Pre-training language model-based similar relationship network drug target interaction prediction method and system

Abstract

The application relates to the field of drug-target interaction prediction, and discloses a drug-target interaction prediction method and system based on a pre-trained language model similarity relation network, which comprises the following steps: extracting initial drug embedding features and initial protein embedding features through a pre-trained language model; constructing a drug similarity network and a protein similarity network; applying a graph neural network to the drug similarity network and the protein similarity network; extracting drug structure features by using a directed message passing neural network, and extracting protein structure features by using a convolutional neural network; cross-fusing drug similarity relation features and protein similarity relation features, splicing the cross-fusion results with the drug structure features and the protein structure features respectively; and inputting drug-protein pair features into a classifier to perform drug-target interaction prediction. The generated embedding representation is combined with the similarity relation network, so that the complementarity between semantic representation and relation representation can be effectively mined.

Description

Technical Field

[0001] This invention relates to the field of drug-target interaction prediction, and more specifically, to a method and system for predicting drug-target interactions based on a similarity network using a pre-trained language model. Background Technology

[0002] Drug discovery is a crucial pathway to treating human diseases and improving human health. However, traditional drug development processes are often lengthy, costly, and have low success rates. Therefore, efficiently and accurately identifying potential drug-target interactions (DTIs) in the early stages of drug development has become key to improving efficiency and reducing costs. In recent years, with the rapid development of computational biology and artificial intelligence technologies, DTI prediction methods based on traditional machine learning and deep learning have been widely applied in drug development.

[0003] Traditional machine learning-based drug-target interaction (DTI) prediction methods typically require feature engineering of drugs and proteins using prior knowledge or domain experience. Then, methods such as Support Vector Machines (SVM), Random Forests (RF), and Logistic Regression (LR) are employed to predict the drug-target interaction. While traditional machine learning methods have achieved some success in DTI prediction, they often rely on manually designed molecular fingerprints, physicochemical properties, or amino acid features. These hand-designed features are difficult to effectively capture the complex nonlinear relationships between drugs and proteins, thus limiting the predictive performance of the models.

[0004] With the rapid development of deep learning technology, in drug-target interaction (DTI) prediction tasks, researchers mainly focus on prediction methods based on structural features and those based on relational features. Structural feature-based prediction methods typically utilize models such as convolutional neural networks (CNNs), graph neural networks (GNNs), and Transformers to extract structural information from the sequence or graph representations of drugs and proteins for prediction. Deep Convolutional Drug-Target Interaction Prediction (DeepConv-DTI) extracts drug molecular fingerprint features and protein sequence features through a one-dimensional convolutional neural network and calculates interaction scores using fully connected layers. Deep Learning-Based Drug-Target Affinity Prediction (DeepDTA) uses CNNs to process drug SMILES strings and protein sequences separately, extracting structural features to predict DTI. Attention-Based Drug-Target Affinity Prediction (AttentionDTA) introduces an attention mechanism to DeepDTA, fusing the structural features of drugs and proteins to enhance sequence feature representation capabilities. The drug target interaction prediction method based on molecular Transformer (MT-DTI) combines Transformer and CNN to extract features from drug smile strings and protein sequences, and uses a multi-layer predictor for DTI prediction. The drug target interaction prediction method based on graph neural network (GraphDTA), the drug target interaction prediction method based on graph Transformer (GraphomerDTI), and the drug target interaction prediction method based on molecular interaction Transformer (MolTrans) represent feature learning methods based on drug molecule graphs and protein sequences. GraphDTA takes drug molecule graphs and protein sequences as input, using graph convolutional networks (GCN), graph attention networks (GAT), graph isomorphic networks (GIN), and graph attention-graph convolutional networks (GAT-GCN) to learn drug molecule graph features, while simultaneously using CNN to learn protein sequence features, and then concatenates these two structural features for DTI prediction. GraphomerDTI, building on AttentionDTA, replaces CNN-based drug sequence features with drug molecule graph features learned based on a graph Transformer model. MolTrans extracts structural features from drug molecule subgraphs and protein subsequences using a frequent continuous subsequence mining algorithm and enhances the embedding representation with Transformer to improve prediction accuracy. While these methods improve DTI prediction performance to some extent, they focus only on structural information of a single modality, ignoring the potential relationship between the drug and the target. To overcome these limitations, researchers have begun to focus on prediction models based on relationship features, attempting to enhance the model's reasoning ability by constructing and fusing multiple relationship networks.The Drug Target Interaction Prediction Method Based on Network Integration (DTINet) constructs heterogeneous networks using various relationships such as drug-protein interactions and drug-disease associations, and achieves DTI prediction through network diffusion algorithms, feature compression, and matrix completion strategies. The Drug Target Interaction Prediction Method Based on Feature Representation Learning and Deep Neural Networks (DTI-CNN) introduces a feature selector based on a denoising autoencoder and an interaction predictor based on a convolutional neural network to perform DTI prediction on top of DTINet. The Drug Target Interaction Prediction Method Based on Meta-Path Heterogeneous Neural Networks (MHGNN-DTI) and the Drug Target Interaction Prediction Method Based on Adaptive Meta-Graph Learning (AMGDTI) aggregate relationship information from heterogeneous networks using meta-path GNNs and adaptive meta-graph methods to perform DTI prediction on multi-source heterogeneous biological information networks. The Drug Target Interaction Prediction Method Based on Multi-View Learning and Graph Autoencoders (MULGA) further proposes a graph autoencoder framework based on multi-view fusion and combines it with a reliable negative sample screening strategy to effectively integrate the relationship features between drugs and proteins. Drug Target Interaction Prediction (MIDTI) based on multi-view similarity network fusion and interactive attention mechanisms more accurately captures the complex relationships between different entities. While relationship-based models have significant advantages in revealing the latent semantics in biological networks, effectively extracting semantically rich drug and protein features and thus improving DTI prediction accuracy, their core limitation lies in their over-reliance on high-quality prior network information. This prior information, such as drug-protein, drug-drug interactions, or drug-disease association networks, is often difficult to obtain for newly discovered or rare proteins, greatly restricting the practical application of such methods. Furthermore, because these models typically lack deep modeling capabilities for molecular structural features, their generalization performance is severely limited in cold-start scenarios or when the relationship network is incomplete.

[0005] In recent years, the successful application of pretrained language models (PLMs) in the field of natural language processing has driven its rapid development in bioinformatics, providing a new breakthrough for DTI representation learning. The general embedding representations constructed by PLMs possess certain transfer capabilities and, compared to sequence-based features, have significant advantages in cold-start scenarios, meaning the model still exhibits strong discriminative ability when predicting drugs that have not been seen or unknown targets. However, most current works only use PLMs as initial encoders, ignoring the structural dependencies between drug molecules and proteins, and lacking cross-modal deep fusion mechanisms, thus limiting the model's comprehensive expressive power.

[0006] There are currently no effective solutions to the problems in the relevant technologies. Summary of the Invention

[0007] (a) Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a method and system for predicting drug target interactions based on a similarity network using a pre-trained language model. This method has the advantage of enhancing the ability to express relational features, thereby solving the problems in existing technologies that ignore the structural dependencies between drug molecules and proteins, and lack a cross-modal deep fusion mechanism, which limits the comprehensive expressive ability of the model.

[0009] (II) Technical Solution

[0010] To achieve the aforementioned advantages in enhancing relational feature representation capabilities, the specific technical solution adopted in this invention is as follows:

[0011] According to one aspect of the present invention, a method for predicting drug target interactions based on a pre-trained language model using similarity networks is provided, the method comprising:

[0012] S1. Input the drug SMILES string and protein sequence into the pre-trained language model respectively, and extract the initial drug embedding features and the initial protein embedding features through the pre-trained language model.

[0013] S2. The initial drug embedding features and initial protein embedding features are used as the initial feature matrix, and a drug similarity network and a protein similarity network are constructed based on the low-rank sparse representation technique of multiple views.

[0014] S3. Apply graph neural networks to drug similarity networks and protein similarity networks to extract drug similarity relationship features and protein similarity relationship features; use directed message passing neural networks to extract drug structural features, and use convolutional neural networks to extract protein structural features.

[0015] S4. Based on the cross-fusion attention mechanism, cross-fusion of drug similarity features and protein similarity features is performed. The cross-fusion results are then concatenated with drug structural features and protein structural features to obtain drug-protein pair features.

[0016] S5. Input the drug-protein pair features into the classifier to predict drug-target interactions and obtain the prediction results of drug-target interactions.

[0017] Preferably, the step of inputting the drug SMILES string and protein sequence into a pre-trained language model, and extracting initial drug embedding features and initial protein embedding features through the pre-trained language model includes:

[0018] S11. Input the drug SMILES string into the drug pre-trained language model, and use the drug pre-trained language model to extract the initial drug embedding features of the drug SMILES string.

