DTI prediction method based on dynamic grouping interaction

Abstract

The invention discloses a DTI prediction method based on dynamic grouping interaction. The method comprises the following steps: acquiring an original SMILES of a medicine and an SDF three-dimensional structure file of the medicine; in the multi-modal dynamic alignment preprocessing unit, unified protein features are obtained based on standardized sequence coding vectors and structural feature vectors; in the induced fit dynamic interaction modeling unit, bilinear interaction characteristics are obtained according to unified drug characteristics, unified protein characteristics and dynamic interaction weights, and DTI interaction characteristics are obtained according to the obtained bilinear interaction characteristics; in the domain adaptive feature optimization unit, the feature extractor outputs domain invariance DTI features; and in the evidence uncertainty decoding unit, determining a DTI prediction result according to the calibration prediction probability. According to the DTI prediction method provided by the invention, unification of multi-modal fusion, dynamic interaction modeling and confidence evaluation can be realized at the same time, and a unified prediction method with high precision, high reliability and strong generalization ability is formed.

Description

TECHNICAL FIELD

[0001] The present application belongs to the cross technical field of computer-aided drug design and bioinformatics, and specifically relates to a DTI (Drug-Target Interaction) prediction method based on dynamic grouping interaction. BACKGROUND

[0002] Computer-aided drug design is a key technology for modern drug research and development, and its core is to predict the interaction between drug molecules and target proteins through computational methods, thereby reducing the cost and cycle of experimental screening. Traditional methods such as molecular docking and quantitative structure-activity relationship usually rely on artificially defined molecular descriptors and static protein structure assumptions, which are difficult to fully simulate the real dynamic binding process in vivo, resulting in limitations in prediction accuracy and generalization ability.

[0003] With the progress of deep learning technology, data-driven drug-target interaction prediction methods have become the mainstream research direction. This kind of method uses neural networks to automatically learn the representation of drug molecules (such as SMILES string or molecular graph) and the features of target proteins (such as amino acid sequence), which significantly improves the prediction performance. In the specific technical implementation of drug-target interaction prediction, various deep learning models have been proposed. Graph neural networks are often used to process two-dimensional molecular graph structures of drugs, and convolutional neural networks and recurrent neural networks are often used to extract protein sequence features. In recent years, pre-trained language models (such as ProtTrans for proteins and Mol2Vec for molecules) can extract more generalizable molecular representations through self-supervised learning on large-scale biochemistry datasets. In order to improve the explainability and generalization ability of the model, attention mechanisms, domain adaptation techniques, and confidence evaluation frameworks have also been introduced into drug-target interaction prediction.

[0004] However, existing deep learning methods still face several key challenges: insufficient utilization of multi-modal information, mostly limited to single modal data such as drug two-dimensional structure or protein sequence; lack of dynamic interaction modeling, unable to effectively simulate the conformational changes in the "induced fit" effect; and insufficient generalization ability and prediction reliability of the model in the cold start scenario, lack of effective uncertainty quantification mechanism. Therefore, how to organically integrate these technologies to build a unified framework that can simultaneously solve multi-modal fusion, dynamic interaction modeling, and confidence evaluation is still a technical problem currently faced. SUMMARY

[0005] In order to solve the above problems existing in the prior art, the present application provides a DTI prediction method based on dynamic grouping interaction. The technical problem to be solved by the present application is solved by the following technical scheme: The application provides a DTI prediction method based on dynamic grouping interaction, and the DTI prediction method comprises the following steps: obtaining a drug original SMILES and a drug SDF three-dimensional structure file; In the multi-modal dynamic alignment preprocessing unit, a unified drug feature is obtained based on a 2D topological feature matrix and a substructure feature matrix obtained according to the drug original SMILES and a 3D geometric feature matrix obtained according to the drug SDF three-dimensional structure file, and a unified protein feature is obtained based on a standardized sequence encoding vector obtained according to the original amino acid sequence and a structure feature vector obtained according to the PDB structure file; In the induced fitting dynamic interaction modeling unit, a dynamic interaction weight is obtained based on a gating vector obtained according to the unified protein feature and a basic attention weight obtained according to the unified drug feature and the unified dimension protein feature, a bilinear interaction feature is obtained according to the unified drug feature, the unified protein feature and the dynamic interaction weight, and a DTI interaction feature is obtained according to the obtained bilinear interaction feature; In the domain adaptation feature optimization unit, a feature extractor is obtained according to the trained drug molecule feature encoder and the trained protein feature encoder, the DTI interaction feature is input into a generative adversarial network including the feature extractor and a domain discriminator, and after the joint loss of the generative adversarial network converges, the domain invariance DTI feature is output by the feature extractor; In the evidence uncertainty decoding unit, an original interaction probability is obtained according to the domain invariance DTI feature, a calibrated prediction probability is obtained based on a positive and negative class evidence vector obtained according to the domain invariance DTI feature, and a DTI prediction result is determined according to the calibrated prediction probability.

[0006] Compared with the prior art, the application has the following beneficial effects: The application provides a DTI prediction method based on dynamic grouping interaction, and the DTI prediction method comprises the following steps: Through the deep alignment and fusion of multi-modal features of the multi-modal dynamic alignment preprocessing unit, the input information dimension of the model is fundamentally enriched, laying a foundation for high-precision prediction. By introducing a dynamic interaction reasoning mechanism, the model can simulate real biophysical processes, thereby accurately capturing subtle changes in binding affinity. On this basis, the application designs an uncertainty quantification method that evolves in coordination with dynamic interaction, which can simultaneously output the prediction result and its confidence, providing reliable priority guidance for experimental verification. Finally, by constructing a domain self-adaptive network with multi-modal perception, the application can finely decouple the distribution offset of different modalities, significantly improving the cross-domain generalization ability of the model when facing new targets or new drugs. The DTI prediction method provided by the application can simultaneously realize the unification of multi-modal fusion, dynamic interaction modeling and confidence evaluation, forming a unified prediction method with high precision, high reliability and strong generalization ability.

[0007] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 is a flowchart of a DTI prediction method based on dynamic grouping interaction provided by an embodiment of the present application; Figure 2 is a framework diagram of a DTI prediction model based on dynamic grouping interaction provided by an embodiment of the present application. DETAILED DESCRIPTION

[0009] In order to further illustrate the technical means and effects taken by the present application to achieve the predetermined purpose, a DTI prediction method based on dynamic grouping interaction according to the present application will be described in detail below in combination with the drawings and specific embodiments.

[0010] The foregoing and other technical contents, features and effects of the present application can be clearly presented in the specific embodiment description below in combination with the accompanying drawings. Through the description of the specific embodiments, the technical means and effects taken by the present application to achieve the predetermined purpose can be more deeply and specifically understood. However, the accompanying drawings are provided for reference and illustration only, and are not used to limit the technical solutions of the present application.

