Drug-target interaction prediction method, device and equipment based on comparative learning and Transformer

Abstract

The invention provides a drug-target interaction prediction method, device and equipment based on comparative learning and Transformer, and relates to the technical field of computer artificial intelligence and biological medicine, a to-be-predicted drug-target sequence is acquired and is represented and encoded as to-be-predicted drug sequence representation and to-be-predicted target sequence representation; inputting the to-be-predicted drug sequence representation and the to-be-predicted target sequence representation into a pre-trained prediction model; and sequentially performing feature extraction and feature mapping through a Transform and a multilayer perceptron (MLP) network in the prediction model, calculating an interaction probability between a to-be-predicted drug feature mapping vector and a to-be-predicted target feature mapping vector, and taking a drug-target pair of which the interaction probability is greater than a preset threshold value as a predicted interaction result. According to the method, the accuracy and generalization ability of drug-target interaction prediction are improved, and the method can be widely applied to new target discovery, drug relocation, drug screening and precision medical treatment.

Description

Technical Field

[0001] The present application relates to the fields of computer artificial intelligence and biomedicine technology, and in particular to a drug-target interaction prediction method, device and equipment based on contrastive learning and Transformer. Background Art

[0002] Drug-target interaction (DTI) prediction is a key step in computer-aided drug design and is of great significance to drug development. DTI prediction primarily relies on experimental screening or ligand / receptor-based virtual screening. These traditional methods are limited by the vast size of chemical and protein target spaces and suffer from high costs and long processing times.

[0003] Although artificial intelligence technology has been applied in this field in recent years, existing prediction models still have significant shortcomings. Deep learning architectures represented by Transformer, with their unique self-attention mechanism, overcome the limitations of local feature extraction of traditional convolutional neural networks and the gradient vanishing of recurrent neural networks, providing a new technical path for DTI prediction. This architecture can simultaneously process the SMILES sequence representation of drug molecules and the amino acid sequence characteristics of proteins, effectively capturing the complex interaction relationship between drugs and targets through a global information interaction mechanism. However, in practical applications, existing Transformer models still have technical bottlenecks such as significant data distribution deviations and insufficient prediction performance for unknown drug-target combinations, which restrict their application in real drug screening scenarios. Summary of the Invention

[0004] The present application provides a drug-target interaction prediction method, device and equipment based on contrastive learning and Transformer to solve the above-mentioned problems existing in existing drug target prediction methods.

[0005] In the first aspect, the present application proposes a drug-target interaction prediction method based on contrastive learning and Transformer, the method comprising: Obtaining a drug-target sequence to be predicted and encoding the sequence representation as a drug sequence representation to be predicted and a target sequence representation to be predicted; Inputting the drug sequence representation to be predicted and the target sequence representation to be predicted into a pre-trained prediction model; Performing feature extraction on the sequence representation of the drug to be predicted and the sequence representation of the target to be predicted by using the Transformer in the prediction model to obtain a feature vector of the drug to be predicted and a feature vector of the target to be predicted; Performing feature mapping on the drug feature vector to be predicted and the target feature vector to be predicted by a multi-layer perceptron (MLP) network in the prediction model to obtain a drug feature mapping vector to be predicted and a target feature mapping vector to be predicted; The interaction probability between the drug feature mapping vector to be predicted and the target feature mapping vector to be predicted is calculated, and the drug-target pair with an interaction probability greater than a preset threshold is taken as the predicted interaction result.

