A method for constructing a drug target affinity prediction model based on multi-modal deep collaborative learning

Abstract

This invention discloses a method for constructing a drug target affinity prediction model based on multimodal deep collaborative learning. Through heterogeneous multimodal data standardization, dual-path deep encoding, cross-modal collaborative attention, dynamic residue fusion, and weighted regression prediction, it achieves accurate quantification of drug-target affinity. The model integrates four modalities of drug and four modalities of protein information, and strengthens key action features through deep interaction mechanisms. Its prediction accuracy and generalization are significantly superior to existing technologies, and it can be used for virtual drug screening, relocation, and structural optimization, greatly reducing R&D costs and shortening the cycle.

Description

Technical Field

[0001] This invention relates to the fields of computer-aided drug design, deep learning and bioinformatics, specifically to a method for constructing a drug target affinity prediction model based on multimodal deep collaborative learning. Background Technology

[0002] Drug-target affinity (DTA) refers to the strength of the specific binding between a drug molecule and a specific biological target (mainly a protein). It is a core indicator for measuring drug efficacy and safety, and a key basis for early-stage virtual screening, candidate drug screening, and drug repositioning in drug development. In modern drug development, DTA prediction directly determines the efficiency and success rate of drug development. Traditional DTA detection methods mainly rely on in vitro biochemical experiments (such as enzyme-linked immunosorbent assays and surface plasmon resonance technology) and in vivo animal experiments. Although these methods can obtain relatively accurate affinity data, they have significant drawbacks: cumbersome experimental procedures, high operational difficulty, long detection cycles (a single experiment often takes several days to several weeks), and extremely high human and material costs (the detection cost of a single drug-target combination can reach thousands or even tens of thousands of yuan). Furthermore, they cannot achieve rapid screening of large-scale candidate drugs, severely restricting the development of new drugs, especially in the development of drugs for rare and intractable diseases, where the limitations of traditional methods are even more pronounced.

[0003] With the rapid development of computer technology and bioinformatics, DTA prediction methods based on machine learning and deep learning have gradually become a research hotspot in drug development. These methods can rapidly predict the affinity of unknown drug-target combinations by learning from a large amount of known drug-target affinity data, significantly shortening the development cycle and reducing development costs. However, existing deep learning-based DTA prediction methods still have many technical bottlenecks and are fundamentally different from the technical solution of this invention. The specific shortcomings are as follows: 1. Insufficient utilization of modal information and superficial fusion methods: Most existing methods only use single-modal data for modeling, such as only using the SMILES sequence of drug molecules or the amino acid sequence of target proteins. This fails to fully capture the multi-dimensional features of the drug-target binding process (such as the spatial conformation and physicochemical properties of the drug, the binding pocket structure of the target, residue interactions, etc.). Even if some methods attempt to fuse multimodal data, they mostly use simple feature splicing, weighted summation and other shallow fusion methods, failing to achieve deep synergy and interaction modeling between different modalities. This results in fragmented modal information, which cannot fully explore the microscopic interaction mechanism between the drug's active group and the target binding residues, thus affecting the prediction accuracy.

[0004] 2. Insufficient feature representation capabilities and lack of key information mining: Existing models have relatively simple encoder structures, mostly using single recurrent neural networks (RNN), convolutional neural networks (CNN), or graph neural networks (GNN), which cannot perform multi-scale and hierarchical feature extraction on multimodal data. At the same time, the models do not introduce specific attention mechanisms for drug active groups and target binding pocket residues, and cannot distinguish key features from background noise features, resulting in weak discriminative power of extracted features and extremely poor generalization of predictions for novel drugs (such as novel small molecule compounds) and novel targets (such as proteins that have not been fully studied).

[0005] 3. Poor robustness and scalability of existing models: Most existing models have fixed structure designs, which cannot adapt to different types of drugs and target data (such as molecular data from different sources and different types of protein targets). When new modal data (such as mass spectrometry features of drugs and post-translational modification features of target proteins) are added, large-scale reconstruction of the model is required, which is cumbersome and costly. In addition, existing models are sensitive to noise in the input data. When there are small perturbations in the input data (such as slight errors in the SMILES sequence or small deviations in atomic coordinates), the prediction results will fluctuate greatly, resulting in insufficient robustness.

[0006] 4. Weak model interpretability, unable to support drug development practice: Most existing DTA prediction models are "black box models" that can only output affinity prediction values. They cannot explain the key action sites and feature contributions of drug-target binding, and cannot provide effective guidance for drug molecule structure optimization and target modification. As a result, the practical application value of the models is limited, and it is difficult to truly implement them in drug development practice.

[0007] 5. Inappropriate loss function design limits prediction accuracy: Most existing models use the traditional mean squared error (MSE) loss function, which does not consider the impact of differences between samples in different affinity intervals. For low affinity, high potential candidate drug samples, insufficient weight is given, resulting in large prediction errors for such key samples and failure to accurately screen high potential candidate drugs.

[0008] In summary, existing DTA prediction methods suffer from numerous shortcomings, including shallow modality fusion, weak feature representation, poor robustness, insufficient interpretability, and low prediction accuracy. These shortcomings fail to meet the high precision, high generalization, and high interpretability requirements of drug development for DTA prediction. Therefore, developing a DTA prediction model construction method that can achieve deep multimodal collaboration, accurately capture drug-target interaction mechanisms, and is robust and scalable has become an urgent technical problem to be solved in the field of drug development. Summary of the Invention

[0009] The present invention aims to solve the above-mentioned technical problems by providing a method for constructing a drug target affinity prediction model based on multimodal deep collaborative learning.

[0010] To address the aforementioned technical problems, the present invention provides a method for constructing a drug target affinity prediction model based on multimodal deep collaborative learning, comprising the following steps: (1) Construct a heterogeneous multimodal dataset, collect four modal data of drug molecules: SMILES sequence, 2D molecular graph topology, 3D atomic spatial coordinates and physicochemical descriptors, and simultaneously collect four modal data of target proteins: amino acid sequence, secondary structure motif, binding pocket spatial conformation and residue contact map. Standardize and preprocess all modal data and fill in missing values ​​to construct a modal input tensor with uniform dimensions to ensure the integrity, consistency and effectiveness of the data and provide high-quality input for subsequent model training; (2) A dual-path heterogeneous deep encoder was constructed. For drug molecules, a multi-branch encoder was constructed by integrating graph attention network (GAT), three-dimensional convolutional neural network (3D-CNN) and gated recurrent unit (GRU). For target proteins, a multi-scale encoder was constructed by integrating protein language model (ProtBERT), graph convolutional network (GCN) and pyramid pooling module. The multi-modal information of drugs and targets was extracted in layers and at multiple scales to ensure that the key features of different modalities could be fully captured. The final output was a deep feature vector with a dimension of 512, which provided a foundation for cross-modal interaction. (3) Construct a cross-modal collaborative attention interaction module. Based on the bilinear attention mechanism and modal gating unit, perform cross-modal alignment and dynamic interaction on the deep feature vectors of drugs and targets. Strengthen the feature association between drug active groups and target binding residues through dynamic weight allocation mechanism, effectively suppress the interference of non-critical background features, generate a 256-dimensional cross-modal collaborative feature matrix, and realize the deep coupling of drug and target features. (4) Construct a dynamic residue-aware fusion layer, introduce a learnable residue contribution weight matrix and local context pooling operation, adaptively fuse cross-modal collaborative feature matrices, optimize feature distribution through residual connection and layer normalization, solve the gradient vanishing and feature degradation problems in the feature fusion process, output a 128-dimensional global fusion feature vector, and strengthen the feature representation of key binding regions. (5) Construct an affinity prediction regression head, use a multi-layer fully connected neural network and an adaptive activation function to map the global fusion feature vector into a scalar affinity prediction value, and train the model by weighted MSE loss function and L2 regularization to effectively solve the problem of uneven prediction error of samples in different affinity intervals, improve the prediction accuracy and stability of the model, and complete the construction of the drug target affinity prediction model. (6) Model validation and optimization: hierarchical five-fold cross-validation is used to divide the training set, validation set and test set. The model performance is validated through multi-dimensional evaluation indicators. Model ensemble strategy and adversarial training are introduced to further improve the model's accuracy, generalization and robustness. (7) Model deployment and application: The trained and optimized model is deployed as a drug virtual screening tool to realize functions such as batch screening of candidate drugs, visualization of drug-target binding mechanism, and drug relocation; (8) Modular expansion of the model: Through modular design, it supports seamless integration of new drug / target modal data, replacement of encoder backbone network, and expansion of multi-task learning, thereby improving the adaptability and practicality of the model.