[0011] Based on the in-depth analysis of the prior art, the core defect lies in the failure to systematically solve the multi-level complexity in drug-target interaction prediction. Since the existing methods generally use isolated single-modal features or simple feature splicing, the learning of the model on the key modalities such as drug three-dimensional space information and protein evolution background is incomplete. This information loss at the feature level directly limits the theoretical upper limit of the prediction accuracy. More fundamentally, the existing model simplifies the dynamic biological interaction into static feature matching, which cannot simulate the conformational changes of drug and target adaptation in the "induced fit" process, nor can it quantify the uncertainty and associate it with the dynamic interaction process, resulting in a lack of biophysical interpretability and difficulty in providing reliable confidence evaluation when predicting novel structures. In addition, the single domain adaptation strategy ignores the differentiated distribution shift of different modalities in cross-domain scenarios. This coarse-grained alignment often leads to negative transfer phenomenon, which is the main bottleneck for improving the generalization ability of the model in the cold start scenario.

[0012] Therefore, the present application provides a DTI prediction method based on dynamic grouping interaction, please see Figure 1 and Figure 2 , Figure 1 is a flowchart of a DTI prediction method based on dynamic grouping interaction provided by an embodiment of the present application, Figure 2This is a framework diagram of a DTI prediction model based on dynamic grouping interaction provided by an embodiment of the present invention. The DTI prediction method based on dynamic grouping interaction includes: Step 1: Obtain the original SMILES (Simplified Molecular-Input Line-Entry System) and SDF (Structure-Data File) three-dimensional structure files of the drug.

[0013] Specifically, the original SMILES of a drug is a string, which is a specification for encoding the chemical molecular structure using ASCII characters. It describes three-dimensional chemical information such as atomic connection methods, bond types, and ring structures in string form. The SDF three-dimensional structure file is the structural data file of the drug, which contains the connection information of atoms and bonds and stores the Cartesian coordinates of each atom in three-dimensional space.

[0014] Step 2: In the multimodal dynamic alignment preprocessing unit, unified drug features are obtained based on the 2D topological feature matrix and substructure feature matrix obtained from the original SMILES of the drug, and the 3D geometric feature matrix obtained from the SDF three-dimensional structure file of the drug. Unified protein features are obtained based on the standardized sequence coding vector obtained from the original amino acid sequence and the structural feature vector obtained from the PDB structure file.

[0015] Optionally, the multimodal dynamic alignment preprocessing unit includes a drug multimodal feature processing subunit and a protein multimodal feature processing subunit.

[0016] Step 2.1: In the drug multimodal feature processing subunit, the original SMILES of the drug are molecularly standardized and normalized. The 2D topological feature matrix is ​​determined based on the two-dimensional molecular graph structure obtained from the normalized SMILES string. The high-dimensional spatial features are obtained based on the drug's SDF three-dimensional structure file. The 3D geometric feature matrix is ​​obtained based on the high-dimensional spatial features. The substructure vector mapping dictionary is obtained based on the substructure identifier sequence obtained from the original SMILES of the drug. The substructure feature matrix is ​​obtained based on the substructure vector mapping dictionary. Finally, the unified drug features are obtained based on the 2D topological feature matrix, the dimension-unified 3D geometric feature matrix, and the substructure feature matrix.

[0017] Optionally, the drug multimodal feature processing subunit includes a 2D topological feature extraction module, a 3D geometric feature extraction module, a substructure feature extraction module, and a multimodal feature fusion module.

[0018] Step 2.11: In the 2D topological feature extraction module, the original SMILES of the drug are molecularly normalized using RDKit to obtain a normalized SMILES string.

[0019] Step 2.12: Normalize the standardized SMILES string to obtain a normalized SMILES string.

[0020] Specifically, the Chem.CanonSmiless() function in RDKit is used to normalize the standardized SMILES string, resulting in a normalized SMILES string.

[0021] Step 2.13: Convert the normalized SMILES string into a molecule object in RDKit.

[0022] Specifically, the normalized SMILES string is converted into an RDKit molecule object using the Chem.MolFromSmiles() function in RDKit.

[0023] Step 2.14: Generate a two-dimensional molecular graph structure based on the molecular object in RDKit.

[0024] Specifically, the Chem.Compute2DCoords() function in RDKit is called to process the molecular object in RDKit to obtain a two-dimensional molecular graph structure.

[0025] Step 2.15: Use the two-dimensional molecular graph structure to extract features from each atom in the molecule to obtain a 32-dimensional basic descriptor.

[0026] Specifically, the two-dimensional molecular graph structure is traversed to extract features from each atom in the molecule. The extracted features include atom type, bond order, hybridization state, formal charge, etc., and finally a 32-dimensional basic descriptor is obtained. The 32-dimensional basic descriptor is a spliced ​​vector of 11-dimensional elements, 6-dimensional degree, 1-dimensional charge, 5-dimensional hybridization, 1-dimensional aroma, 1-dimensional ring, 5-dimensional hydrogen, and 2-dimensional chirality.

[0027] Step 2.16: Input the 32-dimensional basic descriptor into a two-layer graph attention network for deep feature learning to obtain a 2D topological feature matrix. The output dimension of the 2D topological feature matrix is ​​25×300.

[0028] Step 2.17: In the 3D geometric feature extraction module, OpenBabel is used to perform file format conversion and coordinate parsing on the SDF three-dimensional structure file of the drug, resulting in a molecular structure file, an atomic type sequence of length N, and an N×3 atomic coordinate matrix, where N is the total number of atoms in the molecule, for example, N=25.

[0029] Specifically, OpenBabel is used to convert the SDF three-dimensional structure file of the drug to obtain the molecular structure file. OpenBabel is also used to perform coordinate analysis on the SDF three-dimensional structure file of the drug to obtain the atom type sequence and the atom coordinate matrix. The atom type sequence records the element symbol of each atom, and each row in the atom coordinate matrix corresponds to the coordinates of one atom.

[0030] Step 2.18: Obtain the bond angle distribution characteristics based on the molecular structure file, and obtain an N×N Euclidean distance matrix based on the Euclidean distance between every two atoms in the atomic coordinate matrix.

[0031] Specifically, the process involves iterating through all three consecutive atoms in the molecular structure file, identifying all three sequentially connected atoms, calculating the angles (bond angles) of the three consecutive atoms using vector dot products, and statistically analyzing all bond angle values ​​to generate a bond angle distribution histogram. This histogram is then converted into vector form to obtain the bond angle distribution characteristics. By calculating the Euclidean distances between any two atoms in the atomic coordinate matrix, an Euclidean distance matrix is ​​obtained, for example, a 25×25 Euclidean distance matrix.

[0032] Step 2.19: Use Gaussian radial basis functions to uniformly map the bond angle distribution features and Euclidean distance matrix to a high-dimensional feature space to obtain high-dimensional space features.

