sequences and target sequences. Given the input sequence, The sequence is masked;

[0074] Step S125: Input the subsequence after the masking operation into the BERT model, extract the feature representation of the subsequence, adjust the parameters of the BERT model by adjusting the size of the loss function, and iterate the training until the convergence condition is met to obtain the trained BERT model.

[0075] During BERT model training, the Adam optimizer can optionally be used to optimize the model parameters. A specific parameter setting example is that the learning rate can be set to 1×102. −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.

[0076] Optionally, the training process of the BERT model employs a masked subsequence prediction task, and its loss function can be expressed as:

[0077] ;

[0078] in, Let represent the i-th subsequence that is masked, and s≠i represent all other subsequences except the i-th subsequence. p(s) i |s≠i) represents the masked subsequence predicted by the BERT model. The probability of.

[0079] Unlike the traditional BERT method, which only uses [CLS] to label the corresponding features, this implementation preserves the feature representations of all subsequences through masking during the training process to obtain richer sequence information.

[0080] It should be noted that the pre-training process of the BERT model described above only needs to be performed once. After obtaining a pre-trained model with good transferability, it can be used for feature extraction in subsequent tasks. This pre-trained model can serve as a general semantic feature extractor, suitable for representation learning and downstream task modeling of different drug target datasets.

[0081] In some embodiments, the method of encoding the drug target sub-sequence after word segmentation and using a pre-trained BERT model to extract the features of the drug target sub-sequence includes the following steps:

[0082] Step 121: Encode the drug target sub-sequence after word segmentation and input it into the trained BERT model to extract the sub-sequence embedding vectors of drug and target respectively, and obtain the feature representation at the word segmentation level;

[0083] Unlike the traditional BERT method, which only uses [CLS] to label the corresponding features, this implementation retains the feature representations of all subsequences to obtain richer sequence information. The feature dimension output by the BERT model can be represented as (batch_size, Lmax, de), where de is the feature dimension of the subsequence. Optionally, de is set to 128.

[0084] Step 122: Divide the extracted drug and target subsequence embedding vectors into drug and target parts, and fill and align them to obtain a feature matrix with uniform dimension, which is the extracted drug and target subsequence features.

[0085] Specifically, the padding alignment method can identify the maximum length of the drug subsequence embedding vector and the target subsequence embedding vector, and pad all the obtained subsequence embedding vectors with a set padding character.

[0086] In step 123, based on the position of the delimiter [SEP] explicitly inserted when inputting into the BERT model, each drug and target subsequence embedding vector is divided into a drug part and a target part.

[0087] Since the lengths of drug sequences and target sequences may differ, in order to meet the alignment requirements of the subsequent cross-attention module, the feature matrices of each are padded and aligned based on the maximum subsequence length of the drug to be identified and the target, respectively.

[0088] After padding and alignment, the dimensions of the drug subsequence feature matrix can be represented as (batch_size, Ld, de), and the dimensions of the target subsequence feature matrix can be represented as (batch_size, Lp, de), where Ld and Lp represent the maximum lengths of the drug and target subsequences within the corresponding batch, respectively. It should be noted that the longest subsequence lengths of the drug and target to be identified are variable, therefore the specific dimensions of the output feature matrices will differ.

[0089] Further, in step 2, a two-dimensional molecular graph of the drug and a three-dimensional structural graph of the target are constructed. The graph structure features of the drug and the target are extracted using a graph-based Transformer model, including the following steps:

[0090] Step 21: Construct a two-dimensional molecular graph based on the SMILES sequence of the drug molecule. The nodes in the two-dimensional molecular graph represent the individual atoms in the drug molecule, and the edges represent the chemical bond connections between the atoms.

[0091] Optionally, drug SMILES sequence data can be obtained directly from publicly available drug-target interaction datasets such as Davis, KIBA, and BIOSNAP, or retrieved from publicly available drug databases such as DrugBank and ChEMBL based on the drug identification information contained therein.