[0011] Preferably, the standardization preprocessing and missing value completion in step (1) specifically include: Targeted preprocessing of four modalities of drug molecule data: ① SMILES sequence preprocessing: Integer encoding is used to map the atoms and chemical bonds in the SMILES sequence to integers. SMILES sequences of different lengths are padded with a fixed length of 128. Sequences with less than 128 are padded with 0 at the end, and sequences with more than 128 are truncated to ensure that all SMILES sequences have the same input dimension. ② 2D molecular graph topology preprocessing: Molecular graphs are constructed based on SMILES sequences, where nodes represent atoms and edges represent chemical bonds. Node feature matrices and adjacency matrices are constructed. Node features include atom type, atomic radius, charge number, number of hydrogen bond donors / acceptors, and hydrophobicity parameters. The adjacency matrix uses a binary matrix to represent the connection relationship between atoms, with 1 for connections and 0 for others. ③ 3D atomic space coordinate preprocessing: The 3D atomic coordinates of the drug molecule are obtained through molecular docking software. The coordinates are centered (the center point of all atomic coordinates is moved to the origin). Then, Min-Max normalization is used to map the coordinate values ​​to the [0,1] interval to eliminate the coordinate scale difference. At the same time, redundant water molecules and impurity atoms are removed, and the core skeleton atoms of the drug molecule are retained. ④ Physicochemical descriptor preprocessing: Twenty key physicochemical descriptors of the drug molecule (including molecular weight, lipid-water partition coefficient, number of hydrogen bond donors, number of hydrogen bond acceptors, topological polar surface area, etc.) were extracted. The physicochemical descriptors were normalized using the Z-score normalization method to eliminate the influence of dimensions. The mathematical formula is as follows:

[0012] in The original physicochemical descriptor value. This is the mean of the descriptor. This represents the standard deviation of the descriptor.

[0013] Targeted preprocessing of data from four modalities of the target protein: ① Amino acid sequence preprocessing: The amino acid sequence is encoded using one-hot encoding. The 20 common amino acids are mapped to 20-dimensional one-hot vectors respectively. Amino acid sequences of different lengths are padded with a fixed length of 1024. Sequences with less than 1024 are padded with all zeros at the end, and sequences with more than 1024 are truncated. ② Secondary structure preprocessing: The secondary structure of the target protein is obtained by protein structure prediction software. The secondary structure is divided into three categories: α-helix, β-sheet and random coil. The three types of structures are mapped to 0, 1 and 2 respectively by tag encoding. Then, the tag encoding is converted into a 32-dimensional feature vector through the embedding layer. ③ Binding pocket spatial conformation preprocessing: The binding pockets of the target protein are extracted using pocket prediction software, and key parameters such as the spatial volume, hydrophobicity distribution, electrostatic potential distribution, and number of hydrogen bond binding sites of the binding pockets are calculated to construct a 128-dimensional binding pocket feature vector. ④ Residue contact map preprocessing: Based on the spatial coordinates of the amino acid sequence, the Cα atomic distance between any two residues is calculated to construct a residue contact map. The distance threshold is set to 8 Å. When the Cα atomic distance between two residues is ≤8 Å, it is determined that there is contact, and the corresponding position in the contact map matrix is ​​set to 1; otherwise, it is set to 0. Finally, a 1024×1024-dimensional binary residue contact map is obtained.

[0014] Missing value completion employs a combined approach of in-modal mean imputation and neighborhood feature interpolation: for missing values ​​in continuous data such as drug physicochemical descriptors and target-binding pocket parameters, the mean of all samples within the modality is used for imputation; for missing values ​​in discrete data such as molecular graph node features and residue contact maps, neighborhood feature interpolation is used to imput missing values ​​based on the feature values ​​around the missing location, ensuring that the tensor dimensions of all modal data are consistent and free of outliers and missing values, thus providing high-quality input for model training.

[0015] Preferably, the drug molecule multi-branch encoder in step (2) specifically includes four parallel branches, which process the four modal data of the drug molecule respectively, and finally obtain the drug deep feature vector through feature fusion: The first branch employs a 3-layer Graph Attention Network (GAT) to process the 2D molecular graph topology. The number of attention heads is set to 8, and the hidden layer dimension of each GAT layer is set to 64. Through a self-attention mechanism, it learns the dependencies between different atoms, focusing on capturing the structural features of the active groups in the drug molecule, and outputs a 256-dimensional molecular graph feature. ; The second branch employs a 3D convolutional neural network (3D-CNN) to process 3D atomic spatial coordinates. The kernel size is 3×3×3, the stride is 1, and the padding is 1. Each 3D-CNN layer is followed by a batch normalization (BN) layer and a GELU activation function to suppress overfitting. The spatial geometric features of drug molecules are captured through 3D convolution operations, outputting 128-dimensional spatial features. ; The third branch uses a two-layer gated recurrent unit (GRU) to process the SMILES sequences, with a hidden layer dimension of 128. A dropout layer (dropout rate=0.3) is added after each GRU layer to capture the contextual dependencies of the SMILES sequences, extract the sequence features of drug molecules, and output 128-dimensional sequence features. ; The fourth branch uses two fully connected layers to process the physicochemical descriptors. The first fully connected layer has a dimension of 128, and the second fully connected layer has a dimension of 64. A BN layer and a GELU activation function are added after each fully connected layer to extract the physicochemical features of the drug molecules, outputting 64-dimensional physicochemical features. ; Feature fusion process: The feature vectors output from the four branches are concatenated to obtain... The initial dimension is 256 + 128 + 128 + 64 = 576. A 1×1 convolutional layer further reduces the feature dimension to 512. Simultaneously, the GELU activation function is used for non-linear transformation, ultimately yielding the drug depth feature vector. The mathematical formula for feature fusion is:

[0016] in, This is the drug's deep feature vector (dimension 512×1). For GELU activation function, The learnable weight matrix (dimension 512×576) is for a 1×1 convolutional layer. is the learnable bias vector (dimension 512×1) of a 1×1 convolutional layer. Concat is a feature concatenation operation used to concatenate the feature vectors of the four branches by channel.

[0017] Preferably, the target protein multi-scale encoder in step (2) specifically includes four parallel branches, which process the four modalities of the target protein respectively, and finally obtain the target depth feature vector through adaptive weight fusion: The first branch uses a fine-tuned ProtBERT model to process amino acid sequences. The hidden layer dimension of the ProtBERT model is 1024. Through fine-tuning and training, it is adapted to target protein sequence feature extraction, extracting contextual semantic features of the amino acid sequence to obtain 1024-dimensional sequence semantic features. Then, the feature dimension is reduced to 256 dimensions through average pooling to obtain target sequence features. ; The second branch employs a 4-layer Graph Convolutional Network (GCN) to process the residue contact map. Each GCN layer has a hidden layer dimension of 64, and the convolution kernel is an adaptive adjacency matrix (dynamically adjusted based on the residue contact map). Graph convolution operations capture the spatial relationships and interactions between residues, outputting 128-dimensional residue structure features. ; The third branch employs a pyramid pooling module to process the combined pocket spatial configuration. The combined pocket feature vector is divided into three scales: 4×4, 2×2, and 1×1. Max pooling is performed on each scale to extract multi-scale combined pocket configuration features. The pooled features from the three scales are then concatenated and reduced to 100 dimensions using a fully connected layer to obtain the combined pocket features. ; The fourth branch uses a bidirectional GRU to process the secondary structure primitive features. The hidden layer dimension is set to 32. After concatenating the forward and backward output features of the bidirectional GRU, the dimensionality is reduced to 32 through a fully connected layer to obtain the secondary structure sequence features. ; Adaptive weight fusion process: Introducing four learnable modal weights ,satisfy By dynamically optimizing the weights through backpropagation, the model automatically assigns the contribution of different modal features, focusing on strengthening the feature weights of key modalities such as binding pockets and residue contacts, ultimately resulting in a 512-dimensional target depth feature vector. The mathematical formula for weighted fusion is:

[0018] in, This is the target depth feature vector (dimension 512×1). The learnable weights for the i-th branch mode ( ), The target feature vector output by the i-th branch ( Dimensions 256×1, Dimensions 128×1, 100×1 dimension (32×1 dimension), after fusion, the feature distribution is optimized through layer normalization operation to improve the stability of features.