[0006] Optionally, the training process of the pre-trained prediction model includes: Obtaining drug-target pair information from a drug-target interaction database, and characterizing and encoding the drug-target pair information to obtain training data, wherein the training data includes binary training data and triple training data; Inputting the binary training data into the Transformer in the prediction model to be trained, performing deep feature extraction on the binary training data to obtain a drug feature vector and a target feature vector; Based on the feature mapping of the multi-layer perceptron MLP network in the prediction model to be trained, mapping the drug feature vector and the target feature vector to a high-dimensional feature space to obtain the drug feature mapping vector and the target feature mapping vector; calculating a drug-target interaction probability between the drug feature mapping vector and the target feature mapping vector, and calculating a drug-target loss based on the drug-target interaction probability; Optimizing the parameters of the Transformer and Multi-Layer Perceptron (MLP) networks in the prediction model to be trained based on the loss between the drug targets to obtain an initially optimized prediction model to be trained; Inputting the triplet training data into the Transformer of the initially optimized prediction model to be trained to obtain an active compound feature vector, an inactive compound feature vector, and a target protein feature vector; Mapping the active compound feature vector, the inactive compound feature vector, and the target protein feature vector to the high-dimensional feature space to obtain an active compound feature mapping vector, an inactive compound feature mapping vector, and a target protein feature mapping vector; calculating a contrastive learning loss between the active compound feature map vector, the inactive compound feature map vector, and the target feature map vector; The parameters of the multi-layer perceptron (MLP) network in the initially optimized prediction model to be trained are optimized based on the contrastive learning loss to obtain the optimized prediction model.

[0007] Optionally, the process of establishing the drug-target interaction database is as follows: Acquiring drug-target pair data, wherein the drug-target pair data includes a drug molecule and a target protein; The drug molecule is represented by a SMILES sequence to obtain the original drug molecule sequence, and the target protein is represented by an amino acid sequence to obtain the original target protein sequence; Preprocessing the original drug molecule sequence and the original target protein sequence to obtain a drug molecule sequence and a target protein sequence; The drug-target interaction database is established using the drug molecule sequence and the target protein sequence.

[0008] Optionally, the drug-target pair information includes a drug molecule sequence and a target protein sequence. The drug-target pair information is obtained from a drug-target interaction database, and the drug-target pair information is represented and encoded to obtain training data, including: Morgan fingerprints are used to characterize and encode the drug molecular sequence to obtain a drug sequence characterization; Using a protein language model (PLM) to represent and encode the target protein sequence to obtain a target sequence representation; Setting the label value of the drug-target pair whose interaction value is less than a first preset threshold as a first label value, wherein the drug molecule in the drug-target pair corresponding to the first label value is an active compound; Setting the label value of the drug-target pair whose interaction value is greater than the first preset threshold as a second label value, wherein the drug molecule in the drug-target pair corresponding to the second label value is an inactive compound; The drug sequence representation, target sequence representation and label value are used as the binary training data, and the active compound sequence, inactive compound sequence and target protein sequence are used as the triplet training data.

[0009] Optionally, the binary training data is input into a Transformer in a prediction model to be trained, and deep feature extraction is performed on the binary training data to obtain a drug feature vector and a target feature vector, including: Inputting the drug sequence representation and the target sequence representation into the Transformer, passing through the linear layer and the linear embedding layer, to obtain the drug embedding vector and the target embedding vector; The drug embedding vector and the target embedding vector are passed through an encoding network to obtain an initial drug feature vector and the target feature vector; The drug embedding vector and the target embedding vector are integrated by using a cross attention mechanism, and the initial drug feature vector is optimized to obtain the drug feature vector.

[0010] Optionally, calculating the drug-target interaction probability between the drug feature map vector and the target feature map vector includes: Calculate drug feature map vector and target feature map vector The high-dimensional space distance is calculated by cosine similarity. To characterize and predict the probability of drug-target interaction , the calculation formula is as follows: .

[0011] Optionally, the drug-target loss can be calculated using the following formula: :

[0012] in, is the binary cross entropy loss, is the drug-target interaction probability, is the label value of the drug-target pair.

[0013] Optionally, the contrastive learning loss is calculated using the following formula:

[0014] in, is the interval loss, is the target protein feature mapping vector, is the active compound feature mapping vector, is the feature map vector of inactive compounds.