[0092] Specifically, in this embodiment, the chemical information processing tool RDKit can be used to analyze the SMILES sequence and generate a standard molecular diagram structure.

[0093] Optionally, the node features of the drug's two-dimensional molecular graph may include chemical attributes such as atom type, atomic valence state, hybrid orbital type, aromaticity marker, and number of hydrogen atoms; edge features may include information such as chemical bond type (e.g., single bond, double bond, triple bond), whether it is an aromatic bond, conjugation status, and whether it is located in a ring structure.

[0094] 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 nodes and their neighborhoods, a high-dimensional graph structure feature representation of the drug is obtained, i.e., the graph structure feature of the drug.

[0095] Furthermore, the graph-based Transformer model includes multiple stacked Transformer encoders, with each layer of Transformer encoder connected to multiple attention heads; each layer's output is connected to a Dropout regularization module with a regularization probability set to 0.2.

[0096] The graph-based Transformer model is stacked with Lg layers of Transformer encoders, each layer containing hg attention heads, and the node feature dimension is set to dg. In this embodiment, optionally, Lg = 6, hg = 8, and dg = 128.

[0097] In the Transformer model based on graph structure, the first... The feature update method for the i-th node of the layer can be represented as:

[0098] ;

[0099] in, Let N(i) represent the feature representation of the i-th node in the l-th layer, and let N(i) represent the set of neighboring nodes of the i-th node.

[0100] After processing by the above graph-based Transformer model, the final dimension of the graph features of each batch of drug molecules can be represented as (batch_size, nd, dg), where nd represents the maximum number of atomic nodes in the drug molecules within the corresponding batch.

[0101] Step 23: Construct a target graph structure based on the three-dimensional structural data of the target. Nodes in the target graph structure represent amino acid residues, and edges represent the spatial interaction relationships between residues.

[0102] Optionally, the three-dimensional structural data of the target can be obtained from public databases such as the protein database PDB or the AlphaFold database, and residue-level graph structural representations can be obtained using existing graph construction methods.

[0103] During the target map structure construction process, node features can be selected based on residue type, secondary structure type, physicochemical properties (such as hydrophobicity and charge), and the coordinates of the central atoms of the residues. Edge features can be defined based on the spatial distance relationship between residues. Specifically, when the minimum distance between the central atoms of two residues is less than 10 Å, an edge connection is established.

[0104] Step 24: Input the target graph structure obtained in Step 23 into the graph-based Transformer model, aggregate and transform the node features and edge features in the graph to obtain the high-dimensional graph structure feature representation of the target, that is, the graph structure feature of the target.

[0105] After processing by the graph-based Transformer model, the final dimension of the target graph structure feature for each batch can be represented as (batch_size, np, dg), where np represents the maximum number of residue nodes contained in the target structure within the corresponding batch.

[0106] Furthermore, in step 3, through a cross-attention mechanism, the sub-sequence features and graph structure features of the drug and target are fused to obtain a multimodal fusion feature representation at the drug atomic level and the target structural level, including the following steps:

[0107] Step 31: For drug subsequence features and drug graph structure features, map the drug graph structure features as query vectors and map the drug subsequence features as key vectors and value vectors, perform cross-attention operation to obtain the atomic-level multimodal fusion feature representation of drugs;

[0108] 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 scaling dot product attention mechanism is used for feature fusion.

[0109] like Figure 2 As shown, in this embodiment, the drug graph structural features are linearly mapped to serve as the query vector, and the drug subsequence features are linearly mapped to serve as the key vector and value vector, respectively. All three types of vectors are 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 as follows:

[0110] ;

[0111] 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 drug atomic-level multimodal fusion feature representation is (batch_size, nd, dg).

[0112] Step 32: Based on the target subsequence features and target graph structure features, map the target graph structure features as the query vector, and map the target subsequence features as the key vector and value vector, perform cross-attention operation to obtain the target structure-level multimodal fusion feature representation;

[0113] Specifically, the target subsequence features extracted in step 1 and the target map structure features extracted in step 2 are input into the cross-attention module, and the scaling dot product attention mechanism is used for feature fusion.