[0019] S12. Input the protein sequence into the protein pre-trained language model, and use the multi-layer Transformer architecture of the protein pre-trained language model to encode the protein sequence and generate initial protein embedding features.

[0020] Preferably, the step of using the initial drug embedding features and initial protein embedding features as initial feature matrices, and constructing drug similarity networks and protein similarity networks based on multi-view low-rank sparse representation techniques includes:

[0021] S21. Introduce a lightweight convolutional neural network module to adjust the initial drug embedding features and the initial protein embedding features respectively, and obtain the adjusted drug embedding features and protein embedding features.

[0022] S22. Use the adjusted drug embedding features as the initial drug feature matrix of the drug similarity network, and use the adjusted protein embedding features as the initial protein feature matrix of the protein similarity network.

[0023] S23. Based on the initial drug feature matrix and the initial protein feature matrix, and combined with the low-rank sparse representation technique of multiple views, construct drug similarity network and protein similarity network respectively.

[0024] Preferably, the construction of the drug similarity network and protein similarity network based on the initial drug feature matrix and the initial protein feature matrix, combined with multi-view low-rank sparse representation technology, includes:

[0025] S221. Select drug molecule descriptors from molecular access system fingerprints, Morgan fingerprints, and topological fingerprints, and encode and embed each drug molecule descriptor to convert the drug molecule descriptor into a corresponding feature matrix.

[0026] S222. Solve the affinity matrix corresponding to the feature matrix based on the low-rank sparse representation technique based on multiple views, and obtain the drug similarity matrix based on the affinity matrix;

[0027] S223. Select protein representation methods based on amino acid composition, combined triplet descriptors, autocorrelation descriptors, and pseudo-amino acid composition. Convert the protein representation methods into corresponding feature matrices and obtain the protein similarity matrix in accordance with step S222.

[0028] S224. Normalize the initial drug feature matrix and drug similarity matrix to obtain the normalized drug feature matrix and drug adjacency matrix, and construct a drug similarity network based on the normalized drug feature matrix and drug adjacency matrix.

[0029] S225. Normalize the initial protein feature matrix and protein similarity matrix to obtain the normalized protein feature matrix and protein adjacency matrix, and construct a protein similarity network based on the normalized protein feature matrix and protein adjacency matrix.

[0030] Preferably, the application of graph neural networks to drug similarity networks and protein similarity networks to extract drug similarity relationship features and protein similarity relationship features; the extraction of drug structural features using directed message passing neural networks, and the extraction of protein structural features using convolutional neural networks, include:

[0031] S31. Based on the preset similarity relationship network feature embedding learning module, the graph convolutional neural network is applied to the drug similarity network and the protein similarity network respectively to obtain drug similarity relationship features and protein similarity relationship features;

[0032] S32. Convert the drug SMILES string into a molecular graph, use a directed message passing neural network to update the message passing using the directed bonds in the molecular graph as the basic unit, obtain the drug molecular graph features, and use the drug molecular graph features as drug structural features.

[0033] S33. Encode the protein amino acid sequence into a digital vector, map the digital vector to the embedding space through an embedding layer, obtain the protein sequence features in the embedding space through a one-dimensional convolutional layer, and use the protein sequence features as the protein structural features.

[0034] Preferably, the step of converting the drug SMILES string into a molecular graph and using a directed message-passing neural network to perform message-passing updates with directed bonds in the molecular graph as the basic units to obtain drug molecular graph features includes:

[0035] S321. Use the Cheminformatics Toolkit to convert the drug SMILES string into a molecular graph, where the set of nodes in the molecular graph is the set of atoms in the drug molecule, and the set of edges in the molecular graph is the set of chemical bonds between atoms.

[0036] S322. Using a directed message-passing neural network with directed bonds in the molecular graph as the basic unit, and combining the hidden state and initial hidden state of nodes and the hidden state and initial hidden state of edges for message-passing updates.

[0037] S323. In message passing updates, obtain node messages through the edge hidden state when aggregating nodes;

[0038] S324. Obtain the hidden state of atoms based on the node message, and aggregate all the hidden states of atoms in the molecular graph to obtain the molecular-level representation of the drug.

[0039] Preferably, the method of cross-fusion attention mechanism for drug similarity features and protein similarity features, and the concatenation of the cross-fusion results with drug structural features and protein structural features respectively, to obtain drug-protein pair features, includes:

[0040] S41. Based on the preset similarity relationship feature fusion module, cross-fusion of drug similarity relationship features and protein similarity relationship features is performed to obtain drug similarity relationship fusion features and protein similarity relationship fusion features.

[0041] S42. Combine drug similarity relationship fusion features with drug structure features to obtain drug combination features; and combine protein similarity relationship fusion features with protein structure features to obtain protein combination features.

[0042] S43. Combine the drug combination features and the protein combination features to obtain the drug-protein pair features.

[0043] Preferably, the preset similarity feature fusion module performs cross-fusion of drug similarity features and protein similarity features to obtain drug similarity fusion features and protein similarity fusion features, including:

[0044] S411. Construct a similarity relationship feature fusion module consisting of several cross-fusion attention layers, and each cross-fusion attention layer consists of a self-attention module, a drug-protein cross-attention module, and a protein-drug cross-attention module.

[0045] S412. Input the drug similarity relationship features and protein similarity relationship features into the cross-fusion attention module. The drug similarity relationship features and protein similarity relationship features are processed by the self-attention module to obtain the self-attention drug similarity relationship features and the self-attention protein similarity relationship features.

[0046] S413. Based on the self-attention drug similarity relationship features, calculate the first query matrix through the drug-protein cross-attention module; based on the self-attention protein similarity relationship features, calculate the first key matrix and the first value matrix through the drug-protein cross-attention module.

[0047] S414. Calculate drug fusion features based on the first query matrix, the first key matrix, and the first value matrix.

[0048] Preferably, after calculating the drug fusion features based on the first query matrix, the first key matrix, and the first value matrix, the method further includes:

[0049] S415. Based on the self-attention protein similarity relationship features, calculate the second query matrix through the protein-drug cross-attention module; based on the self-attention drug similarity relationship features, calculate the second bond matrix and the second value matrix through the protein-drug cross-attention module.

[0050] S416. Calculate protein fusion features based on the second query matrix, the second bond matrix, and the second value matrix;

[0051] S417. Use drug fusion features and protein fusion features as input to the next cross-fusion attention layer and perform a new round of feature fusion. Then, concatenate and map the drug fusion features and protein fusion features output from all cross-fusion attention layers in sequence to obtain drug similarity relationship fusion features and protein similarity relationship fusion features.

[0052] According to another aspect of the present invention, a drug target interaction prediction system based on a pre-trained language model similarity network is also provided, the system comprising:

[0053] The initial embedding feature extraction module is used to input the drug SMILES string and protein sequence into the pre-trained language model, and extract the initial drug embedding features and the initial protein embedding features through the pre-trained language model.

[0054] A similarity network construction model is used to construct drug similarity networks and protein similarity networks based on initial drug embedding features and initial protein embedding features as initial feature matrices and low-rank sparse representation techniques with multiple views.

[0055] The similarity relationship and structural feature extraction module is used to apply graph neural networks to drug similarity networks and protein similarity networks to extract drug similarity relationship features and protein similarity relationship features; it uses directed message passing neural networks to extract drug structural features and convolutional neural networks to extract protein structural features.

[0056] The feature fusion module is used to cross-fuse drug similarity features and protein similarity features based on the cross-fusion attention mechanism. The cross-fusion results are then concatenated with drug structural features and protein structural features to obtain drug-protein pair features.

[0057] The prediction result generation module is used to input drug-protein pair features into the classifier to predict drug-target interactions and obtain the prediction results of drug-target interactions.

[0058] (III) Beneficial Effects

[0059] Compared with existing technologies, this invention provides a method and system for predicting drug target interactions based on a pre-trained language model-based similarity network, which has the following advantages:

[0060] (1) This invention utilizes protein similarity networks and drug similarity networks to extract similarity relationship features and applies them to drug-target interaction prediction. The introduction of such similarity features can significantly improve the prediction accuracy of DTI. More importantly, this invention cleverly utilizes the inherent similarity measure to construct the relationship network, making it highly adaptable and successfully breaking the limitation of traditional relationship-based DTI models that are highly dependent on external prior knowledge.

[0061] (2) This invention combines the embedded representation generated by the biological big language model with the similarity relation network, thereby effectively mining the complementarity between semantic representation and relation representation. Unlike the traditional DTI model that only relies on molecular or sequence features, this invention uses the contextual semantics captured by the big language model to further enhance the model's ability to mine potential dependencies between drugs and proteins using the similarity network. This approach generates more discriminative and generalizable representation features for DTI prediction.