[0015] In a second aspect, the present application proposes a drug-target interaction prediction device based on contrastive learning and Transformer, the device comprising: An acquisition module is used to acquire the drug-target sequence to be predicted and represent and encode it into a representation of the drug sequence to be predicted and a representation of the target sequence to be predicted; An input module, configured to input the drug sequence representation to be predicted and the target sequence representation to be predicted into a pre-trained prediction model; A feature extraction module is used to extract features from the sequence representation of the drug to be predicted and the sequence representation of the target to be predicted through the Transformer in the prediction model to obtain a feature vector of the drug to be predicted and a feature vector of the target to be predicted; A feature mapping module is used to perform feature mapping on the drug feature vector to be predicted and the target feature vector to be predicted through a multi-layer perceptron (MLP) network in the prediction model to obtain a drug feature mapping vector to be predicted and a target feature mapping vector to be predicted; The calculation prediction module is used to calculate the interaction probability between the feature mapping vector of the drug to be predicted and the feature mapping vector of the target to be predicted, and take the drug-target pair with an interaction probability greater than a preset threshold as the predicted interaction result.

[0016] In the third aspect, the present application proposes an electronic device comprising a memory, a processor, and a computer program stored on the memory, wherein the processor executes the computer program to implement the drug-target interaction prediction method based on contrastive learning and Transformer as described in any one of the first aspects above.

[0017] The present application includes the following advantages: the drug-target interaction prediction method, device and equipment based on contrastive learning and Transformer proposed in the present application first obtains the drug-target sequence to be predicted and characterizes and encodes it into a drug sequence representation to be predicted and a target sequence representation to be predicted; then, the drug sequence representation to be predicted and the target sequence representation to be predicted are input into a pre-trained prediction model; the Transformer in the prediction model is used to extract features of the drug sequence representation to be predicted and the target sequence representation to be predicted to obtain a drug feature vector to be predicted and a target feature vector to be predicted; the multi-layer perceptron MLP network in the prediction model is used to feature map the drug feature vector to be predicted and the target feature vector to be predicted to obtain a drug feature mapping vector to be predicted and a target feature mapping vector to be predicted; finally, the interaction probability between the drug feature mapping vector to be predicted and the target feature mapping vector to be predicted is calculated, and the drug-target pairs with an interaction probability greater than a preset threshold are taken as predicted interaction results.

[0018] This method uses the Transformer architecture to capture the deep feature relationship between drugs and targets, and combines contrastive learning to optimize feature representation, so that the prediction model can more effectively distinguish different drug-target pair combinations, and introduces the protein language model PLM to obtain high-quality protein feature representation, thereby enhancing adaptability to new targets and low-resource targets. This method improves the accuracy and generalization ability of drug-target interaction prediction, and can be widely used in new target discovery, drug repositioning, drug screening and precision medicine, and is expected to reduce experimental screening costs and accelerate the process of innovative drug development. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments of the present application. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0020] Figure 1 This is a schematic flow chart of the steps of a drug-target interaction prediction method based on contrastive learning and Transformer provided in an embodiment of the present application; Figure 2 Schematic diagram of a flow chart of a drug-target interaction prediction method based on contrastive learning and Transformer provided in an embodiment of the present application; Figure 3 Schematic diagram of the overall workflow of a drug-target interaction prediction method provided in an embodiment of the present application; Figure 4 This is a schematic diagram of the architecture of a prediction model provided in an embodiment of the present application; Figure 5 Schematic diagram of the implementation steps of a drug-target interaction prediction method provided in an embodiment of the present application; Figure 6 This is a schematic diagram of the functional modules of a drug-target interaction prediction device based on contrastive learning and Transformer provided in an embodiment of the present application; Figure 7 This is a schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0022] Considering that contrastive learning is a technology that has been widely used in representation learning in recent years, it aims to optimize the feature space by constructing positive and negative sample pairs, so that similar samples are close and irrelevant samples are far away. This method has been successfully applied to fields such as image recognition, natural language processing, and bioinformatics. In the task of drug-target interaction prediction, contrastive learning can be used to enhance the representation of the interaction between drugs and targets, enabling the model to more effectively capture the feature differences of different drug-target combinations and reduce potential biases in the training data. In addition, the Transformer architecture combined with contrastive learning can further enhance the generalization ability of the model, making it perform better in low-resource scenarios or prediction tasks of new targets. However, most current drug-target interaction prediction methods have not yet fully combined contrastive learning strategies, resulting in the model still facing greater uncertainty when dealing with new molecules or new targets.