[0033] Step 2.110: Use a three-layer fully connected neural network to perform nonlinear transformation on the high-dimensional spatial features to obtain a 3D geometric feature matrix, for example, a 25×256 3D geometric feature matrix.

[0034] Step 2.111: In the substructure feature extraction module, the original SMILES of the drug are processed sequentially using RDKit and Morgan fingerprint algorithms to obtain the substructure identifier sequence.

[0035] Specifically, the original SMILES of the drug are processed using RDKit to obtain Mol objects, and then the Mol objects are processed using the Morgan fingerprint algorithm to obtain substructure identifier sequences.

[0036] Step 2.112: Use the pre-trained Mol2Vec model to perform substructure embedding on each substructure identifier in the substructure identifier sequence to obtain a substructure vector mapping dictionary.

[0037] Step 2.113: For the substructure vector mapping dictionary, extract the substructure embedding vector of each atom within a preset radius, and combine the extracted substructure embedding vectors of all atoms into the first substructure local feature vector.

[0038] Here, the substructure embedding vector refers to the local chemical environment feature, which is a small chemical structural unit composed of atoms and chemical bonds directly connected to an atom within a predetermined radius. The predetermined radius is, for example, 1. For example, initially 18 substructure embedding vectors are obtained, each with a dimension of 300.

[0039] Step 2.114: When the number of all substructure embedding vectors in the local feature vector of the substructure is less than N, the zero-value padding vector and the local feature vector of the first substructure are concatenated to obtain the substructure feature matrix.

[0040] Specifically, when the number of all substructure embedding vectors in the substructure local feature vectors is less than N, a virtual substructure padding mechanism is introduced. The SMARTS pattern "*" is used to define a zero-value padding vector (i.e., a zero vector). This zero-value padding vector and the substructure local feature vectors are concatenated (pasted) to obtain a substructure feature matrix with N substructure embedding vectors. When the number of all substructure embedding vectors is equal to N, the substructure local feature vectors are directly used as the substructure feature matrix. For example, a 25×300 substructure feature matrix has a sequence length of 25.

[0041] Step 2.115: Use the linear projection matrix to perform dimension unification processing on the 3D geometric feature matrix to obtain the dimension-unified 3D geometric feature matrix.

[0042] Here, the 3D geometric feature matrix D after dimensional unification 3d (proj) is: D 3d (proj)= D 3d •W 3d D 3d W is a 3D geometric feature matrix. 3d W is a linear projection matrix. 3d ∈R^(256×300).

[0043] Step 2.116: Perform weighted feature fusion processing on the 2D topological feature matrix, the dimension-unified 3D geometric feature matrix, and the substructure feature matrix to obtain unified drug features.

[0044] Here, the unified drug characteristic D unif For: D unif = w1·D 2d + w2·D 3d (proj) + w3·D sub w1, w2, and w3 are weight coefficients calculated based on information entropy theory, and D unif ∈R^(25×300).

[0045] Step 2.2: In the protein multimodal feature processing subunit, the original amino acid sequence is sequence-validated to obtain a sequence-uniformed amino acid sequence. Based on the length-normalized amino acid sequence obtained from the sequence-uniformed amino acid sequence, a standardized sequence encoding vector is obtained. Based on the PDB structure file with structure completion obtained from the PDB structure file, a structural feature vector is obtained. Based on the sequence embedding vector and structural feature vector obtained from the standardized sequence encoding vector, a unified protein feature is obtained.

[0046] Optionally, the protein multimodal feature processing subunit includes a sequence preprocessing module, a structural feature extraction module, and a feature unification processing module.

[0047] Step 2.21: In the sequence preprocessing module, the original amino acid sequence is validated to verify whether the sequence is consistent. If it is inconsistent, the original amino acid sequence is processed to obtain a consistent amino acid sequence. Then, the consistent amino acid sequence is standardized to obtain a standardized amino acid sequence.

[0048] Specifically, the original amino acid sequence is the amino acid sequence of the target protein. When the original amino acid sequences are not uniform, BioPython is used to unify the original amino acid sequences to obtain a uniform amino acid sequence. If the original amino acid sequences are uniform, the original amino acid sequences are directly used as the uniform amino acid sequences.

[0049] Step 2.22: Perform length normalization on the standardized amino acid sequence to obtain a length-normalized amino acid sequence. For example, fix the sequence length to 1000 residues.

[0050] Step 2.23: In the integer encoding mapping table, query the integer value of each amino acid in the length-normalized amino acid sequence in sequence to obtain a standardized sequence encoding vector based on all the queried integer values.

[0051] Specifically, firstly, for each of the 23 standard amino acids, an integer value is assigned, with the integer values ​​in the order of 0, 1, 2, ... Then, a mapping table is established between the 23 standard amino acids and their corresponding integer values, i.e., an integer encoding mapping table is established. Then, the integer value of each amino acid in the length-normalized amino acid sequence is queried in the integer encoding mapping table in sequence, and the integer values ​​are combined in order. The vector obtained by combining the integer values ​​is used as the standardized sequence encoding vector.

[0052] Step 2.24: In the structural feature extraction module, perform structural completion processing on the PDB (Protein Data Bank) structural file to obtain the structurally completed PDB structural file.

[0053] Specifically, the PDB structure file is the original amino acid sequence PDB file. For incomplete PDB structure files, the 3D structure needs to be completed. AlphaFold3 can be used to complete the PDB structure file to obtain the completed PDB structure file.

[0054] Step 2.25: Based on the spatial distance threshold, identify potential binding residues in the structure-completed PDB structure file to obtain a binary structure hint vector marked with binding sites based on all identified potential binding residues.

[0055] Specifically, for the PDB structure file used for structural completion, the 3D coordinates of the binding region are determined by ligand localization or cavity prediction. The minimum Euclidean distance from each residue to the binding region (in proteins, small molecules typically bind within a specific depression, crack, or cavity; this region is the binding region) is then calculated. It is determined whether the minimum Euclidean distance for each residue is less than or equal to a spatial distance threshold. If so, the residue is identified as a potential binding residue. All selected potential binding residues form a binary binding residue list, which is then normalized in length, for example, to 1000, to obtain a binary structure cue vector. This vector is then used to specifically label binding sites. The spatial distance threshold is, for example, 4 Å.

[0056] Step 2.26: Process the binary structure hint vector through the embedding layer to obtain the structure feature vector. The dimension of the structure feature vector is, for example, 1000×128.

[0057] Step 2.27: In the feature unification processing module, the standardized sequence encoding vector is mapped to a continuous vector space to obtain the sequence embedding vector.

[0058] Specifically, an embedding layer is used to convert the standardized sequence encoding vector into a continuous vector space such as 128-dimensional vector space to obtain the sequence embedding vector.

[0059] Step 2.28: Add the sequence embedding vector and the structural feature vector element by element to obtain the initial fused features.