[0062] (3) This invention organically combines structural features with similarity features, which can maintain the semantic information of the molecular and protein structure itself, while incorporating global similarity information between molecules and targets, thereby forming a more discriminative comprehensive representation in the feature space, further improving the accuracy and generalization performance of DTI prediction.

[0063] (4) The drug-target interaction prediction method provided by the present invention has designed a feature fusion module, which integrates the similarity relationship features of drugs and proteins through the relationship cross-fusion attention mechanism and combines them with the structural features of drugs and proteins to enhance the overall feature representation ability, and thus shows stronger generalization ability in cold start prediction tasks. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 This is a flowchart of a drug target interaction prediction method based on a similarity network using a pre-trained language model, according to an embodiment of the present invention.

[0066] Figure 2This is a block diagram illustrating the principle of a drug target interaction prediction system based on a pre-trained language model using a similarity network, according to an embodiment of the present invention.

[0067] Figure 3 This is an overall architecture diagram of a drug target interaction prediction method based on a similarity network using a pre-trained language model, according to an embodiment of the present invention.

[0068] Figure 4 This is a structural diagram of the feature fusion module in the similarity network-based drug target interaction prediction method based on a pre-trained language model according to an embodiment of the present invention.

[0069] Figure 5 This is a structural diagram of self-attention (SA), drug-protein cross-attention (DPCA), and protein-drug cross-attention (PDCA) in the similarity network-based drug target interaction prediction method based on a pre-trained language model according to an embodiment of the present invention.

[0070] Figure 6 This is a schematic diagram illustrating the impact of embedding-dims and cross-fusion attention layers (CA-layers) on evaluation metrics in a drug target interaction prediction method based on a pre-trained language model according to an embodiment of the present invention.

[0071] Figure 7 This is a schematic diagram illustrating the impact of the number of attention heads (Att-heads) and the number of neighbors (Neighbor-num) of nodes in a similarity network based on a pre-trained language model on the evaluation index in an embodiment of the present invention.

[0072] Figure 8 This is a ROC curve on the Human dataset in the similarity network-based drug target interaction prediction method based on a pre-trained language model according to an embodiment of the present invention.

[0073] Figure 9 This is a PR curve on the Human dataset in the similarity relation network-based drug target interaction prediction method based on a pre-trained language model according to an embodiment of the present invention.

[0074] Figure 10 This is an ROC curve on the C. elegans dataset in the similarity network-based drug target interaction prediction method based on a pre-trained language model according to an embodiment of the present invention.

[0075] Figure 11 This is a PR curve on the C. elegans dataset in the similarity relation network-based drug target interaction prediction method based on a pre-trained language model according to an embodiment of the present invention;

[0076] Figure 12 This is a ROC curve on the BioSNAP dataset of the similarity network-based drug target interaction prediction method based on a pre-trained language model according to an embodiment of the present invention.

[0077] Figure 13 This is a PR curve on the BioSNAP dataset for the similarity network-based drug target interaction prediction method based on a pre-trained language model according to an embodiment of the present invention.

[0078] Figure 14 This is a ROC curve on the DrugBank dataset in the similarity network-based drug target interaction prediction method based on a pre-trained language model according to an embodiment of the present invention.

[0079] Figure 15 This is a PR curve on the DrugBank dataset in the similarity relation network-based drug target interaction prediction method based on a pre-trained language model according to an embodiment of the present invention.

[0080] Figure 16 This is a schematic diagram comparing experimental results on an imbalanced dataset in a drug target interaction prediction method based on a similarity network using a pre-trained language model, according to an embodiment of the present invention.

[0081] Figure 17 This is one of the cold-start prediction performance comparison charts in the drug target interaction prediction method based on a similarity network using a pre-trained language model according to an embodiment of the present invention.

[0082] Figure 18 This is the second comparison chart of cold-start prediction performance in the similarity network-based drug target interaction prediction method based on a pre-trained language model according to an embodiment of the present invention.

[0083] Figure 19 This is one of the visualization diagrams of drug-target embedding in the similarity network-based drug-target interaction prediction method based on a pre-trained language model according to an embodiment of the present invention;

[0084] Figure 20 This is the second visualization of drug-target embedding in the similarity network-based drug-target interaction prediction method based on a pre-trained language model according to an embodiment of the present invention.

[0085] In the picture:

[0086] 1. Initial embedding feature extraction module; 2. Similar network model construction module; 3. Similarity relationship and structural feature extraction module; 4. Feature fusion module; 5. Prediction result generation module. Detailed Implementation

[0087] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0088] According to embodiments of the present invention, a method and system for predicting drug target interactions based on a similarity network using a pre-trained language model are provided.

[0089] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to an embodiment of the present invention, a method for predicting drug target interactions based on a pre-trained language model using a similarity network is provided. The method includes:

[0090] S1. Input the drug SMILES string and protein sequence into the pre-trained language model respectively, and extract the initial drug embedding features and the initial protein embedding features through the pre-trained language model.

[0091] S2. The initial drug embedding features and initial protein embedding features are used as the initial feature matrix, and a drug similarity network and a protein similarity network are constructed based on the low-rank sparse representation technique of multiple views.

[0092] S3. Apply graph neural networks to drug similarity networks and protein similarity networks to extract drug similarity relationship features and protein similarity relationship features; use directed message passing neural networks to extract drug structural features, and use convolutional neural networks to extract protein structural features.

[0093] S4. Based on the cross-fusion attention mechanism, cross-fusion of drug similarity features and protein similarity features is performed. The cross-fusion results are then concatenated with drug structural features and protein structural features to obtain drug-protein pair features.

[0094] S5. Input the drug-protein pair features into the classifier to predict drug-target interactions and obtain the prediction results of drug-target interactions.

[0095] In one embodiment, the step of inputting the drug SMILES string and protein sequence into a pre-trained language model, and extracting initial drug embedding features and initial protein embedding features through the pre-trained language model includes:

[0096] S11. Input the drug SMILES string into the drug pre-trained language model, and use the drug pre-trained language model to extract the initial drug embedding features of the drug SMILES string.

[0097] S12. Input the protein sequence into the protein pre-trained language model, and use the multi-layer Transformer architecture of the protein pre-trained language model to encode the protein sequence and generate initial protein embedding features.

[0098] In one embodiment, the step of using initial drug embedding features and initial protein embedding features as initial feature matrices, and constructing drug similarity networks and protein similarity networks based on multi-view low-rank sparse representation techniques includes:

[0099] S21. Introduce a lightweight convolutional neural network module to adjust the initial drug embedding features and the initial protein embedding features respectively, and obtain the adjusted drug embedding features and protein embedding features.

[0100] S22. Use the adjusted drug embedding features as the initial drug feature matrix of the drug similarity network, and use the adjusted protein embedding features as the initial protein feature matrix of the protein similarity network.

[0101] S23. Based on the initial drug feature matrix and the initial protein feature matrix, and combined with the low-rank sparse representation technique of multiple views, construct drug similarity network and protein similarity network respectively.

[0102] In one embodiment, the construction of the drug similarity network and protein similarity network based on the initial drug feature matrix and the initial protein feature matrix, combined with multi-view low-rank sparse representation technology, includes:

[0103] S221. Select drug molecule descriptors from molecular access system fingerprints, Morgan fingerprints, and topological fingerprints, and encode and embed each drug molecule descriptor to convert the drug molecule descriptor into a corresponding feature matrix.

[0104] S222. Solve the affinity matrix corresponding to the feature matrix based on the low-rank sparse representation technique based on multiple views, and obtain the drug similarity matrix based on the affinity matrix;

[0105] S223. Select protein representation methods based on amino acid composition, combined triplet descriptors, autocorrelation descriptors, and pseudo-amino acid composition. Convert the protein representation methods into corresponding feature matrices and obtain the protein similarity matrix in accordance with step S222.

[0106] S224. Normalize the initial drug feature matrix and drug similarity matrix to obtain the normalized drug feature matrix and drug adjacency matrix, and construct a drug similarity network based on the normalized drug feature matrix and drug adjacency matrix.

[0107] S225. Normalize the initial protein feature matrix and protein similarity matrix to obtain the normalized protein feature matrix and protein adjacency matrix, and construct a protein similarity network based on the normalized protein feature matrix and protein adjacency matrix.

[0108] In one embodiment, applying a graph neural network to a drug similarity network and a protein similarity network to extract drug similarity relationship features and protein similarity relationship features; extracting drug structural features using a directed message passing neural network and extracting protein structural features using a convolutional neural network includes:

[0109] S31. Based on the preset similarity relationship network feature embedding learning module, the graph convolutional neural network is applied to the drug similarity network and the protein similarity network respectively to obtain drug similarity relationship features and protein similarity relationship features;

[0110] S32. Convert the drug SMILES string into a molecular graph, use a directed message passing neural network to update the message passing using the directed bonds in the molecular graph as the basic unit, obtain the drug molecular graph features, and use the drug molecular graph features as drug structural features.