[0023] Based on this, the present application proposes a drug-target interaction prediction method based on contrastive learning and Transformer architecture. Specifically, based on the Transformer architecture and contrastive learning strategy, an efficient drug-target interaction prediction model is constructed. Through the drug-target interaction database, the feature information of active compounds (positive), inactive compounds (negative) and target proteins (anchor) are extracted respectively, and the sequence representation of drugs and targets is modeled as feature vectors using the Transformer architecture. Subsequently, the feature representation of drugs and targets is optimized in combination with the contrastive learning strategy, so that drug-target pairs with real interactions are closer in the shared high-dimensional feature space, while non-interacting pairs maintain a greater distance. Finally, based on the optimized feature representation, a drug-target interaction prediction model is constructed to improve the predictive ability of drug-target combinations.

[0024] In the first aspect, the present application embodiment proposes a drug-target interaction prediction method based on contrastive learning and Transformer, see Figure 1 and Figure 2 , Figure 1 This is a schematic diagram of the steps of a drug-target interaction prediction method based on contrastive learning and Transformer proposed in this application. Figure 2 : This is a flow chart of a drug-target interaction prediction method based on contrastive learning and Transformer proposed in an embodiment of the present application, which includes the following steps: Step 101: Obtain a drug-target sequence to be predicted and characterize and encode it into a drug sequence representation to be predicted and a target sequence representation to be predicted.

[0025] The drug in the drug-target sequence to be predicted is the drug molecule sequence to be predicted, and the target in the drug-target sequence to be predicted is the target protein sequence corresponding to the drug molecule to be predicted. In the process of drug-target prediction, different target proteins are usually predicted for the same drug molecule, wherein the target protein can be all human target proteins.

[0026] When making predictions, the drug molecules to be predicted can be combined with the human target protein sequences in sequence and predicted in batches.

[0027] Step 102: Input the drug sequence representation to be predicted and the target sequence representation to be predicted into a pre-trained prediction model.

[0028] The above-mentioned pre-trained prediction model includes a transformer and a multi-layer perceptron (MLP) network. The transformer is used to extract features from the representation of the drug sequence to be predicted and the representation of the target sequence to be predicted, and the multi-layer perceptron (MLP) network is used to perform reinforcement learning on the extracted features to obtain the corresponding predicted interaction results.

[0029] Step 103: extracting features from the sequence representation of the drug to be predicted and the sequence representation of the target to be predicted by using the Transformer in the prediction model to obtain a feature vector of the drug to be predicted and a feature vector of the target to be predicted.

[0030] The drug sequence representation to be predicted and the target sequence representation to be predicted are input into the feature extractor based on the Transformer architecture. The multi-head attention mechanism is used to extract the deep feature vectors of the drug sequence representation to be predicted and the target sequence representation to be predicted respectively. The cross-attention mechanism is used to realize the feature interaction from target features to drug features, and the feature vector of the drug to be predicted and the feature vector of the target to be predicted are obtained.

[0031] Specifically, first, the sequence representation of the drug to be predicted and the sequence representation of the target to be predicted pass through the linear layer and the linear embedding layer to obtain the embedding vector of the drug to be predicted and the embedding vector of the target to be predicted; then, the embedding vector of the drug to be predicted and the embedding vector of the target to be predicted pass through the encoding network to obtain the initial feature vector of the drug to be predicted and the feature vector of the target to be predicted; finally, the cross-attention mechanism is used to integrate the embedding vector of the drug to be predicted and the embedding vector of the target to be predicted, and the initial feature vector of the drug to be predicted is optimized to obtain the feature vector of the drug to be predicted.