[0060] Step 2.29: Extract local features from the initial fused features to obtain the extracted local features.

[0061] Specifically, local features are extracted from the initial fused features by using a one-dimensional convolutional layer (kernel size of 3, stride of 1, padding of 1).

[0062] Step 2.210: The local extracted features and the global sequence features extracted from the original amino acid sequence are fused by mean fusion to obtain a unified protein feature.

[0063] Specifically, firstly, the trained ESM-2 model is used to extract global sequence features from the original amino acid sequence. Then, the locally extracted features and the global sequence features are added element-wise. Finally, the result of the addition is divided by 2 to obtain a unified protein feature, P. unif ∈ R^(1000×1024).

[0064] In the multimodal dynamic alignment preprocessing unit, this invention introduces a linear projection layer to uniformly map drug 3D features and protein features to a representation space of the same dimension, and employs a sequence filling strategy to address the issue of inconsistent sequence lengths for different features. Based on this, a weighted fusion mechanism is constructed using information gain theory to generate comprehensive global drug features. This scheme significantly improves the accuracy of binding affinity prediction, demonstrates marked improvement in drug screening performance under cold-start scenarios, and exhibits high consistency between the prediction results and in vitro experimental data.

[0065] Step 3: In the induced fit dynamic interaction modeling unit, dynamic interaction weights are obtained based on the gating vector obtained from the unified protein features, the basic attention weights obtained from the unified drug features and the unified dimension of the protein features. Bilinear interaction features are obtained based on the unified drug features, the unified protein features and the dynamic interaction weights. DTI interaction features are obtained based on the obtained bilinear interaction features.

[0066] Optionally, the induced fit dynamic interaction modeling unit includes a dimension adaptation module, a dynamic weight generation module, and a bilinear pooling module.

[0067] Step 3.1: In the dimension adaptation module, a fully connected projection layer is used to project the uniform protein features onto a uniform dimension to obtain the uniform protein features.

[0068] Specifically, protein features P with unified dimensions proj1 = P unif ·W proj W proj For a fully connected projection layer, W proj ∈R^(1024×512).

[0069] Step 3.2: Use the unsqueeze and repeat operations to perform feature dimension expansion on the uniform drug features and uniform protein features to obtain expanded drug features and expanded protein features.

[0070] Specifically, the unsqueeze operation inserts a new dimension of length 1 at a specified position, and the repeat operation copies data along a specified dimension. The unsqueeze and repeat operations are used to perform feature dimension expansion on uniform drug features to obtain expanded drug features. The unsqueeze and repeat operations are also used to perform feature dimension expansion on uniform protein features to obtain expanded protein features.

[0071] Here, D expand = D unif .unsqueeze(1).repeat(1, 1000, 1),D expand ∈R^(25×1000×512), D expand For extended drug characteristics.

[0072] P expand = P proj1 .unsqueeze(0).repeat(25, 1, 1),P expand ∈R^(25×1000×512), P expand For extended protein characteristics.

[0073] Step 3.3: In the dynamic weight generation module, the gating vector is obtained based on the extended drug features.

[0074] Here, Z = σ(D) expand ·W g + b g ), where b g Here, W represents the bias weights, σ ​​is the activation function, and W represents the weights. g Let W be the weight matrix. g ∈R^(512×512), using Xavier's algorithm on W g Perform initialization.

[0075] Step 3.4: Obtain the basic attention weights based on the unified drug characteristics and the unified protein characteristics.

[0076] Here, attn base = Softmax((D unif ·P unif T ) / ), attn base The basic attention weights are T, where T is the transpose operation.

[0077] Step 3.5: Obtain the dynamic interaction weights based on the gating vector and the basic attention weights.

[0078] Here, W = Z⊙attnbase W represents the dynamic interaction weight, and ⊙ represents the Hadamard product.

[0079] Step 3.6: Sparsify the dynamic interaction weights to obtain a sparse dynamic interaction weight matrix.

[0080] Specifically, the dynamic interaction weights of the gated vector Z<0.3 are reset to zero to highlight key interaction pairs, thereby obtaining a sparse dynamic interaction weight matrix.

[0081] Step 3.7: In the bilinear pooling module, the first projection feature is obtained based on the first bilinear projection matrix and the unified drug feature, and the second projection feature is obtained based on the second bilinear projection matrix and the unified protein feature.

[0082] Here, the first projection feature and the second projection feature are respectively represented as: D proj = D unif ·U,P proj2 = P proj1 ·V,D proj For the first projection feature, P proj2 Let U be the second projection feature, U be the first bilinear projection matrix, and V be the second bilinear projection matrix, where U and V ∈ R^(512×1024).

[0083] Step 3.8: Obtain the bilinear interaction features based on the first projection features, the second projection features, and the dynamic interaction weights.

[0084] Here, the bilinear interaction feature is represented as: F raw = D proj T ·W·P proj2 F raw ∈ R^(1024×1024).

[0085] Step 3.9: Perform feature compression on the bilinear interaction features using a non-overlapping summation pooling operation with a preset step size to obtain the compressed interaction features.

[0086] Specifically, based on a preset step size, the bilinear interaction feature is divided into several non-overlapping regions. All values ​​within each region are summed to obtain the compressed interaction feature. For example, with a preset step size of 2, the 1024×1024 bilinear interaction feature is divided into 512×512 2×2 non-overlapping regions. Then, the four values ​​within each non-overlapping region are summed to obtain the 512×512 compressed interaction feature. Step 3.10: Perform global average pooling on the compressed interaction feature to obtain the DTI interaction feature, for example, the DTI interaction feature F. dti ∈R^512.

[0087] In the induced fit dynamic interaction modeling unit, this invention adapts the feature representation to the hidden layer dimension of the deep learning model by defining the dimension of the composite feature matrix and designing a dedicated projection layer; then, it utilizes global pooling to eliminate sequence length interference and extract dynamic interaction features with strong representational capabilities. This approach effectively improves the prediction performance in protein cold-start scenarios, enhances the model's adaptability to different target structures, and makes cross-target prediction performance more stable.

[0088] Step 4: In the domain adaptation feature optimization unit, a feature extractor is obtained based on the trained drug molecule feature encoder and the trained protein feature encoder. The DTI interactive features are input into the generative adversarial network including the feature extractor and the domain discriminator. After the joint loss of the generative adversarial network converges, the feature extractor outputs the domain-invariant DTI features.

[0089] Optionally, the domain adaptation feature optimization unit includes a pre-training fine-tuning module and a conditional domain adversarial training module.

[0090] Step 4.1: In the pre-training fine-tuning module, obtain the SMILES string of the disease-related compound dataset, divide the SMILES string into substructures, and use the masking language model paradigm to randomly mask a preset proportion of substructure identifiers. The task is to predict the masked substructure identifiers to train a drug molecule feature encoder. During the training process, the parameters of the drug molecule feature encoder are optimized by minimizing the reconstruction cross-entropy loss to obtain the trained drug molecule feature encoder.