[0111] S33. Encode the protein amino acid sequence into a digital vector, map the digital vector to the embedding space through an embedding layer, obtain the protein sequence features in the embedding space through a one-dimensional convolutional layer, and use the protein sequence features as the protein structural features.

[0112] In one embodiment, converting the drug SMILES string into a molecular graph and using a directed message-passing neural network to perform message-passing updates with directed bonds in the molecular graph as the basic unit to obtain drug molecular graph features includes:

[0113] S321. Use the Cheminformatics Toolkit to convert the drug SMILES string into a molecular graph, where the set of nodes in the molecular graph is the set of atoms in the drug molecule, and the set of edges in the molecular graph is the set of chemical bonds between atoms.

[0114] S322. Using a directed message-passing neural network with directed bonds in the molecular graph as the basic unit, and combining the hidden state and initial hidden state of nodes and the hidden state and initial hidden state of edges for message-passing updates.

[0115] S323. In message passing updates, obtain node messages through the edge hidden state when aggregating nodes;

[0116] S324. Obtain the hidden state of atoms based on the node message, and aggregate all the hidden states of atoms in the molecular graph to obtain the molecular-level representation of the drug.

[0117] In one embodiment, the cross-fusion of drug similarity features and protein similarity features based on the cross-fusion attention mechanism, and the concatenation of the cross-fusion results with drug structural features and protein structural features respectively, yields drug-protein pair features including:

[0118] S41. Based on the preset similarity relationship feature fusion module, cross-fusion of drug similarity relationship features and protein similarity relationship features is performed to obtain drug similarity relationship fusion features and protein similarity relationship fusion features.

[0119] S42. Combine drug similarity relationship fusion features with drug structure features to obtain drug combination features; and combine protein similarity relationship fusion features with protein structure features to obtain protein combination features.

[0120] S43. Combine the drug combination features and the protein combination features to obtain the drug-protein pair features.

[0121] In one embodiment, the preset similarity feature fusion module performs cross-fusion of drug similarity features and protein similarity features to obtain drug similarity fusion features and protein similarity fusion features, including:

[0122] S411. Construct a similarity relationship feature fusion module consisting of several cross-fusion attention layers, and each cross-fusion attention layer consists of a self-attention module, a drug-protein cross-attention module, and a protein-drug cross-attention module.

[0123] S412. Input the drug similarity relationship features and protein similarity relationship features into the cross-fusion attention module. The drug similarity relationship features and protein similarity relationship features are processed by the self-attention module to obtain the self-attention drug similarity relationship features and the self-attention protein similarity relationship features.

[0124] S413. Based on the self-attention drug similarity relationship features, calculate the first query matrix through the drug-protein cross-attention module; based on the self-attention protein similarity relationship features, calculate the first key matrix and the first value matrix through the drug-protein cross-attention module.

[0125] S414. Calculate drug fusion features based on the first query matrix, the first key matrix, and the first value matrix.

[0126] S415. Based on the self-attention protein similarity relationship features, calculate the second query matrix through the protein-drug cross-attention module; based on the self-attention drug similarity relationship features, calculate the second bond matrix and the second value matrix through the protein-drug cross-attention module.

[0127] S416. Calculate protein fusion features based on the second query matrix, the second bond matrix, and the second value matrix;

[0128] S417. Use drug fusion features and protein fusion features as input to the next cross-fusion attention layer and perform a new round of feature fusion. Then, concatenate and map the drug fusion features and protein fusion features output from all cross-fusion attention layers in sequence to obtain drug similarity relationship fusion features and protein similarity relationship fusion features.

[0129] According to another embodiment of the invention, such as Figure 2 As shown, a drug target interaction prediction system based on a pre-trained language model similarity network is also provided. This system includes:

[0130] Initial embedding feature extraction module 1 is used to input the drug SMILES string and protein sequence into the pre-trained language model, respectively, and extract the initial drug embedding feature and the initial protein embedding feature through the pre-trained language model.

[0131] Similarity network construction model 2 is used to construct drug similarity networks and protein similarity networks based on the initial drug embedding features and initial protein embedding features as the initial feature matrix and the low-rank sparse representation technique of multiple views.

[0132] The similarity relationship and structural feature extraction module 3 is used to apply graph neural networks to drug similarity networks and protein similarity networks to extract drug similarity relationship features and protein similarity relationship features; it uses directed message passing neural networks to extract drug structural features and convolutional neural networks to extract protein structural features.

[0133] Feature fusion module 4 is used to cross-fuse drug similarity features and protein similarity features based on the cross-fusion attention mechanism, and then splices the cross-fusion results with drug structural features and protein structural features respectively to obtain drug-protein pair features.

[0134] The prediction result generation module 5 is used to input drug-protein pair features into the classifier to predict drug-target interactions and obtain the prediction results of drug-target interactions.

[0135] To facilitate understanding of the above technical solutions of the present invention, the following further explains the above technical solutions of the present invention from the perspective of architecture and principle, as follows:

[0136] The overall architecture diagram of the drug-target interaction prediction method proposed in this invention is as follows: Figure 3As shown, it consists of seven modules: (A) a feature extraction module based on a pre-trained language model, (B) a similarity network construction module, (C) a feature extraction module based on a graph neural network, (D) a drug molecule graph feature extraction module, (E) a protein sequence feature extraction module, (F) a multi-feature fusion module, and (G) a classifier module.

[0137] exist Figure 3 In the table, Human, C. elegans, BioSNAP, and DrugBank represent four datasets, and Drug SMILES and Protein Sequences represent drug SMILES strings and protein sequences, respectively. The symbols represent convolutional neural networks, D-MPNN represents directed message passing neural networks, GCN represents graph convolutional neural networks, SA represents self-attention modules, DPCA represents drug-protein cross-attention modules, PDCA represents protein-drug cross-attention modules, stack represents stacking operations, concatenate represents concatenation operations, and Score represents DTI prediction scores.

[0138] Specifically, the drug-target interaction prediction method provided by this invention first extracts pre-trained features of drugs and proteins using the drug language model BPE and the protein language model ESM-2, respectively; then, it uses a CNN for feature fine-tuning and initializes drug similarity networks and protein similarity networks using the pre-trained features, and applies a graph neural network to the similarity networks to obtain drug relationship features and protein relationship features; it then uses a cross-fusion attention module to cross-fuse these two relationship features to obtain drug relationship fusion features and protein relationship fusion features; simultaneously, it uses a D-MPNN to extract drug structural features and a CNN to extract protein structural features; finally, it concatenates the relationship features and structural features and inputs them into a classifier module for DTI prediction, specifically including:

[0139] I. Relationship Network Feature Encoding;

[0140] ① Drug characteristic coding:

[0141] For drug feature encoding, this invention inputs the SMILES string (Simplified Molecular Input Line Entry System) of drug molecules into a pre-trained drug language model and extracts initial embedding features as initial features for the drug similarity network:

[0142] (1)

[0143] In the formula, Indicates the drug SMILES sequence. Indicates the number of drugs, This represents a drug pre-training model. Represents a set of numbers.

[0144] This invention maps the SMILES strings of drug molecules into low-dimensional embedding vectors with rich semantic information through a drug pre-trained language model. This not only captures the chemical structure information of drug molecules, but also extracts latent semantic features through the deep learning capabilities of the pre-trained language model, thereby enhancing the expressive power of drug features.

[0145] ② Protein feature encoding:

[0146] Since the features extracted by the protein pre-trained language model can effectively capture complex features such as protein structure and function information, this invention uses the protein pre-trained language model to encode these features, generating a low-dimensional vector representation of the protein as the initialization feature for the protein similarity network. Specifically, this invention inputs the protein sequence into the protein pre-trained language model, and encodes the sequence through its multi-layer Transformer architecture to generate a low-dimensional vector representation:

[0147] (2)

[0148] In the formula, Represents the amino acid sequence of a protein. Indicates the number of proteins. This represents the protein pre-trained language model ESM2.

[0149] II. Relationship Network Construction:

[0150] Relationship networks provide crucial information for predicting drug-protein interactions. This invention introduces drug similarity networks and protein similarity networks to capture implicit associations between drugs and proteins. Drugs with similar chemical structures often exhibit similar physicochemical properties and biological activities, while proteins with similar sequences or structures often have similar functions and binding sites. Drug similarity networks can characterize the potential functional connections between drugs with similar chemical structures, while protein similarity networks can reveal the functional consistency between proteins with similar sequences or structures. By aggregating neighbor features on these similarity networks, the model can enhance the representational power of drugs and proteins, thereby improving the accuracy and robustness of interaction predictions. Figure 3As shown in module (B), the initial drug embeddings and initial protein embeddings obtained by using module (A) are adjusted by CNN and used as the initial feature matrices of the drug similarity network and the protein similarity network. The two similarity networks contain rich semantic features of drugs and proteins. At the same time, this invention obtains the adjacency matrix of the drug similarity network and the protein similarity network based on the Multi-View Low-Rank Sparse Representation (MV-LRSR) method to construct the inter-drug similarity network and the inter-protein similarity network.