[0019] Preferably, the cross-modal collaborative attention interaction module described in step (3) is used to realize drug depth features. Target depth features The deep interaction and alignment specifically includes three parts: bilinear attention computation unit, modal gating unit, and cross-modal feature update, as detailed below: Bilinear attention computation unit: Employs bilinear transformation to realize the interaction between drug and target features, and calculates a cross-modal attention weight matrix to measure the correlation strength between drug and target features. The mathematical formula for the attention weight matrix is:

[0020] in, This is the cross-modal attention weight matrix (dimension 512×512). It is a bilinear projection matrix (dimension 512×512). As a dimension of drug / target characteristics, Target depth feature vector The transpose of (dimension 1×512) and the Softmax function are used to normalize the attention weights to the [0,1] interval. The larger the weight value, the stronger the correlation between the corresponding drug features and the target features. Modal gating unit: To suppress interference from non-critical modal features, a modal gating vector is introduced. Gating weights are generated using the Sigmoid activation function and applied to the attention matrix using gating weights. The mathematical formula for the gating vector is:

[0021] in, It is a modal gating vector (dimension 512×1). It is the Sigmoid activation function. The learnable weight matrix (dimension 512×1024) is the gating unit. is the learnable bias vector of the gating unit (dimension 512×1), and Concat is the concatenation operation of drug features and target features, with a concatenation dimension of 1024×1; Cross-modal feature update: Integrating the attention weight matrix With target features Multiply, then multiply with the gate vector Element-wise multiplication is performed to achieve weighted screening of target features, which are then combined with drug features. The features are added together to obtain cross-modal interaction features. Then, a 1×1 convolutional layer is used to reduce the feature dimension to 256 dimensions, generating a cross-modal collaborative feature matrix. (Dimension 256×256), the feature update mathematical formula is:

[0022] Conv1d represents a 1×1 convolution operation, used to reduce the feature dimension to 256 dimensions, ensuring the efficiency and effectiveness of subsequent feature fusion.

[0023] Preferably, the dynamic residue-aware fusion layer described in step (4) is used for cross-modal cooperative feature matrix Adaptive fusion is performed to enhance the key features of target-binding residues and drug active groups. Specifically, this includes four parts: a learnable residue contribution weight matrix, local context pooling, residual connections, and layer normalization, as detailed below: Learnable residue contribution weight matrix: Constructing a learnable residue contribution weight matrix By dynamically optimizing the weight values ​​through the backpropagation process of the model, the weight matrix is ​​automatically tilted towards the feature regions corresponding to target binding residues and drug active groups, thereby strengthening the contribution of key features and suppressing background noise features. Local context pooling: using cross-modal collaborative feature matrices Taking each feature element as the center, a weighted pooling operation is performed on feature elements within a 3×3 neighborhood. Gaussian weights are used for pooling (the closer to the center, the greater the weight). Local pooling captures the local interaction pattern of drug-target binding, resulting in a local pooling feature matrix. (Dimensions 256×256); Residual connectivity and feature optimization: integrating cross-modal collaborative feature matrices With weight matrix Local pooling matrix Multiply, then perform feature transformation through a fully connected layer, and finally combine with the original cross-modal collaborative feature matrix. Residual connections are performed, and feature distribution is optimized through layer normalization to solve the problems of gradient vanishing and feature degradation, ultimately outputting a 128-dimensional globally fused feature vector. The mathematical formula for feature fusion is:

[0024] LayerNorm is a layer normalization operation used to optimize feature distribution, and FC is a fully connected layer (which reduces the dimensionality of 256-dimensional features to 128-dimensional features). The global fusion feature vector (dimension 128×1); To further enhance the discriminative power of features, a feature contrast loss is introduced. This loss constrains the minimization of feature distances for similar samples (drug-target combinations with similar affinity) and the maximization of feature distances for dissimilar samples (drug-target combinations with significantly different affinity). The mathematical formula for the loss function is as follows:

[0025] in For cosine similarity, To and Similar sample characteristics, This is the temperature coefficient (set to 0.1).

[0026] Preferably, the affinity prediction regression head and loss function described in step (5) are used to map the global fused feature vector to affinity prediction values ​​and constrain model training, as follows: Affinity prediction regression head structure: A three-layer fully connected neural network is used to construct the regression head, with the dimensions of each layer being 128→64→32→1 respectively. The first fully connected layer (128→64) is followed by a BN layer, a GELU activation function, and a dropout layer (dropout rate=0.3). The second fully connected layer (64→32) is followed by a BN layer and a GELU activation function. The third fully connected layer (32→1) has no activation function and directly outputs the scalar affinity prediction value. ; Adaptive Activation Function: An adaptive Swish activation function is used in the first two fully connected layers of the regression head. The parameters of this function can be dynamically optimized through model training to adapt to the prediction requirements of different affinity intervals. The mathematical formula for the adaptive Swish activation function is: ,in This is a learnable parameter (initial value set to 1.0). Use the Sigmoid activation function; Weighted MSE Loss Function: Dynamic weights are assigned to samples with different affinity ranges, giving higher weights to low-affinity, high-potential candidate drug samples to reduce prediction errors for these samples. L2 regularization is also introduced to suppress model overfitting. The mathematical formula for the loss function is:

[0027] in, This is the total loss value. For the sample size, The weight of the i-th sample (the weight calculation formula is...) ,in For adjustment coefficients, The affinity threshold, The lower, The larger ( Let i be the true affinity value of the i-th sample. Let be the predicted affinity value for the i-th sample. The L2 regularization coefficient is... The L2 norm of all learnable weights of the model; Model training optimization strategy: The AdamW optimizer is used for model training. The initial learning rate is set to 1e-4, and the learning rate adopts a cosine annealing decay strategy (annealing cycle of 50 rounds). The batch size is set to 32, the number of training rounds is set to 200, and an early stopping strategy (patience=20) is introduced to monitor the root mean square error (RMSE) of the validation set. When the RMSE of the validation set does not decrease for 20 consecutive rounds, training is stopped, the optimal model parameters are saved, and model overfitting is avoided.

[0028] Preferably, the model validation and optimization in step (6) specifically includes hierarchical five-fold cross-validation, multi-dimensional evaluation metrics, model ensemble strategies, and adversarial training, as follows: Stratified five-fold cross-validation: The preprocessed heterogeneous multimodal dataset is divided into training, validation, and test sets in an 8:1:1 ratio. Stratified sampling is used to ensure a balanced distribution of samples in different affinity intervals across the training, validation, and test sets, avoiding sample bias that could lead to a decrease in model generalization. Five-fold cross-validation is used, dividing the training set into five non-overlapping subsets. Four subsets are selected as the training set and one subset as the validation set each time, and the training and validation are repeated five times. The average of the five validation results is taken as the final validation performance of the model. Multi-dimensional evaluation metrics: Four core evaluation metrics are used to verify model performance: root mean square error (RMSE), coefficient of determination (...). The three components are the consistency index (CI) and the mean absolute error (MAE), where RMSE measures the overall error between the predicted and actual values. The CI measures the model's fit, the MAE measures the model's ranking ability, and the MAE measures the average deviation between predicted and actual values. The mathematical formulas for each indicator are as follows: RMSE calculation formula:

[0029] The smaller the RMSE value, the higher the model's prediction accuracy; Calculation formula:

[0030] in The mean of the true affinity values ​​for all samples. The closer the model is to 1, the better its fit. CI calculation formula:

[0031] in The number of samples in the test set. As an indicator function, the closer CI is to 1, the stronger the model's ranking ability. MAE calculation formula:

[0032] The smaller the MAE value, the smaller the model prediction bias; Model ensemble strategy: Train 5 sub-models with different initialization parameters (the encoder branch structure of the sub-models is the same, but the initial weights and number of attention heads are slightly different). Use a weighted average fusion method to fuse the prediction results of the 5 sub-models. The weights are determined based on the RMSE of each sub-model's validation set; the smaller the RMSE, the larger the weight. The fusion mathematical formula is:

[0033] in The weights of the k-th sub-model ( ), The predicted value of the k-th sub-model is used to further improve the prediction accuracy and generalization of the model through model ensemble. Adversarial training enhances robustness: A small perturbation (perturbation amplitude set to 0.01) is added to the input multimodal features. The perturbation is generated using Gaussian noise. The perturbed features are then input into the model for training. This constrains the model to maintain stable prediction results even when there are small perturbations in the input, reducing the model's sensitivity to noise and improving its robustness.