[0114] like Figure 3 As shown, in this embodiment, the target map structural features are linearly mapped and used as the query vector, while the target subsequence features are linearly mapped and used as the key vector and value vector, respectively. The fusion calculation process is similar to that in step 31. The calculated multimodal fusion feature representation matrix of target residue levels within each batch has dimensions (batch_size, np, dg).

[0115] Furthermore, in step 4, a bidirectional collaborative attention mechanism is employed to interactively fuse the obtained multimodal fusion feature representations at the drug atomic level and target structural level, including the following steps:

[0116] Step 41: Input the drug atomic-level multimodal fusion feature representation and target structural-level multimodal fusion feature representation obtained in Step 3 into the bidirectional collaborative attention module. Based on the set learnable matrix, perform bilinear interaction scoring on the drug and target features to obtain the initial interaction matrix. ;

[0117] The atomic-level multimodal characteristics of a drug are represented as follows: ;

[0118] The target residue-level multimodal features are represented as follows: ;

[0119] Furthermore, bilinear interactive scoring uses a learnable matrix. Projecting and matching drug and target features yields an initial interaction matrix:

[0120] ;

[0121] in:

[0122] ;

[0123] in, This represents the feature vector of the i-th drug atom, with the superscript T indicating transpose; This represents the feature vector of the j-th target residue. This represents the interaction matching score between the i-th drug atom and the j-th target residue;

[0124] In this embodiment, a learnable parameter matrix is ​​used. It achieves learnable feature transformation capabilities, enabling the learning of linear mapping relationships between different dimensions; The parameter matrix can be initialized with a normal distribution and updated during training using gradient descent.

[0125] Step 42: Based on the initial interaction matrix Calculate the first attention weight matrix of the drug on the target respectively. and the second attention weight matrix of the target to the drug ;

[0126] The first attention weight matrix of the drug on the target The construction method is as follows: Fix the drug node i, and perform weighted normalization on all target nodes j to measure which target nodes the drug node focuses on more. That is, adjust the interaction matrix... Calculate the element in the i-th row:

[0127] ;

[0128] The second attention weight matrix of the target to the drug The construction method is as follows: Fix the target node j, and perform weighted normalization on all drug nodes i to measure which drug nodes the target node focuses on. That is, adjust the interaction matrix... Calculate the element in the j-th column:

[0129] ;

[0130] in, Represents the matrix Perform exponential calculations with base e;

[0131] Step 43: Perform element-wise multiplication and normalization on the elements of the first attention weight matrix and the second attention weight matrix to obtain the final bidirectional collaborative attention weights. The formula is as follows:

[0132] ;

[0133] Step 44: Based on bidirectional collaborative attention weights The drug and target node features are weighted and fused to obtain the joint feature vector of each drug-target pair. ;

[0134] , ;

[0135] in, , All of these are learnable parameters.

[0136] Furthermore, the joint feature vector Input the multilayer perceptron model and perform regression prediction of drug-target binding affinity to obtain the prediction results of drug-target binding affinity;

[0137] In the above-described scheme of this embodiment, the attention weights of the drug to the target and the target to the drug are calculated separately, and the two sets of attention weights are multiplied and normalized element-wise to obtain a consistent interaction weight matrix that takes into account both perspectives. This allows for the accurate screening of key atom-residue pairs with bidirectional recognition between the drug and the target. Compared to traditional unidirectional attention mechanisms that only model interaction relationships from a single direction, the bidirectional collaborative attention mechanism proposed in this embodiment introduces a cross-screening strategy based on a bidirectional collaborative perspective. This effectively filters out unidirectional bias and false pairings with low confidence, strengthens the key biological connections, achieves accurate capture of key atom-residue pairs, and improves the model's predictive performance.