[0151] ① Construct a similar network feature matrix:

[0152] This invention first introduces a lightweight convolutional neural network module to fine-tune the embedding representation generated by the drug pre-trained language model:

[0153] (3)

[0154] In the formula, This represents the initial drug embedding features of the pre-trained language model. This represents the adjusted drug embedding features. For proteins, the same method is used to obtain fine-tuned features:

[0155] (4)

[0156] In the formula, This represents the initial protein embedding features of the pre-trained language model. Adjusted protein embedding features. To fully explore the semantic features between drugs and proteins, the fine-tuned embedding features are used as the initial feature matrices for constructing drug similarity networks and protein similarity networks.

[0157] ② Construct an adjacency matrix for similar networks:

[0158] The adjacency matrix of a similarity network is used to represent the similarity relationships between drugs or proteins. Therefore, it is necessary to first employ an appropriate similarity calculation method to measure the similarity between drugs or proteins, and then construct the adjacency matrix based on this. Commonly used similarity calculation methods mainly include the Jaccard coefficient method and the cosine similarity method. However, these methods can only characterize shallow similarities within a single feature space, making it difficult to effectively handle redundancy and noise between features, and even more difficult to fully explore the complementary information between multi-view features. In contrast, multi-view-based low-rank sparse representation methods can fuse different feature views within a unified framework, effectively mitigating the impact of single fingerprint representation bias on similarity calculation. Therefore, this invention utilizes a multi-view-based low-rank sparse representation method to obtain the adjacency matrices of drug similarity networks and protein similarity networks. For drugs, this invention uses three molecular descriptors—Molecular Access System (MACCS) fingerprint, Morgan fingerprint, and topological fingerprint—to represent drug molecules. Each descriptor can be viewed as an independent view, characterizing the features of the drug molecule from different perspectives. Then, each drug molecule descriptor is encoded and embedded to transform it into a feature matrix. ,in, This represents the feature dimension of the v-th view. The quantity of drugs is represented. Finally, the feature matrix is ​​solved using a multi-view low-rank sparse representation method. The corresponding affinity matrix :

[0159] (5)

[0160] In the formula, Represents the nuclear norm. Describing the L1 norm, , and These represent constraint coefficients, used to control the affinity matrix. Low-rank constraint strength, sparsity constraint strength and multi-view Figure 1 Devotion, Frobenius norm Matrix transpose This represents taking the absolute value of each element of the matrix. The affinity matrix is ​​obtained through optimization. Then, for the affinity matrix Further processing yields the drug similarity matrix. :

[0161] (6)

[0162] For proteins, this invention uses four representation methods to describe proteins: Amino Acid Composition (AAC), Composition / Transition / Distribution (CTD), Moran Autocorrelation, and Pseudo-Amino Acid Composition (PAAC). These four protein representation methods are then converted into corresponding feature matrices. ,in This represents the feature dimension of the v-th view. This represents the number of proteins. Finally, a protein similarity matrix is ​​obtained using the same method as for drugs. .

[0163] Obtaining the drug similarity matrix Protein similarity matrix Then, the drug feature matrix was analyzed. and similarity matrix Normalization yields the characteristic matrix and adjacency matrix The protein feature matrix was obtained in the same way. and adjacency matrix Then, a drug similarity network is constructed using the feature matrix and the adjacency matrix. Protein similarity network In the construction of similarity relation networks, directly using the adjacency matrix can introduce noise due to the large number of weak connections within the matrix, thereby weakening the feature learning ability of the graph neural network. To address this, this invention employs a threshold filtering strategy, retaining only the top 50 neighbors with the highest similarity to each node. This reduces noise interference while maintaining network connectivity, enhancing the reliability of representation learning.

[0164] III. Feature Embedding Learning:

[0165] ① Similar network feature embedding learning: Figure 3 Module (C) shows the similarity relation network feature embedding learning module, which uses a graph convolutional neural network. Applied to drug similarity networks respectively Protein similarity network To extract drug embedding representations and protein embedding representations:

[0166] (7)

[0167] (8)

[0168] In the formula, and They represent the first Drug and protein characteristics of the layer and They represent the first The drug and protein characteristics of layer +1, and the characteristics of layer 0 are... and , and Let represent the degree matrices of the drug similarity network and the protein similarity network, respectively. and They represent the identity matrix, and These represent the learnable weight matrices, This represents the activation function. and Let represent the adjacency matrices of the drug similarity network and the protein similarity network, respectively. The final output of the graph neural network is:

[0169] (9)

[0170] (10)

[0171] In the formula, The first in the drug similarity network Layer features, This represents the first protein in a protein similarity network. Layer features, Indicates characteristics of drug similarity. This indicates the characteristics of protein similarity relationships.

[0172] ② Drug molecule graph feature embedding learning:

[0173] Drug similarity features characterize drugs at the relational level, while drug molecular structure features directly reflect the intrinsic chemical properties and atomic-level composition of drugs. Therefore, this invention introduces drug molecular structure features on top of similarity features to achieve complementarity between relational and structural information, thereby significantly improving the expressive power of drug features. To obtain drug molecular structure features, this invention first utilizes the RDKit tool (RDKit is a widely used open-source cheminformatics and molecular modeling toolkit specifically designed for processing chemical information, molecular structures, and related data computation) to convert the drug SMILES string into a molecular diagram. ,in, It represents the collection of atoms in a drug molecule. This represents the set of chemical bonds in a molecular diagram. (Atoms) eigenvectors It can contain information such as element type, degree, and valence; edges eigenvectors It can include structural information such as its chemical bond type and whether it is conjugated. Then, this invention uses D-MPNN to learn the drug structural features. Unlike traditional atom-centric message passing, D-MPNN updates information using directed bonds as the basic unit, effectively avoiding the repeated propagation of information in ring structures. In D-MPNN, Represents a node The hidden state, Representing an edge The message at step t, Representing an edge The message at step t+1, Indicates the edge at step t. The hidden state, Indicates the edge at step t+1. Hidden state of nodes The initial hidden state is ,side The initial hidden state is The message passing and update process is as follows:

[0174] (11)

[0175] (12)

[0176] In the formula, This indicates a splicing operation. This represents the activation function. Represents a node The neighborhood, Node messages Through aggregation nodes At step T, the edges are hidden. get:

[0177] (13)

[0178] Then, by node information Its characteristics obtain atomic hidden states :

[0179] (14)

[0180] Finally, the hidden states of all nodes in the graph are aggregated to obtain a molecular-level representation of the drug. :

[0181] (15)

[0182] ③ Protein sequence feature embedding learning:

[0183] The sequence characteristics of a protein not only determine its spatial structure but also directly reflect its inherent biological functions and properties. Therefore, this invention introduces sequence features based on protein similarity characteristics to construct a protein characterization that combines similarity information with its own attributes. Specifically, the protein amino acid sequence is first encoded into a digital vector. The protein sequence features are then mapped to an embedding space using an embedding layer. Subsequently, protein sequence features are obtained through one-dimensional convolution.

[0184] (16)

[0185] In the formula, Indicates a convolutional layer. Indicates the embedding layer. Indicates protein sequence characteristics, Indicates the sequence length. Represents the dimension of sequence features.

[0186] IV. Multi-feature fusion:

[0187] By fusing drug similarity features, drug structural features, and protein similarity features with protein structural features, drugs and proteins can be characterized from multiple perspectives. The relationship features and structural features complement each other. Combining them improves the robustness of the features and mitigates biases introduced by a single data source. Therefore, this invention proposes a multi-feature fusion module, as shown in Figure 4. This module cross-fuses drug relationship features and protein relationship features through a similarity feature fusion module to obtain drug relationship fusion features and protein relationship fusion features respectively. These features are then concatenated with their respective structural features to obtain the final combined features.

[0188] exist Figure 4 middle, Indicates the characteristics of drug molecular diagrams. Indicates protein sequence characteristics, and These represent drug similarity features and protein similarity features, respectively. and These represent drug similarity fusion features and protein similarity fusion features, respectively. and These represent drug combinatorial features and protein combinatorial features, respectively, representing fusion structural features and relational features. The text represents drug-protein pair features, SA stands for self-attention module, DPCA stands for drug-protein cross-attention module, PDCA stands for protein-drug cross-attention module, Cat stands for splicing operation, and Linear stands for linear layer.