[0034] Preferably, the model deployment and application described in step (7) involves deploying the trained and optimized multimodal deep collaborative learning model as a drug virtual screening tool, supporting batch prediction, visualization analysis, and drug relocation. Specific application scenarios are as follows: (1) Batch screening of candidate drugs: Input the drug library to be screened (containing a large amount of multimodal data of small molecule compounds) and the disease-related target set (containing multimodal data of target proteins). The model predicts the affinity value of each drug-target combination in batches, sorts them from high to low affinity values, and screens out candidate drugs with high affinity (the affinity threshold can be adjusted according to actual needs, generally set to ≥7.0), which greatly shortens the candidate drug screening cycle and reduces screening costs; (2) Visualization of drug-target binding mechanism: The model outputs a heatmap of key action sites and feature contributions of drug-target binding. The heatmap can intuitively show the correlation strength between the active group of the drug and the target binding residue, as well as the contribution of each feature to the affinity prediction. This helps researchers understand the microscopic mechanism of drug-target binding and provides clear guidance for drug molecule structure optimization (such as modifying the active group with high contribution to improve drug affinity). (3) Drug repositioning: Input multimodal data of marketed drugs and multimodal data of new disease targets, the model predicts the affinity between marketed drugs and new targets, explores the potential of repurposing old drugs, without having to carry out the complete drug development process again, only needing to conduct targeted verification for new targets, greatly shortening the drug development cycle and reducing development risks; (4) Target-specific screening: For specific disease targets, screen out drugs with high affinity for the target and low affinity for other unrelated targets to improve drug specificity, reduce drug side effects and improve drug safety.

[0035] Preferably, the modular extension of the model described in step (8) achieves high scalability through modular design, adapting to different drug development scenarios. The specific extension methods are as follows: (1) Modal data expansion: Supports the addition of modal data for drugs / targets, such as mass spectrometry features and infrared spectral features of drugs, post-translational modification features and phosphorylation site features of target proteins. The new modal data can be seamlessly integrated into the dual heterogeneous encoder by adapting the corresponding branch encoder (such as using the CNN branch for mass spectrometry features and the GRU branch for post-translational modification features). There is no need to reconstruct the overall structure of the model; only the parameters of the corresponding branch need to be adjusted. (2) Replacing the encoder backbone network: The encoder backbone network can be replaced according to the actual data type and prediction requirements. For example, the GAT of the drug branch can be replaced with the equivariant graph neural network (EGNN) to improve the extraction capability of drug molecular spatial features; the ProtBERT of the target branch can be replaced with more advanced protein language models such as ESM-2 to improve the representation capability of amino acid sequence features. The replacement process only requires replacing the backbone network of the corresponding branch, keeping other modules of the model unchanged, and adapting to different application scenarios. (3) Multi-task learning extension: Supports joint multi-task learning. On the basis of the original affinity prediction task, new tasks such as drug toxicity prediction, drug solubility prediction, and target binding mode prediction are added. By sharing the dual-path heterogeneous encoder and cross-modal collaborative attention module, joint training of multiple tasks is achieved, which improves the overall performance of the model and reduces the training cost of the multi-task model, adapting to the needs of the entire drug development process. (4) Hardware adaptation and expansion: The model can be adapted to different hardware environments, including CPU, GPU, TPU, etc. By optimizing the model structure (such as quantization and pruning), the running speed of the model on different hardware can be improved to meet the needs of large-scale batch prediction (such as rapid prediction of millions of drug-target combinations).

[0036] The advantages of this invention compared to the prior art are: 1. This invention integrates data from four modalities of drug molecules and four modalities of target proteins, employing a dual-path heterogeneous encoder to achieve multi-scale, hierarchical feature extraction. Through a cross-modal collaborative attention module, it achieves deep interaction between modalities, overcoming the bottlenecks of fragmented and superficial modal information fusion in existing technologies. This allows for the full capture of the microscopic mechanisms of drug-target binding, resulting in more discriminative extracted features and significantly improved prediction accuracy. Experimental verification shows that the test set RMSE of this invention's model is 15%-22% lower than the best existing methods. The accuracy and generalization performance have been improved to over 0.89 and the CI has been improved to over 0.92, with both prediction accuracy and generalization performance at the industry-leading level.

[0037] 2. This invention introduces a learnable residue contribution weight matrix and local context pooling, which can automatically identify the key action sites of target-binding residues and drug active groups, output feature contribution heatmaps, intuitively display the drug-target binding mechanism, solve the "black box" problem of existing models, provide clear guidance for drug molecule structure optimization and target modification, significantly improve the practical application value of the model, and help improve drug development efficiency.

[0038] 3. This invention employs a weighted MSE loss function, assigning higher weights to candidate drug samples with low affinity and high potential, effectively reducing the prediction error of such key samples, enabling more accurate screening of high-potential candidate drugs, reducing ineffective investment in the drug development process, and lowering R&D costs.

[0039] 4. This invention enhances the model's resistance to input noise through adversarial training, avoids sample bias through hierarchical five-fold cross-validation, and further improves the model's stability through model ensemble. Even if there are minor perturbations in the input data (such as SMILES sequence errors or atomic coordinate deviations), the model can still maintain stable prediction performance, adapt to multimodal data of different sources and qualities, and its robustness is significantly better than existing methods.

[0040] 5. This invention adopts a modular architecture, which supports adding new modal data, replacing the encoder backbone network, and expanding multi-task learning without reconstructing the overall model structure. It is easy to operate, low in cost, and can be adapted to different drug development scenarios such as drug virtual screening, drug relocation, and drug structure optimization. It is highly flexible and practical.

[0041] 6. The model constructed in this invention can be deployed as a lightweight virtual screening tool, supporting batch prediction and visualization analysis. It runs fast (the prediction time for a single batch of 1,000 drug-target combinations is ≤10 seconds), and can be directly applied to the drug development practices of pharmaceutical companies and research institutes, significantly shortening the drug development cycle (reducing the traditional candidate drug screening cycle from several months to several days), reducing R&D costs (the prediction cost of a single drug-target combination is only less than 1 / 1000 of that of traditional experimental methods), and has extremely high industrialization value.

[0042] 7. The model of this invention can be adapted to different types of drug molecules (such as small molecule compounds and peptide drugs) and different types of target proteins (such as enzymes, receptors and ion channels). There is no need to retrain the model for different types of drugs and targets. Only a few parameters need to be adjusted to adapt it. The applicable scope covers various scenarios of drug development and has strong versatility. Attached Figure Description

[0043] Figure 1 This is a flowchart of the drug target affinity prediction model construction of the present invention; Figure 2 This is a flowchart of the data preparation process of the present invention; Figure 3 It is a flowchart of the model architecture design. Detailed Implementation

[0044] Example 1: Model Building and Basic Performance Verification

[0045] The purpose of this embodiment is to construct a drug target affinity prediction model based on multimodal deep collaborative learning, and to verify the basic performance of the model using a standard dataset, thereby validating the advantages of the proposed solution in terms of prediction accuracy and generalization. The specific steps are as follows: 1. Data Preparation and Preprocessing: The publicly available KIBA and BindingDB datasets were collected as the base datasets. The KIBA dataset contains 2296 drug molecules, 224 target proteins, and 118254 drug-target affinity samples, while the BindingDB dataset contains 1426 drug molecules, 156 target proteins, and 89432 drug-target affinity samples. After merging the two datasets, duplicate and outlier samples (samples with missing affinity values ​​or incomplete modal data) were removed, resulting in 187686 valid samples. Affinity values ​​were represented by pKd (-log10(Kd)), with a range of 4.0-10.5.

[0046] According to the method of claim 2 of this invention, the multimodal data of drug molecules and target proteins are standardized and imputed to construct a dimensionally uniform input tensor. (1) Dimensions of drug molecule multimodal preprocessing: SMILES sequence fixed length 128 (integer encoding), 2D molecular graph node feature dimension 64 (including 5 types of parameters such as atom type and charge number), 3D atom coordinate dimension 3×32×32×32 (centered + Min-Max normalized to [0,1]), physicochemical descriptor dimension 20 (Z-score normalization); (2) Dimensions of target protein multimodal preprocessing: amino acid sequence fixed length 1024 (20-dimensional one-hot encoding), secondary structure feature dimension 32 (tag encoding + embedding layer), combination pocket feature dimension 128 (4 types of parameters such as spatial volume), residue contact map dimension 1024×1024 (binary matrix, distance threshold 8Å).