[0138] The overall model that implements the above steps includes the BERT model, a graph-based Transformer model, a cross-attention module, a bidirectional collaborative attention module, and a multilayer perceptron model. During the training of this overall model, the mean squared error between the predicted affinity value and the actual affinity value is used as the loss function, defined as follows:

[0139] ;

[0140] in, Let i be the true affinity value of the i-th sample. Let be the predicted affinity value of the i-th sample, and n represent the number of samples in the current batch.

[0141] During model training, the parameters of both the BERT pre-training module and the graph-based Transformer model participate in joint optimization without freezing or staged fine-tuning. The entire model is trained end-to-end using the aforementioned loss function.

[0142] To illustrate the effectiveness of the method in this embodiment, experimental verification was conducted to evaluate the performance of the method in the drug target interaction prediction task.

[0143] During model training, in this embodiment, the model performance is evaluated using the test set obtained in step 2 after each training round. The model performance is continuously recorded during the evaluation process, and the model with the best performance on the test set is selected as the final output model for subsequent comparison and application analysis.

[0144] Model training employed the Adam optimizer for parameter optimization and used differentiated learning rate settings. The learning rate for the pre-trained BERT module was set to 1×10−5, while the learning rates for the remaining modules (including the graph-based Transformer model, cross-attention module, bidirectional collaborative attention module, and multilayer perceptron model) were set to... .

[0145] To further improve model training stability, a learning rate decay strategy is introduced. When model performance shows no significant improvement after 50 consecutive training epochs, the current learning rate is automatically reduced to 0.8 times its original value until the minimum learning rate is reached. .

[0146] After the model training was completed, based on the evaluation results of the test set, a visual analysis was performed on the overall performance trend of the model in this embodiment throughout the training process. Figures 4(a) to 4(d) show four key metrics: Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Concordance Index (CI), and Corrected Coefficient of Judgment (CQJ). The curve showing the change of training epochs.

[0147] Experimental results show that as the number of training rounds increases, the MSE and RMSE indices continuously decrease, while the CI and RMSE indices decrease. The indicators gradually rise and tend to stabilize, indicating that the method proposed in this embodiment can stably learn the deep interaction relationship between drugs and targets, and has good training effect and generalization ability.

[0148] To further verify the predictive advantages of the method in this embodiment in practical applications, the above four indicators were used to compare its performance with that of 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.

[0149] The GraphDTA method was adopted from 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.”

[0150] Table 1. Comparison of drug target affinity prediction performance between the method in this embodiment and the GraphDTA method;

[0151]

[0152] Experimental results show that the proposed method, based on the proposed multimodal feature fusion and combined with a bidirectional collaborative attention mechanism for drug target interaction modeling, outperforms the existing GraphDTA method in all key performance indicators, demonstrating stronger predictive power and generalization performance. In particular, the mean squared error (MSE) is significantly reduced from 0.213 in the GraphDTA method to 0.193 in this method, further validating the effectiveness and advancement of the proposed method in drug target affinity prediction.

[0153] Example 2

[0154] Based on Example 1, this example provides a drug target interaction prediction system based on multimodal feature fusion, including:

[0155] The subsequence feature extraction module is configured to perform word segmentation and encoding on the acquired drug sequence to be identified and the target sequence, and extract the subsequence features of the drug and the target respectively.

[0156] 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 the target respectively through a graph structure-based Transformer model;

[0157] The cross-attention module is configured to fuse the sub-sequence features and graph structure features of the drug and the target through the cross-attention mechanism to obtain multimodal fusion feature representations at the atomic level of the drug and the structural level of the target.

[0158] The bidirectional collaborative attention module is configured to use a bidirectional collaborative attention mechanism to calculate the attention weights of the drug to the target and the target to the drug in two directions, respectively. It then performs interactive fusion of the obtained multimodal fusion feature representations at the atomic level of the drug and the structural level of the target to obtain the predicted binding affinity between the drug and the target.