[0189] To fully integrate the relationship features between drugs and proteins, the RelationshipFeature Fusion Module (RFFM) utilizes a Cross-Fusion Attention Module (CFAM). This module consists of two Self-Attention Modules (SA), a Drug to Protein Cross-Attention Module (DPCA), and a Protein to Drug Cross-Attention Module (PDCA). The SA structure diagram is shown below. Figure 5 As shown in (A), the structure diagrams of DPCA and PDCA are as follows: Figure 5 As shown in (B).

[0190] exist Figure 5 middle, and These represent drug similarity features and protein similarity features, respectively. and This indicates that the self-attention features are obtained after passing through the SA module. and These represent drug similarity fusion features and protein similarity fusion features, respectively. Linear and Linear_merge represent linear layers, Multi-Attention represents a multi-head attention module, Dropout represents a Dropout layer, Layer Norm represents a normalization layer, and Scaled Dot-Product represents a scaled dot product attention mechanism.

[0191] For each cross-fusion attention module, drug similarity features Similarity characteristics to proteins After passing through the SA module, self-attention features are obtained. and The DPCA module uses self-attention drug similarity features. Calculate the query matrix (Q), and then use self-attention protein similarity features. Calculate the key matrix (K) and value matrix (V):

[0192] (17)

[0193] Finally, drug fusion characteristics were calculated. :

[0194] (18)

[0195] PDCA uses self-attention protein similarity features. Calculate the query matrix Q, and then use self-attention drug similarity features. Calculate the bond matrix K and value matrix V, and obtain the protein fusion characteristics. The obtained drug fusion characteristics Protein fusion characteristics As input to the next cross-fusion attention layer, further feature fusion is performed. The relationship feature fusion module consists of L layers of cross-fusion attention, which concatenate and map the outputs of these layers to obtain drug similarity relationship fusion features. Fusion features related to protein similarity :

[0196] (19)

[0197] (20)

[0198] Then, these features are combined with structural features to obtain drug combination features. and protein combination features :

[0199] (twenty one)

[0200] (twenty two)

[0201] Finally, the drug combination features and protein combination features are concatenated to obtain the drug-protein pair features. :

[0202] (twenty three)

[0203] V. DTI Prediction:

[0204] Drug-protein pair characteristics The input is fed into a multilayer perceptron (MLP) for prediction:

[0205] (twenty four)

[0206] In the formula, It consists of four alternating linear layers and Dropout layers. The model loss function uses the binary cross-entropy loss function:

[0207] (25)

[0208] In the formula, Represents the actual value. Indicates the predicted value. This represents the number of samples.

[0209] To facilitate understanding of the above technical solutions of the present invention, the following further explains the above technical solutions of the present invention from the perspective of architecture and principle, as follows:

[0210] This invention selects four datasets—Human, C. elegans, BioSNAP, and DrugBank—to evaluate the performance of the PLMSRNDTI model (a similarity network-based drug target interaction prediction model based on a pre-trained language model). The Human and C. elegans datasets are relatively small, while the BioSNAP and DrugBank datasets are larger. Details of the datasets are shown in Table 1. To comprehensively evaluate the predictive ability of the PLMSRNDTI model, seven evaluation metrics—ACC, AUC, AUPR, Precision, Recall, F1-score, and MCC—were used.

[0211] Table 1: Specific details of the dataset

[0212] Parameter analysis:

[0213] On the C.elegans dataset, this invention conducted experiments on the embedding dimension, number of attention heads, number of cross-fusion attention layers, and number of neighbors of similar network nodes to determine their optimal settings. For the embedding dimension, four values ​​were selected for experiments: 128, 256, 512, and 1024. The results are as follows: Figure 6 As shown in (a), when the embedding dimension is 512, the values ​​of ACC, AUC, AUPR, F1, and MCC reach their maximum. For the number of attention heads, experiments were conducted with 1, 4, 8, and 16 cases. When the number of attention heads was set to 8, all evaluation metrics except Precision reached their maximum values, as shown in (a). Figure 7 As shown in (b). Regarding the number of cross-fusion attention layers, experiments were conducted using layers 1, 2, 3, and 4, and the trends of various indicators are as follows. Figure 6 As shown in (c), the trend is one of initial increase followed by a decrease, with all indicators reaching their maximum when the number of layers is 3. For the number of neighbors of nodes in a network with similar relationships, experiments were conducted using four scenarios: Top 10, Top 50, Top 100, and no neighbor filtering. When the number of neighbors is 50, all indicators reach their optimal levels, as shown in (c). Figure 7As shown in (d), based on experimental analysis, the optimal hyperparameter configuration determined by the model is as follows: embedding dimension of 512, number of attention heads of 8, number of cross-fusion attention layers of 3, and number of neighboring nodes of similar network relationships of 50.

[0214] exist Figures 6-7 In this context, ACC, AUC, AUPR, Precision, Recall, F1, and MCC are seven evaluation metrics. ACC is accuracy, AUC and AUPR are the areas under the ROC curve and PR curve, respectively, Precision is precision, Recall is recall, F1 is the F1 score, and MCC is the Matthews correlation coefficient.

[0215] Ablation experiment:

[0216] To evaluate the contribution of key components of the model to prediction performance, this invention constructed three ablation experiments on the C. elegans dataset. "PLMSRNDTI-Random" indicates that node features in the relation similarity network are generated by random initialization, replacing the initialization method based on the pre-trained language model; "Without Relation" indicates the removal of features from the drug relation similarity network and the protein relation similarity network; and "Without Structure" indicates the removal of drug molecule graph features and protein sequence features. The experimental results are shown in Table 7. The results show that the loss of relation features or structural features leads to a decrease in model performance. Specifically, the model's MCC decreased by 1.23% and 1.37%, respectively, indicating that multimodal feature fusion plays an important role in improving model performance. Furthermore, when relation similarity network features are replaced with randomly initialized pre-trained language model features, all indicators decrease significantly, with the MCC decreasing by 1.26%, further highlighting the advantages of pre-trained language features in feature representation.

[0217] Table 7: Comparison of ablation results of PLMRNDTI on C. elegans

[0218] Comparative experiment:

[0219] To comprehensively evaluate the predictive performance of the PLMSRNDTI model, this invention selects CoaDTI, MHSADTI, GIFDTI, FMCADTI, and BINDTI as baseline models and evaluates them using a 5-fold cross-validation strategy on the Human, C. elegans, BioSNAP, and DrugBank datasets.

[0220] The model PLMSRNDTI of this invention was compared with the baseline model on the Human dataset. The experimental results are shown in Table 2. The highest index value is bolded, and the second highest value is underlined. On the Human dataset, the model PLMSRNDTI of this invention ranked first in five metrics: ACC, AUC, AUPR, F1, and MCC. Compared with the second-ranked model FMCADTI, it improved by 0.63%, 0.11%, 0.03%, 0.80%, and 1.11%, respectively. Among them, the improvements in MCC and F1 were relatively large, indicating that the model has a good advantage in balancing precision and recall. The precision and recall of the model PLMSRNDTI of this invention ranked second among the baseline models, indicating that the model also has good competitiveness in these two metrics. To visually display the performance differences between the models, this invention provides ROC curves and PR curves, as shown in Table 2. Figures 8-9 As shown.

[0221] Table 2: Comparison of experimental results on the Human dataset

[0222] exist Figures 8-9 In this context, PLMSRNDTI represents the model of this application, CoaDTI, MHSADTI, GIFDTI, FMCADTI, and BINDTI represent five comparative models, True Positive Rate and False Positive Rate represent the true positive rate and false positive rate, and Precision and Recall represent the precision and recall rates.

[0223] Experimental results on the C. elegans dataset further validate the superior performance of this invention on small datasets. Compared to the baseline model, the PLMSRNDTI model of this invention outperforms all metrics, showing improvements of 0.76%, 0.07%, 0.12%, 0.02%, 1.54%, 0.96%, and 1.38% in ACC, AUC, AUPR, Precision, Recall, F1, and MCC, respectively, compared to the best-performing baseline model FMCADTI. The improvements in Recall and MCC are particularly significant, as shown in the corresponding ROC and PR curves. Figures 10-11 As shown.

[0224] Table 3: Comparison of experimental results on the C.elegans dataset

[0225] exist Figures 10-11In this context, PLMSRNDTI represents the model of this application, CoaDTI, MHSADTI, GIFDTI, FMCADTI, and BINDTI represent five comparative models, True Positive Rate and False Positive Rate represent the true positive rate and false positive rate, and Precision and Recall represent the precision and recall rates.