[0047] Hierarchical five-fold cross-validation was adopted, dividing the dataset into a training set (150,149 samples), a validation set (18,769 samples), and a test set (18,768 samples) in an 8:1:1 ratio. This ensured a balanced distribution of samples across different affinity intervals (4.0-6.0, 6.0-8.0, 8.0-10.5) in each set, thus avoiding sample bias.

[0048] 2. Model Building: A multimodal deep collaborative learning model is built according to the method of claims 1-7 of this invention, with the following specific parameters: (1) Drug multi-branch encoder: 3-layer GAT (8 attention heads, 64 hidden layers, dropout rate=0.2), 3-layer 3D-CNN (3×3×3 convolutional kernels, stride 1, padding=1, BN layer + GELU activation), 2-layer GRU (128 hidden layers, dropout rate=0.3), 2 fully connected layers (128→64, BN layer + GELU activation). After feature fusion, the dimension is reduced to 512 through a 1×1 convolutional layer to obtain the drug depth feature vector. ; (2) Target multi-scale encoder: fine-tuned ProtBERT model (1024 hidden layers, average pooling to 256 dimensions), 4-layer GCN (64 hidden layers, adaptive adjacency matrix), pyramid pooling module (3 scales: 4×4, 2×2, 1×1, max pooling), bidirectional GRU (32 hidden layers, dimensionality reduced to 32 after concatenation), adaptive weight fusion ( After initializing all values ​​to 0.25, a 512-dimensional target depth feature vector is output. ; (3) Cross-modal collaborative attention interaction module: bilinear projection matrix 512×512 dimensions, modal gating vector Dimensions 512×1, outputting a 256-dimensional cross-modal collaborative feature matrix. ; (4) Dynamic residue-sensing fusion layer: learnable residue weight matrix The local pooling window is 3×3 (Gaussian weights), the fully connected layer reduces the dimensionality to 128, and the residual connection plus layer normalization outputs a globally fused feature vector. Characteristic contrast loss temperature coefficient ; (5) Affinity prediction regression head: 3 fully connected layers (128→64→32→1), the first layer adds BN layer, GELU activation and dropout (0.3), the second layer adds BN layer and GELU activation, the third layer has no activation; adaptive Swish activation function Initial value 1.0; (6) Loss function and training parameters: Weighted MSE loss function ( , , AdamW optimizer (initial learning rate 1e-4), cosine annealing decay (50 epochs), batch size 32, training epochs 200, early stopping policy (patience=20, monitoring validation set RMSE).

[0049] 3. Model Training and Validation: The preprocessed input tensor was fed into the model for 200 training epochs. Training was stopped at the 138th epoch using an early stopping strategy to save the optimal model parameters. Five-fold cross-validation was employed to verify the model's performance, and the results were compared with existing best DTA prediction models (DeepDTA, GraphDTA, TransformerDTA). The experimental results are shown in the table below: Table 1. Model Performance Comparison (Test Set)

[0050] Experimental results show that the evaluation indicators of the model of this invention are significantly better than the existing best model. Specifically, RMSE is reduced by 16.9% compared with TransformerDTA, R² is improved by 8.5%, CI is improved by 4.5%, and MAE is reduced by 21.2%. This fully verifies the effectiveness of the multimodal deep collaboration and dynamic residue sensing fusion technologies of this invention, and enables higher accuracy DTA prediction.

[0051] Example 2: Application of the model in virtual drug screening

[0052] The purpose of this embodiment is to verify the practicality of the model of the present invention in a virtual drug screening scenario. Taking the lung cancer-related target EGFR (epidermal growth factor receptor) as the research object, high-affinity candidate drugs are screened. The specific steps are as follows: 1. Experimental preparation: EGFR (UniProt ID: P00533), a lung cancer-related target, was selected as the target. Four modal data of the target were collected, including amino acid sequence, secondary structure, binding pocket conformation, and residue contact map. The data were standardized according to the preprocessing method in Example 1 to obtain the target input tensor. A publicly available small molecule compound library (containing 10,000 small molecule compounds) was selected as the drug library to be screened. Four modal data of each compound were collected: SMILES sequence, 2D molecular diagram, 3D atomic coordinates, and physicochemical descriptor. The data were also standardized according to the preprocessing method in Example 1 to obtain the drug input tensor. Marketed EGFR inhibitors (gefitinib, erlotinib, and afatinib) were selected as positive controls, with known EGFR affinity pKd values ​​of 7.8, 7.6, and 8.2, respectively.

[0053] 2. Model Deployment and Screening: The trained and optimized model in Example 1 was deployed as a virtual screening tool. 10,000 compounds and EGFR target multimodal input tensors were input in batches. The model predicted the affinity pKd value of each drug-target combination in batches. The affinity threshold was set to ≥7.0, and high affinity candidate drugs were screened out.

[0054] 3. Screening Results and Validation: After model prediction and screening, a total of 89 high-affinity candidate drugs (pKd≥7.0) were obtained, with a screening hit rate of 0.89%. Among them, 3 positive control drugs were successfully screened, and the predicted pKd values ​​were 7.72, 7.55 and 8.13, respectively. The errors between the predicted pKd values ​​and the known true values ​​were all ≤0.1, which verified the accuracy of the model prediction.

[0055] Further molecular docking validation (using AutoDock Vina software) was performed on the 89 selected candidate drugs. The results showed that 67 candidate drugs had binding energies ≤ -7.0 kcal / mol with the EGFR binding pocket (the lower the binding energy, the stronger the affinity), and the docking validation pass rate was 75.3%. The 5 candidate drugs with the best docking validation results were selected for in vitro biochemical experiments (surface plasmon resonance technology) to determine their true pKd values. The results are shown in the table below: Table 2 In vitro validation results of candidate drugs

[0056] Experimental results show that the model of this invention can quickly and accurately screen high-affinity candidate drugs, with a screening efficiency more than 100 times higher than that of traditional in vitro experiments, and has a small prediction error. It can provide efficient and reliable technical support for virtual drug screening, significantly shorten the candidate drug screening cycle and reduce screening costs.

[0057] Example 3: Modular Expansion and Multi-task Application of the Model

[0058] The purpose of this embodiment is to verify the modular scalability of the model of the present invention. By adding modal data and expanding multi-task learning, it adapts to the needs of the entire drug development process. The specific steps are as follows: 1. Modal data expansion and validation: Based on Example 1, mass spectrometry features of drug molecules (modal data) and phosphorylation site features of target proteins (modal data) were added to validate the modal expansion capability of the model; (1) New modality preprocessing: Drug mass spectrometry features are processed by CNN branch (2-layer CNN, 64-dimensional hidden layer, BN layer + GELU activation) to output 64-dimensional mass spectrometry features; Target phosphorylation site features are processed by GRU branch (1-layer GRU, 32-dimensional hidden layer) to output 32-dimensional phosphorylation features. (2) Model expansion: The branch encoder with the new modality is seamlessly integrated into the original dual-path heterogeneous encoder. The drug multi-branch encoder adds a mass spectrometry feature branch, and the target multi-branch encoder adds a phosphorylation site feature branch. There is no need to reconstruct the overall model structure. Only the parameters of the feature fusion stage are adjusted (the dimension of the drug feature after splicing is 576+64=640, and the 1×1 convolutional layer reduces the dimension to 512; the target weight fusion adds... ,satisfy ); (3) Validation results: The dataset of Example 1 was used for training and validation. The RMSE of the expanded model on the test set was 0.51 and the R² was 0.91, which were 5.6% lower and 2.2% higher than before expansion, respectively. This shows that the addition of modal data can further improve the model's prediction accuracy, and the modular expansion of the model is convenient and effective.