[0091] Specifically, the process begins by acquiring SMILES strings from a disease-related compound dataset and standardizing them. Then, the standardized SMILES strings are divided into substructures. Next, a masked language model paradigm is used to randomly mask a predetermined proportion (e.g., 20%) of the substructure identifiers. Predicting the masked substructures is the training task. Model parameters are optimized by minimizing the substructure reconstruction cross-entropy loss, allowing the model to accurately reconstruct the masked substructure features from the unmasked molecular local structure. This forces the model to learn the intrinsic relationships and general chemical rules between molecular substructures. Finally, a pre-trained encoder—the drug molecule feature encoder—is output, capable of deeply understanding chemical language and generating high-quality drug feature representations. This encoder transforms the input drug SMILES strings into corresponding deep feature representations. The disease-related compound dataset covers drug molecules of multiple disease types and structural types to ensure general representation versatility. Step 4.2: Obtain homologous sequence data of the target protein family, perform binary classification annotation on each amino acid residue in the homologous sequence data to generate binding residue annotation data, use the binding residue annotation data as training data to train a protein feature encoder to perform the binding residue recognition task, optimize the parameters of the protein feature encoder by minimizing the classification cross-entropy loss, and obtain the trained protein feature encoder.

[0092] Specifically, firstly, homologous sequence data of the target protein family (a group of proteins that are evolutionarily related, have similar three-dimensional structures, and usually function through similar mechanisms) are acquired, and the binding status of each residue in the homologous sequence data is labeled using a residue binary classification task, i.e., labeling "potentially binding residues" and "non-binding residues" in the protein sequence. Then, the identification of binding residues is used as a supervised task, and the protein feature encoder is trained by minimizing the classification cross-entropy loss, so that the model learns to identify key residues in the protein that are related to drug binding, thus obtaining a trained protein feature encoder, which is used to transform the input protein sequence into a deep feature representation containing binding site information.

[0093] Step 4.3: In the conditional domain adversarial training module, the DTI interactive features are input into the generative adversarial network, which includes a feature extractor and a domain discriminator. The feature extractor converts the DTI interactive features into features to be aligned. The features to be aligned are input into the domain discriminator, which outputs the classification result of the features to be aligned to calculate the adversarial loss. Based on the joint loss obtained from the adversarial loss and the DTI classification task loss, the parameters of the feature extractor and the domain discriminator are updated alternately and iteratively. When the joint loss converges, the feature extractor outputs domain-invariant DTI features. The feature extractor consists of a trained drug molecule feature encoder and a trained protein feature encoder.

[0094] Specifically, a Generative Adversarial Network (GAN) comprises a feature extractor and a domain discriminator. The training method for the GAN is alternating iterative optimization. First, the parameters of the domain discriminator are fixed, and the parameters of the feature extractor are updated, i.e., the joint loss is minimized, enabling the feature extractor to generate DTI classification while outputting domain-invariant features. Then, the parameters of the feature extractor are fixed again, and the parameters of the domain discriminator are updated, i.e., the adversarial loss is maximized, allowing the domain discriminator to improve its ability to distinguish the source domain. The iteration terminates when the domain classification accuracy of the domain discriminator approaches 50%, indicating that the feature distributions of the source and target domains are essentially aligned, at which point the iteration stops. This ultimately yields the domain-invariant DTI features F output by the feature extractor. adapted F adapted ∈R^512.

[0095] In this embodiment, the joint loss of the generative adversarial network is: L total = Ltask +λ·L adv L task = L BCE (DTI) L adv = E[log D()] + E[log(1-D(G(x)))] Among them, L total For joint losses, L task For the DTI classification task loss, L BCE (DTI) is the binary cross-entropy loss used for the DTI prediction task, G(x) is the feature to be aligned, D(·) is the output of the domain discriminator (domain classification probability), and λ=0.01.

[0096] Here, the domain discriminator consists of two fully connected layers with a dimensionality mapping of 512→256→2. The output is the source / target domain activation function. The domain discriminator uses LeakyReLU activation (negative slope = 0.2), which retains more negative feature information compared to ReLU, thus improving the domain discriminator's ability to capture weak domain differences.

[0097] In the domain adaptation feature optimization unit, this invention achieves refined distribution alignment for different modal features by quantitatively analyzing multimodal domain offset differences and dynamically allocating adversarial training weights. This scheme significantly improves cross-domain prediction performance, maintains stable prediction performance in transfer learning scenarios with different target families, and effectively meets the adaptation requirements for data distribution differences in actual drug development.

[0098] Step 5: In the evidence uncertainty decoding unit, the original interaction probability is obtained based on the domain invariant DTI feature, and the calibrated prediction probability is obtained based on the positive and negative evidence vectors obtained based on the domain invariant DTI feature, so as to determine the DTI prediction result based on the calibrated prediction probability.

[0099] Optionally, the evidence uncertainty decoding unit includes a DTI classification branch module, an evidence uncertainty quantification module, and a decision and priority allocation module.

[0100] Step 5.1: In the DTI classification branch module, input the domain-invariant DTI features into the main classifier to obtain the original interaction probabilities, p. raw ∈[0,1].

[0101] In this embodiment, during model debugging and analysis, the original interaction probabilities and the calibrated predicted probabilities can be compared. If the performance of the calibrated predicted probabilities consistently outperforms the original interaction probabilities, it indicates that evidence learning is effective.

[0102] Optionally, the main classifier consists of two fully connected layers with dimensions changing sequentially from 512 to 256 to 1, and the activation function is Sigmoid activation.

[0103] Step 5.2: In the evidence uncertainty quantification module, the domain-invariant DTI feature is input into the evidence layer to obtain positive and negative class evidence vectors, where the positive and negative class evidence vectors are e = [e pos e neg ], e pos To support the amount of evidence that there is an interaction (category 1), e neg The amount of evidence supporting the absence of interaction (category 0).

[0104] Optionally, the evidence layer is a fully connected layer with a dimension change of 512→2 and the activation function is Softplus activation.

[0105] Step 5.3: Based on the positive and negative evidence vectors, obtain the concentration parameter vector composed of the positive class concentration parameter vector and the negative class concentration parameter vector. Specifically, a Dirichlet distribution is parameterized using positive and negative evidence vectors, i.e., α = e + 1 = [α pos , α neg ], α pos For positive concentration parameters, α neg This is the negative class concentration parameter vector.

[0106] Step 5.4: Based on the positive and negative concentration parameter vectors, obtain the cognitive uncertainty and the random uncertainty, which are expressed as follows: u epi = 2 / (α pos + α neg ) u ale = / ( (α pos +α neg ) 2 / (α pos +α neg +1) Among them, u epi To understand uncertainty, u ale It is due to chance and uncertainty.