[0226] Experimental results on the Human and C. elegans datasets demonstrate that the PLMSRNDTI model of this invention achieves excellent prediction performance on small-scale datasets. Furthermore, the performance of the model was further validated on two larger datasets: BioSNAP and DrugBank. The experimental results on the BioSNAP dataset are shown in Table 4. The model of this invention outperforms the baseline model in all metrics. The metrics ACC, AUC, AUPR, Recall, F1, and MCC are improved by 4.02%, 2.79%, 2.66%, 2.90%, 3.77%, and 7.91% respectively compared to the second-place model GIFDTI. Simultaneously, the precision of the model of this invention is 0.8717, which is 3.04% higher than the FMCADTI model. These metrics show significant improvements over the baseline model, indicating that the model of this invention has a greater advantage on larger datasets. To more intuitively illustrate the model performance, Figures 12-13 ROC and PR curves on the BioSNAP dataset are presented.

[0227] Table 4: Comparison of experimental results on the BioSNAP dataset

[0228] exist Figure 12-13 In this context, PLMSRNDTI represents the model used in this application, CoaDTI, MHSADTI, GIFDTI, FMCADTI, and BINDTI represent five comparative models, True Positive Rate and False Positive Rate represent the true positive rate and false positive rate, and Precision and Recall represent the precision and recall rates.

[0229] Further experiments were conducted on the DrugBank dataset, and the results are shown in Table 5. The model PLMSRNDTI of this invention also outperforms the baseline model in all metrics. The metrics ACC, AUC, AUPR, Recall, F1, and MCC are improved by 4.51%, 2.92%, 2.94%, 0.73%, 3.53%, and 8.91% respectively compared to the excellent baseline model GIFDTI. Simultaneously, the precision of the model of this invention is improved by 0.64% compared to the baseline model FMCADTI. Furthermore, Figures 14-15 The ROC and PR curves for the DrugBank dataset are presented. Experimental results further demonstrate that the PLMSRNDTI model also possesses excellent predictive capabilities on larger datasets with greater predictive difficulty.

[0230] Table 5: Comparison of experimental results on the DrugBank dataset

[0231] exist Figures 14-15 In this context, PLMSRNDTI represents the model of this application, CoaDTI, MHSADTI, GIFDTI, FMCADTI, and BINDTI represent five comparative models, True Positive Rate and False Positive Rate represent the true positive rate and false positive rate, and Precision and Recall represent the precision and recall rates.

[0232] To evaluate the predictive ability of the PLMSRNDTI model in this invention on imbalanced datasets, this invention constructed an imbalanced dataset with a positive-to-negative sample ratio of 1:5 based on the BioSNAP and DrugBank datasets, and used AUC and AUPR as evaluation metrics. The experimental results are as follows: Figure 16 As shown in the figure, on the BioSNAP dataset, the PLMSRNDTI model outperforms all baseline models in both metrics, with AUC and AUPR improvements of 1.43% and 3.89% respectively compared to the best-performing baseline model, GIFDT. On the DrugBank dataset, the model's AUC is 0.9466, a 2.07% improvement over the baseline model GIFDTI; and its AUPR is 0.8626, a 2.92% improvement over the baseline model BINDTI. These results demonstrate that the PLMSRNDTI model maintains superior predictive performance even under imbalanced data conditions.

[0233] exist Figure 16In this context, PLMSRNDTI represents the model of this application, CoaDTI, MHSADTI, GIFDTI, FMCADTI, and BINDTI represent five comparative models, and the evaluation metrics AUC and AUPR are the areas under the ROC curve and PR curve, respectively.

[0234] Cold start comparative analysis:

[0235] In real-world scenarios, information about drugs and proteins is not always readily available. With the development of new drugs and the discovery of new proteins, models need to be able to handle unknown objects. Pre-trained models, through learning from large-scale data, can acquire semantically rich feature embeddings, mitigating some of the challenges posed by unknown objects. To verify the performance of the PLMSRNDTI model under cold-start conditions, this invention employs a cold-start partitioning strategy on the Human dataset and compares the prediction performance of the proposed model PLMSRNDTI, the PLMSRNDTI-Random model, and the optimal baseline model FMCADTI. PLMSRNDTI-Random represents replacing pre-trained language model features with randomly initialized features from relational similarity networks. The cold-start experiment also uses five-fold cross-validation for evaluation. In terms of dataset partitioning, for drug cold-start, 10% of drugs are randomly selected as new drugs and added to the test set, ensuring they are not visible in the training set; for protein cold-start, 10% of proteins are selected as new proteins and added to the test set, and these proteins are also not visible in the training set. The experimental results are shown in Table 6. Under the drug cold start condition, compared with the model PLMSRNDTI-Random, the model PLMSRNDTI improved ACC, AUC, AUPR, Precision, Recall, F1, and MCC by 24.99%, 38.57%, 19.56%, 2.63%, 41.63%, 27.24%, and 41.41%, respectively, indicating that the features of the pre-trained language model played a key role in the model performance. Compared with the excellent baseline model FMCADTI, the improvements were 3.09%, 3.08%, 2.94%, 2.41%, 6.75%, 4.43%, and 6.74%, respectively. In the case of protein cold start, the PLMSRNDTI model, compared to the PLMSRNDTI-Random model, improved the scores of ACC, AUC, AUPR, Recall, F1, and MCC by 5.61%, 1.00%, 0.69%, 17.71%, 8.97%, and 8.60%, respectively, which further demonstrates the importance of pre-trained language model features. Compared to the FMCADTI model, the scores of ACC, AUC, AUPR, Recall, F1, and MCC improved by 25.45%, 3.55%, 0.06%, 35.82%, 23.96%, and 32.77%, respectively. Figures 17-18The experimental results show the comparison between drug cold start and protein cold start conditions. The PLMSRNDTI model exhibits excellent performance under both drug cold start and protein cold start conditions.

[0236] Table 6: Comparison Results of Cold Start

[0237] Among them, Difference1=value(OURs)-value(OURs-Random), Difference2=value(OURs)-value(FMCADTI).

[0238] exist Figures 17-18 In this paper, ACC, AUC, AUPR, Precision, Recall, F1, and MCC are seven evaluation metrics. ACC is accuracy, AUC and AUPR are the areas under the ROC curve and PR curve, respectively, Precision is precision, Recall is recall, F1 is the F1 score, MCC is the Matthews correlation coefficient, PLMSRNDTI is the model in this application, PLMSRNDTI-Random means that in a relational similarity network, node features are generated by random initialization to replace the self-comparative model based on the initialization method of the pre-trained language model, and FMCADTI is the contrastive model.

[0239] Embedded visualization:

[0240] To demonstrate the model's ability to represent the features of drug-target pairs, this invention uses the t-SNE method on the C. elegans dataset to map the features of drug-target pairs in the test set to a two-dimensional space, thereby displaying the distribution differences between positive and negative sample features. The visualization results are shown below. Figures 19-20 As shown in the figure, the original drug-target pair features obtained using the pre-trained language model are difficult to distinguish in low-dimensional space. With increasing training epochs, the model gradually demonstrates discriminative ability. By the 50th epoch, the features begin to show a clustering trend, and initial separation appears between positive and negative samples. By the 100th epoch, a relatively clear boundary has formed between positive and negative samples in the feature space. When trained to the 150th epoch, the distribution boundary between positive and negative samples becomes more significant, and the discriminative ability is further enhanced. The visualization results show that the model can gradually learn effective semantic representations during training, such as... Figures 19-20 In this context, Positive Pairs and Negative Pairs represent positive and negative sample pairs, respectively.

[0241] This invention proposes a drug-target interaction (DTI) prediction method based on a pre-trained language model using similarity networks. This method effectively improves DTI prediction accuracy by fully integrating similarity and structural features. The main innovation lies in utilizing a pre-trained language model to extract drug and protein embeddings, enhancing the expressive power of similarity network features, and employing a cross-fusion attention mechanism to fuse the two types of similarity features, while simultaneously combining them with structural features, thus achieving an organic combination of structural and similarity features. Experimental results on four datasets fully demonstrate that the proposed model outperforms other baseline models overall and has strong competitiveness. Ablation experiments verify the effectiveness of key parts of the model and further validate its good performance in drug and protein cold-start scenarios. In summary, this invention demonstrates excellent predictive performance in the field of DTI prediction and has good application potential.

[0242] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting drug target interactions based on a similarity network using a pre-trained language model, characterized in that, The method includes: S1. Input the drug SMILES string and protein sequence into the pre-trained language model respectively, and extract the initial drug embedding features and the initial protein embedding features through the pre-trained language model. S2. The initial drug embedding features and initial protein embedding features are used as the initial feature matrix, and a drug similarity network and a protein similarity network are constructed based on the low-rank sparse representation technique of multiple views. S3. Apply graph neural networks to drug similarity networks and protein similarity networks to extract drug similarity relationship features and protein similarity relationship features; use directed message passing neural networks to extract drug structural features, and use convolutional neural networks to extract protein structural features. S4. Based on the cross-fusion attention mechanism, cross-fusion of drug similarity features and protein similarity features is performed. The cross-fusion results are then concatenated with drug structural features and protein structural features to obtain drug-protein pair features. S5. Input the drug-protein pair features into the classifier to predict drug-target interactions and obtain the prediction results of drug-target interactions.