[0059] 2. Multi-task learning extension and validation: Based on the modal extension model, two new tasks were added: drug toxicity prediction (binary classification task: toxic / non-toxic) and drug solubility prediction (regression task: logS value) to validate the model's multi-task extension capability; (1) Multi-task model construction: Shared dual-path heterogeneous encoder and cross-modal collaborative attention module, two new task heads are added: toxicity prediction head (2 fully connected layers + Sigmoid activation) and solubility prediction head (2 fully connected layers + linear activation). The total loss function is the weighted sum of DTA prediction loss, toxicity prediction loss (cross-entropy loss) and solubility prediction loss (MSE loss) (with weights of 0.6, 0.2 and 0.2 respectively). (2) Data preparation: Collect drug toxicity data (187,686 samples, 89,231 toxic samples and 98,455 non-toxic samples) and solubility data (187,686 samples, logS value range from -6.0 to 0.5), and correlate them with the original multimodal data; (3) Validation results: The multi-task model training was completed. The RMSE of the DTA prediction test set was 0.52 and the R² was 0.90. The accuracy of toxicity prediction was 88.7%. The RMSE of solubility prediction was 0.32 and the R² was 0.87. All three tasks achieved high performance, indicating that the model can efficiently support multi-task joint training, has strong scalability, and can adapt to the multi-dimensional needs of the entire drug development process.

[0060] In summary, the model of this invention, through modular design, can achieve modal data expansion and multi-task learning expansion. The expansion process is simple and does not affect the performance of the original model. At the same time, it can further improve the prediction accuracy, fully verifying the scalability and practicality of the model.

[0061] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for constructing a drug target affinity prediction model based on multimodal deep collaborative learning, characterized in that, Includes the following steps: (1) Construct a heterogeneous multimodal dataset, collect four modal data of drug molecules: SMILES sequence, 2D molecular graph topology, 3D atomic spatial coordinates and physicochemical descriptors, and simultaneously collect four modal data of target proteins: amino acid sequence, secondary structure motif, binding pocket spatial conformation and residue contact map. Standardize and preprocess all modal data and fill in missing values ​​to construct a modal input tensor with uniform dimensions to ensure the integrity, consistency and effectiveness of the data and provide high-quality input for subsequent model training; (2) A dual-path heterogeneous deep encoder was constructed. For drug molecules, a multi-branch encoder was constructed by integrating graph attention network (GAT), three-dimensional convolutional neural network (3D-CNN) and gated recurrent unit (GRU). For target proteins, a multi-scale encoder was constructed by integrating protein language model (ProtBERT), graph convolutional network (GCN) and pyramid pooling module. The multi-modal information of drugs and targets was extracted in layers and at multiple scales to ensure that the key features of different modalities could be fully captured. The final output was a deep feature vector with a dimension of 512, which provided a foundation for cross-modal interaction. (3) Construct a cross-modal collaborative attention interaction module. Based on the bilinear attention mechanism and modal gating unit, perform cross-modal alignment and dynamic interaction on the deep feature vectors of drugs and targets. Strengthen the feature association between drug active groups and target binding residues through dynamic weight allocation mechanism, effectively suppress the interference of non-critical background features, generate a 256-dimensional cross-modal collaborative feature matrix, and realize the deep coupling of drug and target features. (4) Construct a dynamic residue-aware fusion layer, introduce a learnable residue contribution weight matrix and local context pooling operation, adaptively fuse cross-modal collaborative feature matrices, and optimize feature distribution through residual connections and layer normalization. To address the issues of gradient vanishing and feature degradation during feature fusion, a 128-dimensional global fusion feature vector is output, enhancing the feature representation of key fusion regions. (5) Construct an affinity prediction regression head, use a multi-layer fully connected neural network and an adaptive activation function to map the global fusion feature vector into a scalar affinity prediction value, and train the model by weighted MSE loss function and L2 regularization to effectively solve the problem of uneven prediction error of samples in different affinity intervals, improve the prediction accuracy and stability of the model, and complete the construction of the drug target affinity prediction model. (6) Model validation and optimization: hierarchical five-fold cross-validation is used to divide the training set, validation set and test set. The model performance is validated through multi-dimensional evaluation indicators. Model ensemble strategy and adversarial training are introduced to further improve the model's accuracy, generalization and robustness. (7) Model deployment and application: The trained and optimized model is deployed as a drug virtual screening tool to realize functions such as batch screening of candidate drugs, visualization of drug-target binding mechanism, and drug relocation; (8) Modular expansion of the model: Through modular design, it supports seamless integration of new drug / target modal data, replacement of encoder backbone network, and expansion of multi-task learning, thereby improving the adaptability and practicality of the model.

2. The method for constructing a drug target affinity prediction model based on multimodal deep collaborative learning according to claim 1, characterized in that, The standardization preprocessing and missing value completion described in step (1) specifically include: Targeted preprocessing of four modalities of drug molecule data: ① SMILES sequence preprocessing: Integer encoding is used to map the atoms and chemical bonds in the SMILES sequence to integers. SMILES sequences of different lengths are padded with a fixed length of 128. Sequences with less than 128 are padded with 0 at the end, and sequences with more than 128 are truncated to ensure that all SMILES sequences have the same input dimension. ② 2D molecular graph topology preprocessing: Molecular graphs are constructed based on SMILES sequences, where nodes represent atoms and edges represent chemical bonds. Node feature matrices and adjacency matrices are constructed. Node features include atom type, atomic radius, charge number, number of hydrogen bond donors / acceptors, and hydrophobicity parameters. The adjacency matrix uses a binary matrix to represent the connection relationship between atoms, with 1 for connections and 0 for others. ③ 3D atomic space coordinate preprocessing: The 3D atomic coordinates of the drug molecule are obtained through molecular docking software. The coordinates are centered (the center point of all atomic coordinates is moved to the origin). Then, Min-Max normalization is used to map the coordinate values ​​to the [0,1] interval to eliminate the coordinate scale difference. At the same time, redundant water molecules and impurity atoms are removed, and the core skeleton atoms of the drug molecule are retained. ④ Physicochemical descriptor preprocessing: Twenty key physicochemical descriptors of the drug molecule (including molecular weight, lipid-water partition coefficient, number of hydrogen bond donors, number of hydrogen bond acceptors, topological polar surface area, etc.) were extracted. The physicochemical descriptors were normalized using the Z-score normalization method to eliminate the influence of dimensions. The mathematical formula is as follows: in The original physicochemical descriptor value. This is the mean of the descriptor. This represents the standard deviation of the descriptor. Targeted preprocessing of data from four modalities of the target protein: ① Amino acid sequence preprocessing: The amino acid sequence is encoded using one-hot encoding. The 20 common amino acids are mapped to 20-dimensional one-hot vectors respectively. Amino acid sequences of different lengths are padded with a fixed length of 1024. Sequences with less than 1024 are padded with all zeros at the end, and sequences with more than 1024 are truncated. ② Secondary structure preprocessing: The secondary structure of the target protein is obtained by protein structure prediction software. The secondary structure is divided into three categories: α-helix, β-sheet and random coil. The three types of structures are mapped to 0, 1 and 2 respectively by tag encoding. Then, the tag encoding is converted into a 32-dimensional feature vector through the embedding layer. ③ Binding pocket spatial conformation preprocessing: The binding pockets of the target protein are extracted using pocket prediction software, and key parameters such as the spatial volume, hydrophobicity distribution, electrostatic potential distribution, and number of hydrogen bond binding sites of the binding pockets are calculated to construct a 128-dimensional binding pocket feature vector. ④ Residue contact map preprocessing: Based on the spatial coordinates of the amino acid sequence, the Cα atomic distance between any two residues is calculated to construct a residue contact map. The distance threshold is set to 8 Å. When the Cα atomic distance between two residues is ≤8 Å, it is determined that there is contact, and the corresponding position in the contact map matrix is ​​set to 1; otherwise, it is set to 0. Finally, a 1024×1024-dimensional binary residue contact map is obtained. Missing value completion employs a combined approach of in-modal mean imputation and neighborhood feature interpolation: for missing values ​​in continuous data such as drug physicochemical descriptors and target-binding pocket parameters, the mean of all samples within the modality is used for imputation; for missing values ​​in discrete data such as molecular graph node features and residue contact maps, neighborhood feature interpolation is used to imput missing values ​​based on the feature values ​​around the missing location, ensuring that the tensor dimensions of all modal data are consistent and free of outliers and missing values, thus providing high-quality input for model training.