[0107] Step 5.5: Based on cognitive uncertainty and accidental uncertainty, obtain the total uncertainty, u. total =u epi +u ale .

[0108] Step 5.6: Obtain the calibrated prediction probability p based on the positive and negative class concentration parameter vectors. calibrated = α pos / (α pos +α neg ).

[0109] Step 5.7: In the decision and priority allocation module, determine the relationship between the calibrated prediction probability and the decision threshold. If the calibrated prediction probability is greater than the decision threshold, the DTI prediction result is that there is an interaction between the drug and the target; otherwise, the DTI prediction result is that there is no interaction between the drug and the target.

[0110] Specifically, the DTI prediction result y pred = I(p calibrated >τ1), where I(·) is the indicator function, and τ is the decision threshold, i.e., the calibrated prediction probability p calibrated When y is greater than τ1, pred If the value is 1, then the DTI prediction result indicates the existence of a drug-target interaction, and the calibrated prediction probability p calibrated When y is less than or equal to τ1, pred If the value is 0, the DTI prediction result is that there is no interaction between the drug and the target.

[0111] Preferably, the decision threshold τ1 is 0.5.

[0112] Step 5.8: Verify the DTI prediction results according to the priority determined by the calibrated prediction probability and the total uncertainty probability, and obtain the verification results. The priority order from high to low is as follows: the first priority that meets the first condition, the second priority that meets the second condition, and the third priority that meets the third condition. The first condition is that the calibrated prediction probability is greater than the first threshold and the total uncertainty feature is less than the second threshold. The third condition is that the calibrated prediction probability is less than the second threshold and the total uncertainty feature is less than the second threshold. The second condition is any other condition besides the first and third conditions.

[0113] Specifically, the first condition is p calibrated >τ2andu total <τ3, the third condition is p calibrated<τ3, where τ2 is the first threshold (e.g., τ2 = 0.7) and τ3 is the second threshold (e.g., τ3 = 0.3). The first priority is high priority; a valid DTI prediction result indicating the highest probability of drug-target interaction can be adopted or rapidly validated. The second priority is medium priority; a valid DTI prediction result may require manual review or further experimental validation. The third priority is low priority; a valid DTI prediction result indicating no drug-target interaction is also reliable, but because it is a negative result, its validation priority is set lower.

[0114] In the evidence uncertainty decoding unit, this invention constructs a dimensionally consistent total evidence vector calculation framework by unifying the evidence components of each task into scalar form, effectively solving the dimensional conflict problem in existing technologies. This scheme significantly improves the screening accuracy of high-reliability samples, reduces the experimental cost of in vitro validation, and shortens the validation cycle of drug development.

[0115] Compared with existing technologies, the DTI prediction method provided by this invention brings the following significant benefits by systematically integrating multimodal features, dynamic interaction modeling, domain adaptation training, and evidence uncertainty quantification: 1. This invention effectively improves the completeness and effectiveness of multimodal feature fusion, significantly enhancing prediction accuracy. Existing technologies (such as DrugBAN and ColdstartCPI) generally suffer from the problem of missing key modalities, failing to utilize the three-dimensional spatial structural information of drugs. This invention constructs a multimodal dynamic alignment preprocessing unit, employs linear projection and virtual filling mechanisms, unifies the dimensions and sequence lengths of drug 2D topology, 3D geometry, and substructure features, and performs weighted fusion based on information entropy, forming a more complete unified feature representation. This technical solution fundamentally solves the feature heterogeneity problem, enabling the model to fully utilize key combined information such as three-dimensional structure. Its effect is a significant improvement in the accuracy of DTI prediction, with overall performance indicators on public benchmark datasets outperforming existing mainstream solutions.

[0116] 2. This invention achieves accurate simulation of the dynamic interaction process of "induced fit," significantly enhancing the model's generalization ability in cold-start scenarios. Addressing the problem of fixed interaction weights and the inability to simulate dynamic conformational changes during binding in existing technologies (such as DrugBAN), this invention designs an induced fit dynamic interaction modeling unit. This unit generates dynamic weights through a gated attention mechanism and performs a Hadamard product with the basic attention, achieving dynamic screening and reinforcement of key interaction pairs. This accurately simulates the biophysical process of drug-target adaptation, enabling the model to perform effective inference through dynamic adaptation when faced with novel drugs or targets not seen during training (i.e., cold-start scenarios). Its generalization performance, especially in protein cold-start and cross-target prediction tasks, is significantly improved compared to existing technologies, with a marked reduction in variability.

[0117] 3. This invention provides calibrated prediction uncertainty quantification, which provides reliable priority guidance for experimental verification and effectively reduces R&D costs and time.

[0118] Existing technologies (such as ColdstartCPI and Rep-ConvDTI) mostly lack uncertainty assessment, while a few attempts (such as EviDTI) suffer from a disconnect from dynamic interactions. This invention, in its evidence uncertainty decoding unit, inputs domain-invariant features into the evidence layer as a parameterized Dirichlet distribution, and calculates cognitive uncertainty and chance uncertainty accordingly. This method closely links uncertainty quantification with the dynamic interaction process of the model, achieving precise calibration of the reliability of prediction results. This enables efficient differentiation between high-reliability and low-reliability samples, providing a clear priority ranking for subsequent wet experimental validation. This allows valuable experimental resources to be focused on candidate molecules with the highest probability of success, significantly reducing invalid experiments, lowering R&D costs, and shortening the drug discovery cycle.

[0119] 4. This invention significantly improves the robustness and practicality of the model in cross-domain scenarios through a fine-grained domain adaptation mechanism. Addressing the limitation of existing domain adaptation methods (such as the single discriminator in DrugBAN) in handling multimodal differential distribution shifts, this invention forces the feature extractor to learn robust features with domain invariance through conditional domain adversarial training in the domain adaptation feature optimization unit. This scheme can finely align the distributions of the source and target domains in the multimodal feature space, effectively mitigating the performance degradation caused by domain shifts. This enables the prediction system constructed by this invention to maintain stable and excellent prediction performance when facing data distribution differences common in real-world drug development (such as migration from kinase targets to ERBB2 targets), greatly enhancing the practical value of the technology in real-world scenarios.

[0120] In summary, through the synergistic innovation and system integration of the above-mentioned technical aspects, this invention ultimately provides a DTI prediction solution with higher accuracy, stronger reliability, and better generalization ability, offering a powerful tool for computer-aided drug screening.

[0121] It should be noted that the terms "first," "second," etc., are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the above exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the present invention.

[0122] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings and the disclosure in carrying out the claimed invention. In the description of the invention, the word "comprising" does not exclude other components or steps, "a" or "an" does not exclude a plurality, and "a plurality" means two or more, unless otherwise explicitly specified. Furthermore, while different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce good results.