[0159] Furthermore, the bidirectional collaborative attention module includes:

[0160] The interactive scoring module is configured to input the drug's atomic-level multimodal fusion feature representation and the target's structural-level multimodal fusion feature representation into the bidirectional collaborative attention module. Based on a set learnable matrix, it performs bilinear interactive scoring on the drug and target features to obtain an initial interaction matrix. ;

[0161] The first calculation module is configured to be based on the initial interaction matrix. Calculate the first attention weight matrix of the drug on the target respectively. and the second attention weight matrix of the target to the drug ;

[0162] The second calculation module is configured to perform element-wise multiplication and normalization operations on the elements of the first attention weight matrix and the second attention weight matrix to obtain the final bidirectional collaborative attention weights. ;

[0163] The weighted fusion module is configured to be based on bidirectional collaborative attention weights. The drug and target node features are weighted and fused to obtain the joint feature vector of each drug-target pair. .

[0164] It should be noted that each module in this embodiment corresponds one-to-one with each step in embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0165] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0166] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this 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 without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting drug target interaction based on multi-modal feature fusion, characterized in that, The method comprises the following steps: The acquired drug sequence to be identified and the target sequence are segmented and coded, and the subsequence features of the drug and the target are extracted respectively; A two-dimensional molecular graph of the drug and a three-dimensional structure graph of the target are constructed, and the graph structure features of the drug and the target are extracted respectively by a graph structure-based Transformer model; Through a cross-attention mechanism, the subsequence features and the graph structure features of the drug and the target are fused to obtain multi-modal fusion feature representations of the drug at the atomic level and the target at the structure level, including the following steps: For the drug subsequence features and the drug graph structure features, the drug graph structure features are mapped as query vectors, the drug subsequence features are mapped as key vectors and value vectors, cross-attention operation is performed, and drug atomic-level multi-modal fusion feature representations are obtained; For the target subsequence features and the target graph structure features, the target graph structure features are mapped as query vectors, the target subsequence features are mapped as key vectors and value vectors, cross-attention operation is performed, and target structure-level multi-modal fusion feature representations are obtained; A bidirectional collaborative attention mechanism is used to calculate the attention weights of the drug to the target and the target to the drug in two directions respectively, and the obtained multi-modal fusion feature representations of the drug at the atomic level and the target at the structure level are interactively fused, including the following steps: The drug atom-level multi-modal fusion feature representation and the target structure-level multi-modal fusion feature representation are scored by a set of learnable matrices based on a bilinear interaction to obtain an initial interaction matrix ; Based on the initial interaction matrix respectively calculating a first attention weight matrix of the drug to the target and a second attention weight matrix of the target to the drug; The elements in the first attention weight matrix and the second attention weight matrix are multiplied element by element and normalized to obtain the final bidirectional collaborative attention weight; Based on the bidirectional collaborative attention weight, the drug and target node features are weighted and fused to obtain a joint feature vector of each drug-target pair ; The binding affinity prediction result of the drug and the target is obtained. 2.The method of predicting drug target interaction based on multi-modal feature fusion according to claim 1, wherein, The method for extracting subsequence features of the drug and the target comprises the following steps: The acquired drug sequence to be identified and the target sequence are segmented and coded by using a frequent reserved subsequence algorithm; The drug subsequence coding representation and the target subsequence coding representation after segmentation coding are input into the trained BERT model to extract the subsequence features of the drug and the target respectively.