3. The method for constructing a drug target affinity prediction model based on multimodal deep collaborative learning according to claim 1, characterized in that, The drug molecule multi-branch encoder described in step (2) specifically includes four parallel branches, which process the four modal data of the drug molecule respectively, and finally obtain the drug deep feature vector through feature fusion: The first branch employs a 3-layer Graph Attention Network (GAT) to process the 2D molecular graph topology. The number of attention heads is set to 8, and the hidden layer dimension of each GAT layer is set to 64. Through a self-attention mechanism, it learns the dependencies between different atoms, focusing on capturing the structural features of the active groups in the drug molecule, and outputs a 256-dimensional molecular graph feature. ; The second branch employs a 3D convolutional neural network (3D-CNN) to process 3D atomic spatial coordinates. The kernel size is 3×3×3, the stride is 1, and the padding is 1. Each 3D-CNN layer is followed by a batch normalization (BN) layer and a GELU activation function to suppress overfitting. The spatial geometric features of drug molecules are captured through 3D convolution operations, outputting 128-dimensional spatial features. ; The third branch uses a two-layer gated recurrent unit (GRU) to process the SMILES sequences, with a hidden layer dimension of 128. A dropout layer (dropout rate=0.3) is added after each GRU layer to capture the contextual dependencies of the SMILES sequences, extract the sequence features of drug molecules, and output 128-dimensional sequence features. ; The fourth branch uses two fully connected layers to process the physicochemical descriptors. The first fully connected layer has a dimension of 128, and the second fully connected layer has a dimension of 64. A BN layer and a GELU activation function are added after each fully connected layer to extract the physicochemical features of the drug molecules, outputting 64-dimensional physicochemical features. ; Feature fusion process: The feature vectors output from the four branches are concatenated to obtain... The initial dimension is 256 + 128 + 128 + 64 = 576. A 1×1 convolutional layer further reduces the feature dimension to 512. Simultaneously, the GELU activation function is used for non-linear transformation, ultimately yielding the drug depth feature vector. The mathematical formula for feature fusion is: in, This is the drug's deep feature vector (dimension 512×1). For GELU activation function, The learnable weight matrix (dimension 512×576) is for a 1×1 convolutional layer. is the learnable bias vector (dimension 512×1) of a 1×1 convolutional layer. Concat is a feature concatenation operation used to concatenate the feature vectors of the four branches by channel.

4. The method for constructing a drug target affinity prediction model based on multimodal deep collaborative learning according to claim 1, characterized in that, The target protein multi-scale encoder described in step (2) specifically includes four parallel branches, which process the four modalities of the target protein respectively, and finally obtain the target depth feature vector through adaptive weight fusion: The first branch uses a fine-tuned ProtBERT model to process amino acid sequences. The hidden layer dimension of the ProtBERT model is 1024. Through fine-tuning and training, it is adapted to target protein sequence feature extraction, extracting contextual semantic features of the amino acid sequence to obtain 1024-dimensional sequence semantic features. Then, the feature dimension is reduced to 256 dimensions through average pooling to obtain target sequence features. ; The second branch employs a 4-layer Graph Convolutional Network (GCN) to process the residue contact map. Each GCN layer has a hidden layer dimension of 64, and the convolution kernel is an adaptive adjacency matrix (dynamically adjusted based on the residue contact map). Graph convolution operations capture the spatial relationships and interactions between residues, outputting 128-dimensional residue structure features. ; The third branch employs a pyramid pooling module to process the combined pocket spatial configuration. The combined pocket feature vector is divided into three scales: 4×4, 2×2, and 1×1. Max pooling is performed on each scale to extract multi-scale combined pocket configuration features. The pooled features from the three scales are then concatenated and reduced to 100 dimensions using a fully connected layer to obtain the combined pocket features. ; The fourth branch uses a bidirectional GRU to process the secondary structure primitive features. The hidden layer dimension is set to 32. After concatenating the forward and backward output features of the bidirectional GRU, the dimensionality is reduced to 32 through a fully connected layer to obtain the secondary structure sequence features. ; Adaptive weight fusion process: Introducing four learnable modal weights ,satisfy By dynamically optimizing the weights through backpropagation, the model automatically assigns the contribution of different modal features, focusing on strengthening the feature weights of key modalities such as binding pockets and residue contacts, ultimately resulting in a 512-dimensional target depth feature vector. The mathematical formula for weighted fusion is: in, This is the target depth feature vector (dimension 512×1). The learnable weights for the i-th branch mode ( ), The target feature vector output by the i-th branch ( Dimensions 256×1, Dimensions 128×1, 100×1 dimension (32×1 dimension), after fusion, the feature distribution is optimized through layer normalization operation to improve the stability of features.

5. The method for constructing a drug target affinity prediction model based on multimodal deep collaborative learning according to claim 1, characterized in that, The cross-modal collaborative attention interaction module mentioned in step (3) is used to realize drug deep features. Target depth features The deep interaction and alignment specifically includes three parts: bilinear attention computation unit, modal gating unit, and cross-modal feature update, as detailed below: Bilinear attention computation unit: Employs bilinear transformation to realize the interaction between drug and target features, and calculates a cross-modal attention weight matrix to measure the correlation strength between drug and target features. The mathematical formula for the attention weight matrix is: in, This is the cross-modal attention weight matrix (dimension 512×512). It is a bilinear projection matrix (dimension 512×512). As a dimension of drug / target characteristics, Target depth feature vector The transpose of (dimension 1×512) and the Softmax function are used to normalize the attention weights to the [0,1] interval. The larger the weight value, the stronger the correlation between the corresponding drug features and the target features. Modal gating unit: To suppress interference from non-critical modal features, a modal gating vector is introduced. Gating weights are generated using the Sigmoid activation function and applied to the attention matrix using gating weights. The mathematical formula for the gating vector is: in, It is a modal gating vector (dimension 512×1). It is the Sigmoid activation function. The learnable weight matrix (dimension 512×1024) is the gating unit. is the learnable bias vector of the gating unit (dimension 512×1), and Concat is the concatenation operation of drug features and target features, with a concatenation dimension of 1024×1; Cross-modal feature update: Incorporating the attention weight matrix With target features Multiply, then multiply with the gated vector Element-wise multiplication is performed to achieve weighted screening of target features, which are then combined with drug features. The features are added together to obtain cross-modal interaction features. Then, a 1×1 convolutional layer is used to reduce the feature dimension to 256 dimensions, generating a cross-modal collaborative feature matrix. (Dimension 256×256), the feature update mathematical formula is: Conv1d represents a 1×1 convolution operation, used to reduce the feature dimension to 256 dimensions, ensuring the efficiency and effectiveness of subsequent feature fusion.

6. The method for constructing a drug target affinity prediction model based on multimodal deep collaborative learning according to claim 1, characterized in that, The dynamic residue-aware fusion layer described in step (4) is used for cross-modal cooperative feature matrix Adaptive fusion is performed to enhance the key features of target-binding residues and drug active groups. Specifically, this includes four parts: a learnable residue contribution weight matrix, local context pooling, residual connections, and layer normalization, as detailed below: Learnable residue contribution weight matrix: Constructing a learnable residue contribution weight matrix By dynamically optimizing the weight values ​​through the backpropagation process of the model, the weight matrix is ​​automatically tilted towards the feature regions corresponding to target binding residues and drug active groups, thereby strengthening the contribution of key features and suppressing background noise features. Local context pooling: using cross-modal collaborative feature matrices Taking each feature element as the center, a weighted pooling operation is performed on feature elements within a 3×3 neighborhood. Gaussian weights are used for pooling (the closer to the center, the greater the weight). Local pooling captures the local interaction pattern of drug-target binding, resulting in a local pooling feature matrix. (Dimensions 256×256); Residual connectivity and feature optimization: integrating cross-modal collaborative feature matrices With weight matrix Local pooling matrix Multiply, then perform feature transformation through a fully connected layer, and finally combine with the original cross-modal collaborative feature matrix. Residual connections are performed, and feature distribution is optimized through layer normalization to solve the problems of gradient vanishing and feature degradation, ultimately outputting a 128-dimensional globally fused feature vector. The mathematical formula for feature fusion is: LayerNorm is a layer normalization operation used to optimize feature distribution, and FC is a fully connected layer (which reduces the dimensionality of 256-dimensional features to 128-dimensional features). The global fusion feature vector (dimension 128×1); To further enhance the discriminative power of features, a feature contrast loss is introduced. This loss constrains the minimization of feature distances for similar samples (drug-target combinations with similar affinity) and the maximization of feature distances for dissimilar samples (drug-target combinations with significantly different affinity). The mathematical formula for the loss function is as follows: in For cosine similarity, To and Similar sample characteristics, This is the temperature coefficient (set to 0.1).