[0123] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A DTI prediction method based on dynamic grouping interaction, characterized in that, The DTI prediction method includes: Obtain the original SMILES and SDF 3D structure file of the drug; In the multimodal dynamic alignment preprocessing unit, unified drug features are obtained based on the 2D topological feature matrix and substructure feature matrix obtained from the original SMILES of the drug and the 3D geometric feature matrix obtained from the SDF three-dimensional structure file of the drug. Unified protein features are obtained based on the standardized sequence coding vector obtained from the original amino acid sequence and the structural feature vector obtained from the PDB structure file. In the induced fit dynamic interaction modeling unit, dynamic interaction weights are obtained based on the gating vector obtained from the unified protein features, the basic attention weights obtained from the unified drug features and the unified dimension of the protein features. Bilinear interaction features are obtained based on the unified drug features, the unified protein features and the dynamic interaction weights, and DTI interaction features are obtained based on the obtained bilinear interaction features. In the domain adaptation feature optimization unit, a feature extractor is obtained based on the trained drug molecule feature encoder and the trained protein feature encoder. The DTI interactive features are input into the generative adversarial network including the feature extractor and the domain discriminator. After the joint loss of the generative adversarial network converges, the feature extractor outputs domain-invariant DTI features. In the evidence uncertainty decoding unit, the original interaction probability is obtained based on the domain invariant DTI feature, and the calibrated prediction probability is obtained based on the positive and negative evidence vectors obtained based on the domain invariant DTI feature, so as to determine the DTI prediction result based on the calibrated prediction probability.

2. The DTI prediction method according to claim 1, characterized in that, The multimodal dynamic alignment preprocessing unit includes a drug multimodal feature processing subunit and a protein multimodal feature processing subunit; Specifically, unified drug features are obtained based on the 2D topological feature matrix and substructure feature matrix obtained from the original SMILES of the drug, and the 3D geometric feature matrix obtained from the SDF 3D structure file of the drug. Unified protein features are obtained based on the standardized sequence coding vector obtained from the original amino acid sequence and the structural feature vector obtained from the PDB structure file, including: In the drug multimodal feature processing subunit, the original SMILES of the drug are molecularly standardized and normalized. A 2D topological feature matrix is ​​determined based on the two-dimensional molecular graph structure obtained from the normalized SMILES strings. High-dimensional spatial features are obtained based on the drug's SDF three-dimensional structure file. A 3D geometric feature matrix is ​​obtained based on the high-dimensional spatial features. A substructure vector mapping dictionary is obtained based on the substructure identifier sequence obtained from the original SMILES of the drug. A substructure feature matrix is ​​obtained based on the substructure vector mapping dictionary. Finally, a unified drug feature is obtained based on the 2D topological feature matrix, the dimension-unified 3D geometric feature matrix, and the substructure feature matrix. In the protein multimodal feature processing subunit, the original amino acid sequence is sequence-validated to obtain a sequence-uniformed amino acid sequence. Based on the length-normalized amino acid sequence obtained from the sequence-uniformed amino acid sequence, a standardized sequence encoding vector is obtained. Based on the structure-completed PDB structure file obtained from the PDB structure file, a structural feature vector is obtained. Based on the sequence embedding vector and structural feature vector obtained from the standardized sequence encoding vector, a unified protein feature is obtained.

3. The DTI prediction method according to claim 2, characterized in that, In the drug multimodal feature processing subunit, the original SMILES of the drug are molecularly normalized and standardized. A 2D topological feature matrix is ​​determined based on the two-dimensional molecular graph structure obtained from the normalized SMILES strings. High-dimensional spatial features are obtained from the drug's SDF three-dimensional structure file, and a 3D geometric feature matrix is ​​derived from these high-dimensional spatial features, including: In the 2D topological feature extraction module, RDKit is used to perform molecular normalization on the original SMILES of the drug to obtain a normalized SMILES string. The standardized SMILES string is normalized to obtain a normalized SMILES string; Convert the normalized SMILES string into a molecule object in RDKit; A two-dimensional molecular graph structure is generated based on the molecular object in the RDKit; The features of each atom in the molecule are extracted using the two-dimensional molecular graph structure to obtain a 32-dimensional basic descriptor. The 32-dimensional basic descriptor is input into a two-layer graph attention network for deep feature learning to obtain a 2D topological feature matrix. In the 3D geometric feature extraction module, OpenBabel is used to perform file format conversion and coordinate parsing on the SDF three-dimensional structure file of the drug, resulting in a molecular structure file, an atomic type sequence of length N, and an N×3 atomic coordinate matrix, where N is the total number of atoms in the molecule. Based on the molecular structure file, the bond angle distribution characteristics are obtained, and based on the Euclidean distance between every two atoms in the atomic coordinate matrix, an N×N Euclidean distance matrix is ​​obtained. By using Gaussian radial basis functions, the bond angle distribution features and the Euclidean distance matrix are mapped to a high-dimensional feature space to obtain high-dimensional space features. A three-layer fully connected neural network is used to perform nonlinear transformation on the high-dimensional spatial features to obtain a 3D geometric feature matrix.

4. The DTI prediction method according to claim 2, characterized in that, Based on the substructure identifier sequence obtained from the original SMILES of the drug, a substructure vector mapping dictionary is obtained. Then, a substructure feature matrix is ​​derived from the substructure vector mapping dictionary. Finally, unified drug features are obtained based on the 2D topological feature matrix, the dimension-unified 3D geometric feature matrix, and the substructure feature matrix, including: In the substructure feature extraction module, the original SMILES of the drug are processed sequentially using RDKit and Morgan fingerprint algorithms to obtain the substructure identifier sequence; The pre-trained Mol2Vec model is used to perform substructure embedding on each substructure identifier in the substructure identifier sequence to obtain a substructure vector mapping dictionary. For the substructure vector mapping dictionary, the substructure embedding vector of each atom is extracted within a preset radius, with each atom as the center, and the extracted substructure embedding vectors of all atoms are combined into a substructure local feature vector. When the number of all substructure embedding vectors in the substructure local feature vector is less than N, the zero-value padding vector is concatenated with the substructure local feature vector to obtain the substructure feature matrix. In the multimodal feature fusion module, the 3D geometric feature matrix is ​​processed by linear projection matrix to unify the dimensions, resulting in a dimension-unified 3D geometric feature matrix. Weighted feature fusion processing is performed on the 2D topological feature matrix, the dimension-unified 3D geometric feature matrix, and the substructure feature matrix to obtain unified drug features.