2. The method for predicting drug target interactions based on a pre-trained language model using a similarity network according to claim 1, characterized in that, The step of inputting the drug SMILES string and protein sequence into a pre-trained language model, and extracting initial drug embedding features and initial protein embedding features through the pre-trained language model includes: S11. Input the drug SMILES string into the drug pre-trained language model, and use the drug pre-trained language model to extract the initial drug embedding features of the drug SMILES string. S12. Input the protein sequence into the protein pre-trained language model, and use the multi-layer Transformer architecture of the protein pre-trained language model to encode the protein sequence and generate initial protein embedding features.

3. The method for predicting drug target interactions based on a pre-trained language model using a similarity network according to claim 1, characterized in that, The step of using initial drug embedding features and initial protein embedding features as initial feature matrices, and constructing drug similarity networks and protein similarity networks based on multi-view low-rank sparse representation techniques includes: S21. Introduce a lightweight convolutional neural network module to adjust the initial drug embedding features and the initial protein embedding features respectively, and obtain the adjusted drug embedding features and protein embedding features. S22. Use the adjusted drug embedding features as the initial drug feature matrix of the drug similarity network, and use the adjusted protein embedding features as the initial protein feature matrix of the protein similarity network. S23. Based on the initial drug feature matrix and the initial protein feature matrix, and combined with the low-rank sparse representation technique of multiple views, construct drug similarity network and protein similarity network respectively.

4. The method for predicting drug target interactions based on a pre-trained language model using a similarity network according to claim 3, characterized in that, The construction of drug similarity networks and protein similarity networks based on initial drug feature matrices and initial protein feature matrices, combined with multi-view low-rank sparse representation techniques, includes: S221. Select drug molecule descriptors from molecular access system fingerprints, Morgan fingerprints, and topological fingerprints, and encode and embed each drug molecule descriptor to convert the drug molecule descriptor into a corresponding feature matrix. S222. Solve the affinity matrix corresponding to the feature matrix based on the low-rank sparse representation technique based on multiple views, and obtain the drug similarity matrix based on the affinity matrix; S223. Select protein representation methods based on amino acid composition, combined triplet descriptors, autocorrelation descriptors, and pseudo-amino acid composition. Convert the protein representation methods into corresponding feature matrices and obtain the protein similarity matrix in accordance with step S222. S224. Normalize the initial drug feature matrix and drug similarity matrix to obtain the normalized drug feature matrix and drug adjacency matrix, and construct a drug similarity network based on the normalized drug feature matrix and drug adjacency matrix. S225. Normalize the initial protein feature matrix and protein similarity matrix to obtain the normalized protein feature matrix and protein adjacency matrix, and construct a protein similarity network based on the normalized protein feature matrix and protein adjacency matrix.

5. The method for predicting drug target interactions based on a pre-trained language model using a similarity network according to claim 1, characterized in that, The graph neural network is applied to drug similarity networks and protein similarity networks to extract drug similarity features and protein similarity features. Extracting drug structure features using directed message-passing neural networks, and simultaneously extracting protein structure features using convolutional neural networks, includes: S31. Based on the preset similarity relationship network feature embedding learning module, the graph convolutional neural network is applied to the drug similarity network and the protein similarity network respectively to obtain drug similarity relationship features and protein similarity relationship features; S32. Convert the drug SMILES string into a molecular graph, use a directed message passing neural network to update the message passing using the directed bonds in the molecular graph as the basic unit, obtain the drug molecular graph features, and use the drug molecular graph features as drug structural features. S33. Encode the protein amino acid sequence into a digital vector, map the digital vector to the embedding space through an embedding layer, obtain the protein sequence features in the embedding space through a one-dimensional convolutional layer, and use the protein sequence features as the protein structural features.

6. The method for predicting drug target interactions based on a pre-trained language model using a similarity network according to claim 5, characterized in that, The process of converting drug SMILES strings into molecular graphs and using a directed message-passing neural network to update the molecular graph using directed bonds as basic units to obtain drug molecular graph features includes: S321. Use the Cheminformatics Toolkit to convert the drug SMILES string into a molecular graph, where the set of nodes in the molecular graph is the set of atoms in the drug molecule, and the set of edges in the molecular graph is the set of chemical bonds between atoms. S322. Using a directed message-passing neural network with directed bonds in the molecular graph as the basic unit, and combining the hidden state and initial hidden state of nodes and the hidden state and initial hidden state of edges for message-passing updates. S323. In message passing updates, obtain node messages through the edge hidden state when aggregating nodes; S324. Obtain the hidden state of atoms based on the node message, and aggregate all the hidden states of atoms in the molecular graph to obtain the molecular-level representation of the drug.

7. The method for predicting drug target interactions based on a pre-trained language model using a similarity network according to claim 1, characterized in that, The method involves cross-fusion of drug similarity features and protein similarity features based on a cross-fusion attention mechanism, and then concatenating the cross-fusion results with drug structural features and protein structural features, respectively. The characteristics of the drug-protein pair include: S41. Based on the preset similarity relationship feature fusion module, cross-fusion of drug similarity relationship features and protein similarity relationship features is performed to obtain drug similarity relationship fusion features and protein similarity relationship fusion features. S42. Combine drug similarity relationship fusion features with drug structure features to obtain drug combination features; and combine protein similarity relationship fusion features with protein structure features to obtain protein combination features. S43. Combine the drug combination features and the protein combination features to obtain the drug-protein pair features.

8. The method for predicting drug target interactions based on a pre-trained language model using a similarity network according to claim 7, characterized in that, The preset similarity feature fusion module cross-fuses drug similarity features and protein similarity features to obtain drug similarity fusion features and protein similarity fusion features, including: S411. Construct a similarity relationship feature fusion module consisting of several cross-fusion attention layers, and each cross-fusion attention layer consists of a self-attention module, a drug-protein cross-attention module, and a protein-drug cross-attention module. S412. Input the drug similarity relationship features and protein similarity relationship features into the cross-fusion attention module. The drug similarity relationship features and protein similarity relationship features are processed by the self-attention module to obtain the self-attention drug similarity relationship features and the self-attention protein similarity relationship features. S413. Based on the self-attention drug similarity relationship features, calculate the first query matrix through the drug-protein cross-attention module; based on the self-attention protein similarity relationship features, calculate the first key matrix and the first value matrix through the drug-protein cross-attention module. S414. Calculate drug fusion features based on the first query matrix, the first key matrix, and the first value matrix.

9. The method for predicting drug target interactions based on a pre-trained language model using a similarity network according to claim 8, characterized in that, After calculating the drug fusion features based on the first query matrix, the first key matrix, and the first value matrix, the process further includes: S415. Based on the self-attention protein similarity relationship features, calculate the second query matrix through the protein-drug cross-attention module; based on the self-attention drug similarity relationship features, calculate the second bond matrix and the second value matrix through the protein-drug cross-attention module. S416. Calculate protein fusion features based on the second query matrix, the second bond matrix, and the second value matrix; S417. Use drug fusion features and protein fusion features as input to the next cross-fusion attention layer and perform a new round of feature fusion. Then, concatenate and map the drug fusion features and protein fusion features output from all cross-fusion attention layers in sequence to obtain drug similarity relationship fusion features and protein similarity relationship fusion features.

10. A drug target interaction prediction system based on a pre-trained language model similarity network, used to implement the drug target interaction prediction method based on a pre-trained language model based on any one of claims 1-9, characterized in that, The system includes: The initial embedding feature extraction module is used to input the drug SMILES string and protein sequence into the pre-trained language model, and extract the initial drug embedding features and the initial protein embedding features through the pre-trained language model. A similarity network construction model is used to construct drug similarity networks and protein similarity networks based on initial drug embedding features and initial protein embedding features as initial feature matrices and low-rank sparse representation techniques with multiple views. The similarity relationship and structural feature extraction module is used to apply graph neural networks to drug similarity networks and protein similarity networks to extract drug similarity relationship features and protein similarity relationship features; it uses directed message passing neural networks to extract drug structural features and convolutional neural networks to extract protein structural features. The feature fusion module is used to cross-fuse drug similarity features and protein similarity features based on the cross-fusion attention mechanism. The cross-fusion results are then concatenated with drug structural features and protein structural features to obtain drug-protein pair features. The prediction result generation module is used to input drug-protein pair features into the classifier to predict drug-target interactions and obtain the prediction results of drug-target interactions.

Images