[0032] After optimizing the drug feature vector through the cross-attention mechanism, the target protein information is integrated into the drug feature vector to further improve the model performance.

[0033] Step 104: performing feature mapping on the drug feature vector to be predicted and the target feature vector to be predicted by using the multi-layer perceptron (MLP) network in the prediction model to obtain a drug feature mapping vector to be predicted and a target feature mapping vector to be predicted.

[0034] After the multi-layer perceptron (MLP) network is optimized using a contrastive learning strategy, the mapping distances of active drug-target pairs are closer, while the mapping distances of inactive drug-target pairs are farther, enabling more accurate prediction of the target proteins corresponding to drug molecules.

[0035] Step 105: Calculate the interaction probability between the feature mapping vector of the drug to be predicted and the feature mapping vector of the target to be predicted, and take the drug-target pair with an interaction probability greater than a preset threshold as the predicted interaction result.

[0036] Cosine similarity is used to calculate the interaction probability between the feature map vector of the drug being predicted and the feature map vector of the target being predicted. A higher interaction probability indicates a greater likelihood that the target protein is the target of the drug being predicted. Based on this, drug-target pairs with an interaction probability greater than a preset threshold are considered predicted interaction results, while drug-target pairs with an interaction probability less than the preset threshold are considered invalid output. The preset threshold can be set to 0.5 or the lowest value of known target proteins (this value is generally higher than 0.5). After all drug-target pairs are predicted, all predicted interaction results are saved and used for subsequent analysis and application.

[0037] The drug-target interaction prediction method based on contrastive learning and Transformer proposed in the present application first obtains the drug-target sequence to be predicted and characterizes and encodes it into a drug sequence representation to be predicted and a target sequence representation to be predicted; then, the drug sequence representation to be predicted and the target sequence representation to be predicted are input into a pre-trained prediction model; the Transformer in the prediction model is used to perform feature extraction on the drug sequence representation to be predicted and the target sequence representation to be predicted to obtain a drug feature vector to be predicted and a target feature vector to be predicted; the multi-layer perceptron MLP network in the prediction model is used to perform feature mapping on the drug feature vector to be predicted and the target feature vector to be predicted to obtain a drug feature mapping vector to be predicted and a target feature mapping vector to be predicted; finally, the interaction probability between the drug feature mapping vector to be predicted and the target feature mapping vector to be predicted is calculated, and the drug-target pairs with an interaction probability greater than a preset threshold are taken as predicted interaction results.

[0038] This method uses the Transformer architecture to capture the deep-level feature relationship between drugs and targets, and combines contrastive learning to optimize feature representation, enabling the prediction model to more effectively distinguish different drug-target pair combinations. It also introduces the protein language model (PLM) to obtain high-quality protein feature representation, thereby enhancing adaptability to new targets and low-resource targets. This method improves the accuracy and generalization ability of drug-target interaction prediction and can be widely used in new target discovery, drug repositioning, drug screening, and precision medicine. It is expected to reduce experimental screening costs and accelerate the development of innovative drugs. Among them, low-resource targets refer to drug molecules with a small number of target proteins.

[0039] Protein Language Models (PLMs) have become a key research area in bioinformatics, providing a powerful tool for learning protein sequence representations. Drawing on language models from the field of natural language learning, PLMs capture the evolutionary information, structural features, and functional properties of proteins through self-supervised pre-training on large-scale protein sequence data. Compared to traditional methods based on sequence alignment or manual feature construction, PLMs can automatically learn implicit protein features and effectively represent new targets in the absence of structural information or experimental data.

[0040] Considering that the current interaction prediction methods cannot effectively combine the high-dimensional protein features extracted by PLM with the structural characteristics of drugs, and how to optimize the feature representation generated by PLM to make it more suitable for drug-target interaction prediction tasks.

[0041] Based on this, in step 101, the operations of encoding the representation of the drug-target sequence to be predicted are as follows: Morgan fingerprint is used to represent and encode the SMILES sequence of the drug molecule to be predicted, and the representation of the drug sequence to be predicted of the drug molecule sequence to be predicted is obtained; the protein language model PLM is used to represent and encode the target protein sequence to be predicted, and the representation of the target protein sequence to be predicted is obtained.