3. The drug target interaction prediction method based on multi-modal feature fusion according to claim 2, wherein: The method for extracting subsequence features of the drug and the target by inputting the drug sequence coding representation and the target sequence coding representation after segmentation coding into the trained BERT model comprises the following steps: The drug target subsequence coding representation after segmentation coding is input into the trained BERT model to extract the subsequence embedding vectors of the drug and the target respectively; The extracted subsequence embedding vectors of the drug and the target are divided into a drug part and a target part, and are filled and aligned respectively to obtain feature matrices with uniform dimensions, i.e., the extracted subsequence features of the drug and the target. 4.The method of predicting drug target interaction based on multi-modal feature fusion according to claim 1, wherein, A two-dimensional molecular graph of the drug and a three-dimensional structure graph of the target are constructed, and the graph structure features of the drug and the target are extracted respectively by a graph structure-based Transformer model, including the following steps: A two-dimensional molecular graph is constructed for the SMILES sequence of the drug molecule; The obtained two-dimensional molecular graph is input into the graph structure-based Transformer model for feature extraction, the features of the nodes and their neighborhoods are aggregated and transformed, and the graph structure features of the drug are obtained; The target graph structure is constructed based on three-dimensional structure data of the target, and nodes in the target graph structure represent amino acid residues, and edges represent spatial interaction relationships between residues; The obtained target graph structure is input into the graph structure-based Transformer model to aggregate and transform node features and edge features in the target graph structure to obtain graph structure features of the target. 5.The method of predicting drug target interaction based on multi-modal feature fusion according to claim 4, wherein, The graph structure-based Transformer model includes a plurality of stacked Transformer encoders, and a plurality of attention heads are arranged at the connection of each Transformer encoder; and a Dropout regularization module is arranged at the output end of each layer. 6.The method of predicting drug target interaction based on multi-modal feature fusion according to claim 1, wherein, The joint feature vector is obtained An input multi-layer perception model is used to make a regression prediction of the drug-target binding affinity, and a prediction result of the binding affinity of the drug and the target is obtained.

7. A drug target interaction prediction system based on multi-modal feature fusion, characterized in that, It comprises: The subsequence feature extraction module is configured to tokenize and encode the obtained to-be-identified drug sequence 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 structure graph of the target, and extract graph structure features of the drug and the target by the graph structure-based Transformer model, respectively; The cross-attention module is configured to fuse the subsequence features and the graph structure features of the drug and the target by a cross-attention mechanism to obtain multi-modal fusion feature representations of the drug at the atomic level and the target at the structure level, including the following steps: For the drug subsequence features and the drug graph structure features, the drug graph structure features are mapped as query vectors, and the drug subsequence features are mapped as key vectors and value vectors, and cross-attention operation is performed to obtain multi-modal fusion feature representations of the drug at the atomic level; For the target subsequence features and the target graph structure features, the target graph structure features are mapped as query vectors, and the target subsequence features are mapped as key vectors and value vectors, and cross-attention operation is performed to obtain multi-modal fusion feature representations of the target at the structure level; The bidirectional collaborative attention module is configured to adopt a bidirectional collaborative attention mechanism to calculate attention weights in two directions of the drug to the target and the target to the drug, respectively, and interactively fuse the obtained multi-modal fusion feature representations of the drug at the atomic level and the target at the structure level, including the following steps: The drug atom-level multi-modal fusion feature representation and the target structure-level multi-modal fusion feature representation are scored by a set of learnable matrices based on a bilinear interaction to obtain an initial interaction matrix ; based on the initial interaction matrix respectively calculating a first attention weight matrix of the drugs to the targets, and a second attention weight matrix of the targets to the drugs; The elements in the first attention weight matrix and the second attention weight matrix are multiplied and normalized to obtain the final bidirectional collaborative attention weight; Based on the bidirectional collaborative attention weight, the drug and target node features are weighted and fused to obtain a joint feature vector of each drug-target pair ; The binding affinity prediction result of the drug and the target is obtained.

8. The multi-modal feature fusion based drug target interaction prediction system of claim 7, wherein, The bidirectional collaborative attention module comprises: The interaction scoring module is configured to perform a bilinear interaction scoring on the drug atom-level multi-modal fusion feature representation and the target structure-level multi-modal fusion feature representation based on a set learnable matrix, to obtain an initial interaction matrix ; a first computing module configured to compute, based on the initial interaction matrix a first attention weight matrix of the drugs on the targets and a second attention weight matrix of the targets on the drugs, respectively; The second calculation module is configured to multiply and normalize the elements in the first attention weight matrix and the second attention weight matrix to obtain the final bidirectional collaborative attention weight; and The binding affinity prediction result of the drug and the target is obtained. The weighted fusion module is configured to perform weighted fusion on the drug and target node features based on the bidirectional collaborative attention weights to obtain a joint feature vector of each drug-target pair .

Images