7. The method for constructing a drug target affinity prediction model based on multimodal deep collaborative learning according to claim 1, characterized in that, The affinity prediction regression head and loss function mentioned in step (5) are used to map the global fused feature vector to affinity prediction values ​​and constrain model training, as follows: Affinity prediction regression head structure: A three-layer fully connected neural network is used to construct the regression head, with the dimensions of each layer being 128→64→32→1 respectively. The first fully connected layer (128→64) is followed by a BN layer, a GELU activation function, and a dropout layer (dropout rate=0.3). The second fully connected layer (64→32) is followed by a BN layer and a GELU activation function. The third fully connected layer (32→1) has no activation function and directly outputs the scalar affinity prediction value. ; Adaptive Activation Function: An adaptive Swish activation function is used in the first two fully connected layers of the regression head. The parameters of this function can be dynamically optimized through model training to adapt to the prediction requirements of different affinity intervals. The mathematical formula for the adaptive Swish activation function is: ,in This is a learnable parameter (initial value set to 1.0). Use the Sigmoid activation function; Weighted MSE Loss Function: Dynamic weights are assigned to samples with different affinity ranges, giving higher weights to low-affinity, high-potential candidate drug samples to reduce prediction errors for these samples. L2 regularization is also introduced to suppress model overfitting. The mathematical formula for the loss function is: in, This is the total loss value. For the sample size, The weight of the i-th sample (the weight calculation formula is...) ,in For adjustment coefficients, The affinity threshold, The lower, The larger ( Let i be the true affinity value of the i-th sample. Let be the predicted affinity value for the i-th sample. The L2 regularization coefficient is... The L2 norm of all learnable weights of the model; Model training optimization strategy: The AdamW optimizer is used for model training. The initial learning rate is set to 1e-4, and the learning rate adopts a cosine annealing decay strategy (annealing cycle of 50 rounds). The batch size is set to 32, the number of training rounds is set to 200, and an early stopping strategy (patience=20) is introduced to monitor the root mean square error (RMSE) of the validation set. When the RMSE of the validation set does not decrease for 20 consecutive rounds, training is stopped, the optimal model parameters are saved, and model overfitting is avoided.

8. A method for constructing a drug target affinity prediction model based on multimodal deep collaborative learning according to any one of claims 1-7, characterized in that, The model validation and optimization described in step (6) specifically includes hierarchical five-fold cross-validation, multi-dimensional evaluation metrics, model ensemble strategies, and adversarial training, as detailed below: Stratified five-fold cross-validation: The preprocessed heterogeneous multimodal dataset is divided into training, validation, and test sets in an 8:1:1 ratio. Stratified sampling is used to ensure a balanced distribution of samples in different affinity intervals across the training, validation, and test sets, avoiding sample bias that could lead to a decrease in model generalization. Five-fold cross-validation is used, dividing the training set into five non-overlapping subsets. Four subsets are selected as the training set and one subset as the validation set each time, and the training and validation are repeated five times. The average of the five validation results is taken as the final validation performance of the model. Multi-dimensional evaluation metrics: Four core evaluation metrics are used to verify model performance: root mean square error (RMSE), coefficient of determination (...). The three components are the consistency index (CI) and the mean absolute error (MAE), where RMSE measures the overall error between the predicted and actual values. The CI measures the model's fit, the MAE measures the model's ranking ability, and the MAE measures the average deviation between predicted and actual values. The mathematical formulas for each indicator are as follows: RMSE calculation formula: The smaller the RMSE value, the higher the model's prediction accuracy; Calculation formula: in The mean of the true affinity values ​​for all samples. The closer the model is to 1, the better the model fit. CI calculation formula: in The number of samples in the test set. As an indicator function, the closer CI is to 1, the stronger the model's ranking ability. MAE calculation formula: The smaller the MAE value, the smaller the model prediction bias; Model ensemble strategy: Train 5 sub-models with different initialization parameters (the encoder branch structure of the sub-models is the same, but the initial weights and number of attention heads are slightly different). Use a weighted average fusion method to fuse the prediction results of the 5 sub-models. The weights are determined based on the RMSE of each sub-model's validation set; the smaller the RMSE, the larger the weight. The fusion mathematical formula is: in The weights of the k-th sub-model ( ), The predicted value of the k-th sub-model is used to further improve the prediction accuracy and generalization of the model through model ensemble. Adversarial training enhances robustness: A small perturbation (perturbation amplitude set to 0.01) is added to the input multimodal features. The perturbation is generated using Gaussian noise. The perturbed features are then input into the model for training. This constrains the model to maintain stable prediction results even when there are small perturbations in the input, reducing the model's sensitivity to noise and improving its robustness.

9. A method for constructing a drug target affinity prediction model based on multimodal deep collaborative learning according to any one of claims 1-8, characterized in that, The model deployment and application described in step (7) deploys the trained and optimized multimodal deep collaborative learning model as a drug virtual screening tool, supporting batch prediction, visualization analysis, and drug relocation. Specific application scenarios are as follows: (1) Batch screening of candidate drugs: Input the drug library to be screened (containing a large amount of multimodal data of small molecule compounds) and the disease-related target set (containing multimodal data of target proteins). The model predicts the affinity value of each drug-target combination in batches, sorts them from high to low affinity values, and screens out candidate drugs with high affinity (the affinity threshold can be adjusted according to actual needs, generally set to ≥7.0), which greatly shortens the candidate drug screening cycle and reduces screening costs; (2) Visualization of drug-target binding mechanism: The model outputs a heatmap of key action sites and feature contributions of drug-target binding. The heatmap can intuitively show the correlation strength between the active group of the drug and the target binding residue, as well as the contribution of each feature to the affinity prediction. This helps researchers understand the microscopic mechanism of drug-target binding and provides clear guidance for drug molecule structure optimization (such as modifying the active group with high contribution to improve drug affinity). (3) Drug repositioning: Input multimodal data of marketed drugs and multimodal data of new disease targets, the model predicts the affinity between marketed drugs and new targets, explores the potential of repurposing old drugs, without having to carry out the complete drug development process again, only needing to conduct targeted verification for new targets, greatly shortening the drug development cycle and reducing development risks; (4) Target-specific screening: For specific disease targets, screen out drugs with high affinity for the target and low affinity for other unrelated targets to improve drug specificity, reduce drug side effects and improve drug safety.

10. A method for constructing a drug target affinity prediction model based on multimodal deep collaborative learning according to any one of claims 1-9, characterized in that, The modular extension of the model described in step (8) achieves high scalability through modular design, adapting to different drug development scenarios. The specific extension methods are as follows: (1) Modal data expansion: Supports the addition of modal data for drugs / targets, such as mass spectrometry features and infrared spectral features of drugs, post-translational modification features and phosphorylation site features of target proteins. The new modal data can be seamlessly integrated into the dual heterogeneous encoder by adapting the corresponding branch encoder (such as using the CNN branch for mass spectrometry features and the GRU branch for post-translational modification features). There is no need to reconstruct the overall structure of the model; only the parameters of the corresponding branch need to be adjusted. (2) Replacing the encoder backbone network: The encoder backbone network can be replaced according to the actual data type and prediction requirements. For example, the GAT of the drug branch can be replaced with the equivariant graph neural network (EGNN) to improve the extraction capability of drug molecular spatial features; the ProtBERT of the target branch can be replaced with more advanced protein language models such as ESM-2 to improve the representation capability of amino acid sequence features. The replacement process only requires replacing the backbone network of the corresponding branch, keeping other modules of the model unchanged, and adapting to different application scenarios. (3) Multi-task learning extension: Supports joint multi-task learning. On the basis of the original affinity prediction task, new tasks such as drug toxicity prediction, drug solubility prediction, and target binding mode prediction are added. By sharing the dual-path heterogeneous encoder and cross-modal collaborative attention module, joint training of multiple tasks is achieved, which improves the overall performance of the model and reduces the training cost of the multi-task model, adapting to the needs of the entire drug development process. (4) Hardware adaptation and expansion: The model can be adapted to different hardware environments, including CPU, GPU, TPU, etc. By optimizing the model structure (such as quantization and pruning), the running speed of the model on different hardware can be improved to meet the needs of large-scale batch prediction (such as rapid prediction of millions of drug-target combinations).

Images