[0042] The protein language model (PLM) and Morgan fingerprint generation rules are used to encode target proteins and drug molecules, respectively, to achieve more accurate molecular characterization. The protein language model (PLM) can deeply capture the potential functional characteristics of protein sequences, provide a more detailed protein representation, and thus enhance the characterization of target features. At the same time, the Morgan fingerprint encodes the structured representation of drug molecules and extracts the topological features of drug molecules, including ring structures, molecular skeletons, and chemical bond connections, thereby effectively capturing the structural information in drug molecules. This combination enables the predictive model to fully explore the potential laws of drug-target interactions, improving the accuracy and stability of predictions.

[0043] In an optional embodiment of the present application, see Figure 3 、 Figure 4 and Figure 5 , Figure 3 Schematic diagram of the overall workflow of a drug-target interaction prediction method proposed in the examples of this application. Figure 4 is a schematic diagram of the architecture of the prediction model proposed in the embodiment of this application, Figure 5 Schematic diagram of the implementation steps of a drug-target interaction prediction method proposed in an embodiment of the present application. Figure 3The process of training the prediction model and the schematic process of using the prediction model for actual prediction are included in the above step 102. The training process of the prediction model pre-trained in step 102 includes the following steps: Step 1021: Obtain drug-target pair information from a drug-target interaction database, and characterize and encode the drug-target pair information to obtain training data, where the training data includes binary training data and triple training data.

[0044] The acquired drug-target pair information includes the drug molecule sequence and the target protein sequence. After characterizing and encoding the above sequences, the corresponding sequence representation is obtained as training data, wherein the training data includes binary training data and triple training data. The binary training data is used to train the Transformer and MLP networks of the prediction model to be trained and their parameters are optimized. The triple training data is used to train the Transformer and MLP networks of the prediction model to be trained, and the parameters of the MLP network are optimized to obtain the optimized prediction model.

[0045] Step 1022: Input the binary training data into the Transformer in the prediction model to be trained, perform deep feature extraction on the binary training data, and obtain drug feature vectors and target feature vectors.

[0046] The binary training data is input into the prediction model to be trained. It first passes through a feature extractor based on the Transformer architecture, and uses a multi-head attention mechanism to extract the deep feature representations of the drug molecules and target proteins in the binary training data. The cross-attention mechanism is then used to realize the feature interaction from target features to drug features.

[0047] By introducing the multi-head attention mechanism and the cross-attention mechanism, the feature extraction process based on the Transformer architecture is further optimized. The multi-head attention mechanism extracts local and global features from multiple subspaces by computing multiple attention heads in parallel, enabling the prediction model to not only accurately capture the detailed information of the local structure, but also comprehensively consider the macroscopic features of global interactions. This mechanism significantly improves the adaptability and generalization ability of the prediction model in the process of extracting drug and target features, making the prediction model widely applicable to prediction tasks of various types of drug-target pairs. The cross-attention mechanism further enriches the information hierarchy of the extracted feature vectors by realizing dynamic interaction between drug features and target features. Through this deep feature fusion, the prediction model can more comprehensively understand the complex relationship between drugs and targets, thereby improving the robustness and accuracy of the predicted interaction results.

[0048] Step 1023: Based on the feature mapping of the multi-layer perceptron MLP network in the prediction model to be trained, the drug feature vector and the target feature vector are mapped to a high-dimensional feature space to obtain the drug feature mapping vector and the target feature mapping vector.

[0049] The process of drug feature mapping is as follows: the drug feature vector is expressed as , through the MLP network Map it to a high-dimensional shared feature space to obtain the drug feature mapping vector , the process is expressed as follows: .

[0050] The process of target feature mapping is: the target feature vector is expressed as , through the MLP network Map it to a high-dimensional shared feature space to obtain the target feature mapping vector , the process is expressed as follows: .