5. The DTI prediction method according to claim 2, characterized in that, Sequence verification of the original amino acid sequence yields a sequence-uniformed amino acid sequence. Based on the length-normalized amino acid sequence obtained from the sequence-uniformed sequence, a standardized sequence coding vector is obtained. Based on the PDB structure file with structure completion obtained from the PDB structure file, a structural feature vector is obtained. Based on the sequence embedding vector and structural feature vector obtained from the standardized sequence coding vector, unified protein features are obtained, including: In the sequence preprocessing module, the original amino acid sequence is validated to verify whether the sequence is consistent. If it is inconsistent, the original amino acid sequence is processed to obtain a consistent amino acid sequence. Then, the consistent amino acid sequence is standardized to obtain a standardized amino acid sequence. The standardized amino acid sequence is subjected to length normalization to obtain a length-normalized amino acid sequence; The integer value of each amino acid in the length-normalized amino acid sequence is sequentially queried in the integer encoding mapping table to obtain a normalized sequence encoding vector based on all the queried integer values; In the structural feature extraction module, the PDB structural file is processed to complete the structure, resulting in a structurally completed PDB structural file. Based on a spatial distance threshold, potential binding residues in the structure-completed PDB structure file are identified to obtain a binary structure hint vector, which is marked with binding sites, based on all identified potential binding residues. The binary structure cue vector is processed by the embedding layer to obtain the structure feature vector; In the feature unification processing module, the standardized sequence encoding vector is mapped to a continuous vector space to obtain the sequence embedding vector; The initial fused features are obtained by adding the sequence embedding vector and the structural feature vector element by element. Local feature extraction is performed on the initial fused features to obtain the locally extracted features; The local extracted features and the global sequence features extracted from the original amino acid sequence are fused by mean fusion to obtain a unified protein feature.

6. The DTI prediction method according to claim 1, characterized in that, Based on the gating vector obtained from unified protein features, and the basic attention weights obtained from unified drug features and unified-dimensional protein features, dynamic interaction weights are obtained. Bilinear interaction features are then derived based on the unified drug features, unified protein features, and dynamic interaction weights. Finally, DTI interaction features are obtained based on the bilinear interaction features, including: In the dimension adaptation module, a fully connected projection layer is used to project uniform protein features onto a uniform dimension to obtain protein features of a uniform dimension. The unsqueeze and repeat operations are used to perform feature dimension expansion on uniform drug features and uniform protein features to obtain expanded drug features and expanded protein features. In the dynamic weight generation module, the gating vector is obtained based on the extended drug features; The basic attention weights are derived based on unified drug characteristics and unified protein characteristics. Dynamic interaction weights are obtained based on the gating vector and the basic attention weights; The dynamic interaction weights are sparsified to obtain a sparse dynamic interaction weight matrix. In the bilinear pooling module, the first projection feature is obtained based on the first bilinear projection matrix and the unified drug features, and the second projection feature is obtained based on the second bilinear projection matrix and the unified protein features. Bilinear interaction features are obtained based on the first projection features, the second projection features, and the dynamic interaction weights. The bilinear interaction features are compressed by using a non-overlapping summation pooling operation with a preset step size to obtain the compressed interaction features. Global average pooling is performed on the compressed interactive features to obtain DTI interactive features.

7. The DTI prediction method according to claim 1, characterized in that, A feature extractor is obtained based on the trained drug molecule feature encoder and the trained protein feature encoder. The DTI interactive features are input into a generative adversarial network (GAN) including the feature extractor and a domain discriminator. After the joint loss of the GAN converges, the feature extractor outputs domain-invariant DTI features, including: In the pre-training fine-tuning module, the SMILES string of the disease-related compound dataset is obtained, the SMILES string is divided into substructures, and a masked language model paradigm is used to randomly mask a preset proportion of the substructure identifiers. The task is to predict the masked substructure identifiers to train a drug molecule feature encoder. During the training process, the parameters of the drug molecule feature encoder are optimized by minimizing the reconstruction cross-entropy loss to obtain a well-trained drug molecule feature encoder. Obtain homologous sequence data of the target protein family, perform binary classification labeling on each amino acid residue in the homologous sequence data, generate binding residue labeling data, use the binding residue labeling data as training data, train a protein feature encoder to perform the binding residue recognition task, optimize the parameters of the protein feature encoder by minimizing the classification cross-entropy loss, and obtain the trained protein feature encoder. In the conditional domain adversarial training module, DTI interactive features are input into a generative adversarial network including a feature extractor and a domain discriminator. The feature extractor converts the DTI interactive features into features to be aligned. The features to be aligned are input into the domain discriminator, which outputs the classification result of the features to be aligned to calculate the adversarial loss. Based on the joint loss obtained from the adversarial loss and the DTI classification task loss, the parameters of the feature extractor and the domain discriminator are updated alternately and iteratively. When the joint loss converges, the feature extractor outputs domain-invariant DTI features. The feature extractor consists of a trained drug molecule feature encoder and a trained protein feature encoder.

8. The DTI prediction method according to claim 1, characterized in that, The original interaction probabilities are obtained based on the domain-invariant DTI features, including: In the DTI classification branch module, the domain-invariant DTI features are input into the main classifier to obtain the original interaction probabilities.

9. The DTI prediction method according to claim 1, characterized in that, The calibrated prediction probabilities are obtained based on the positive and negative class evidence vectors derived from the domain-invariant DTI features. The DTI prediction results are then determined based on these calibrated prediction probabilities, including: In the evidence uncertainty quantification module, the domain invariant DTI feature is input into the evidence layer to obtain positive and negative class evidence vectors; Based on the positive and negative evidence vectors, a concentration parameter vector is obtained, which consists of the positive class concentration parameter vector and the negative class concentration parameter vector. The calibrated prediction probability is obtained based on the positive class concentration parameter vector and the negative class concentration parameter vector; In the decision-making and priority allocation module, the relationship between the calibrated prediction probability and the decision threshold is determined. If the calibrated prediction probability is greater than the decision threshold, the DTI prediction result indicates that there is an interaction between the drug and the target; otherwise, the DTI prediction result indicates that there is no interaction between the drug and the target.

10. The DTI prediction method according to claim 9, characterized in that, The module for quantifying the uncertainty of evidence also includes: Cognitive uncertainty and random uncertainty are derived from the positive and negative class concentration parameter vectors, and are respectively expressed as: u epi = 2 / (α pos + α neg ) u ale = / ( (α pos +α neg ) 2 / (α pos +α neg +1) ) Among them, u epi To understand uncertainty, u ale For random uncertainty, α pos For positive concentration parameters, α neg This is the negative class concentration parameter vector; The total uncertainty is derived from cognitive uncertainty and accidental uncertainty; The decision-making and priority allocation module also includes: The DTI prediction results are validated based on the priority determined by the calibrated prediction probability and the total uncertainty probability. The validation results are obtained, with the priority order from high to low as follows: first priority for satisfying the first condition, second priority for satisfying the second condition, and third priority for satisfying the third condition. The first condition is that the calibrated prediction probability is greater than the first threshold and the total uncertainty characteristic is less than the second threshold. The third condition is that the calibrated prediction probability is less than the second threshold and the total uncertainty characteristic is less than the second threshold. The second condition is any condition other than the first and third conditions.

Images