[0051] Step 1024: Calculate the drug-target interaction probability between the drug feature mapping vector and the target feature mapping vector, and calculate the drug-target loss based on the drug-target interaction probability.

[0052] The drug-target interaction probability is calculated based on the drug feature mapping vector and the target feature mapping vector mapped to the high-dimensional feature space, and the drug-target loss is further calculated based on the interaction probability.

[0053] Step 1025: Optimize the parameters of the Transformer and Multi-Layer Perceptron (MLP) networks in the prediction model to be trained based on the loss between the drug targets to obtain an initially optimized prediction model to be trained.

[0054] Based on the calculated loss between drug targets, the parameters of the Transformer and multi-layer perceptron MLP networks in the prediction model to be trained, such as weights, are first optimized to obtain the initial optimized prediction model to be trained.

[0055] Step 1026: Input the triplet training data into the Transformer of the initially optimized prediction model to be trained to obtain the active compound feature vector, the inactive compound feature vector and the target protein feature vector.

[0056] The process of acquiring triplet training data is the same as that in step 1021, except that the drug molecules corresponding to the target protein are divided into active compounds and inactive compounds. Using this triplet training data can make the mapping distance of the drug-target pair of the active compound closer, while the mapping distance of the drug-target pair of the inactive compound is farther away.

[0057] Step 1027: Map the active compound feature vector, the inactive compound feature vector, and the target protein feature vector to the high-dimensional feature space to obtain an active compound feature mapping vector, an inactive compound feature mapping vector, and a target protein feature mapping vector.

[0058] The operation in step 1027 is the same as that in step 1023 and will not be repeated here.

[0059] Step 1028: Calculate the contrastive learning loss between the active compound feature mapping vector, the inactive compound feature mapping vector, and the target feature mapping vector.

[0060] Through comparative learning between triplets of active compounds (positive), inactive compounds (negative) and target proteins (anchor), the trained prediction model can more clearly distinguish between interaction and non-interaction patterns.

[0061] Specifically, the calculation process of contrastive learning loss is as follows: first, the cosine similarity between the active compound-target protein and the inactive compound-target protein is calculated respectively, and then the contrastive learning loss is calculated based on the cosine similarity between the active compound-target protein and the inactive compound-target protein.

[0062] Step 1029: Optimize the parameters of the multi-layer perceptron (MLP) network in the initially optimized prediction model to be trained based on the contrastive learning loss to obtain the optimized prediction model.

[0063] The optimized prediction model is obtained by optimizing the parameters of the multi-layer perceptron (MLP) network in the initially optimized prediction model to be trained using contrastive learning loss. This training method improves the prediction model's ability to accurately identify true interactions and enables it to demonstrate higher discrimination and accuracy when faced with complex and diverse drug-target relationships. This approach effectively reduces the false positive rate and significantly enhances overall prediction performance.

[0064] In a preferred embodiment of the present application, after the parameters of the prediction model to be trained are optimized as described above, hyperparameters (such as learning rate, number of attention heads, number of Transformer layers, batch size, etc.) are optimized through a grid search method, and the training model with the best AUC (area under the curve) or AUPR (area under the precision-recall curve) indicator is saved as the optimal training model.

[0065] Based on the above embodiment, the process of establishing the drug-target interaction database in the above step 1021 is as follows: first, drug-target pair data is obtained, where the drug-target pair data includes relevant data of drug molecules and target proteins, and the acquisition method includes collecting drug-target pair data from public data sets such as literature materials and ChEMBL database; then, the drug molecules are represented by SMILES sequences to obtain original drug molecule sequences, and the target proteins are represented by amino acid sequences to obtain original target protein sequences; further, the original drug molecule sequences and the original target protein sequences are preprocessed to obtain drug molecule sequences and target protein sequences, where the preprocessing includes data cleaning, duplicate data deletion, sequence standardization and other processing; the drug molecule sequences and target protein sequences obtained by preprocessing are used to establish the drug-target interaction database, and the drug molecule sequences and target protein sequences in the database can be stored in csv or tsv file formats.

[0066] Optionally, the above training data may be obtained from a general database, or may be obtained by combining a general database with a self-built comparison data set as training data.

[0067] Based on the above embodiment, the present application uses PLM to characterize and encode the target protein sequence in step 1021 to optimize the feature representation of the target protein. Drug-target pair information is obtained from the drug-target interaction database and characterized and encoded to obtain training data, wherein the drug-target pair information includes the drug molecule sequence and the target protein sequence. Specifically, the following process is included: First, the Morgan fingerprint is used to characterize and encode the drug molecule sequence to obtain a drug sequence representation. The drug molecule sequence in the embodiment of the present application is the SMILES sequence of the drug molecule; then, the protein language model PLM is used to characterize and encode the target protein sequence to obtain a target sequence representation. For drug-target pairs with interaction values ​​less than a first preset threshold, the label value is set as the first label value. The drug molecule in the drug-target pair corresponding to the first label value is the active compound. The interaction value refers to an experimental or computational metric used to quantify the binding strength or effect between the drug molecule and the target protein. The first preset threshold can be set to 10 μM, and the first label value can be set to 1. For drug-target pairs with interaction values ​​greater than the first preset threshold, the label value is set as the second label value. The drug molecule in the drug-target pair corresponding to the second label value is the inactive compound. The second preset threshold can be set to 30 μM, and the first label value can be set to 0. Finally, all sequences, representation vectors, and label value data are merged and stored in h5 or pickle file format. The drug sequence representation, target sequence representation, and label value are used as binary training data, and the active compound sequence, inactive compound sequence, and target protein sequence are used as triplet training data. The interaction value is data from the database and can be directly obtained. For example, the interaction value between target A and drug A in the ChEMBL database.

[0068] Based on the above embodiment, in the above step 1022, the two-tuple training data is input into the Transformer in the prediction model to be trained, and deep feature extraction is performed on the two-tuple training data to obtain drug feature vectors and target feature vectors, specifically including: Step 10221: Input the drug sequence representation and the target sequence representation into the Transformer, and pass through the linear layer and the linear embedding layer to obtain the drug embedding vector and the target embedding vector.

[0069] The input drug sequence representation is represented as , the linear layer is represented as , the linear embedding layer is expressed as , the drug embedding vector is expressed as , the input target sequence representation is expressed as , the target embedding vector is expressed as , The process of obtaining the drug embedding vector in the above step 10221 is as follows: .

[0070] The process of obtaining the target embedding vector in step 10221 is as follows: .

[0071] Step 10222: The drug embedding vector and the target embedding vector are passed through an encoding network to obtain an initial drug feature vector and the target feature vector.

[0072] The above encoding network is a Transformer encoding network with a multi-head attention mechanism, which is expressed as , the initial drug feature vector is expressed as , the target feature vector is expressed as The process of obtaining the drug feature vector in step 10222 is as follows: .

[0073] The process of obtaining the target feature vector in step 10222 is as follows: .

[0074] Step 10223: Use a cross-attention mechanism to integrate the drug embedding vector and the target embedding vector, optimize the initial drug feature vector, and obtain the drug feature vector.

[0075] The cross attention mechanism is expressed as , integrating the initial drug embedding vector and target embedding vector , optimize the drug feature vector , the process is expressed as follows: .

[0076] Based on the above embodiment, in the above step 1024, calculating the drug-target interaction probability between the drug feature mapping vector and the target feature mapping vector specifically includes: Calculate drug feature map vector and target feature map vector The high-dimensional space distance is calculated by cosine similarity. To characterize and predict the probability of drug-target interaction , the calculation formula is as follows: .

[0077] Based on the above embodiment, in the above step 1024, the loss of drug-target The calculation is performed using the following formula:

[0078] in, is the binary cross entropy loss, is the drug-target interaction probability, is the label value of the drug-target pair.

[0079] Based on the above implementation, in step 1028, the contrastive learning loss is calculated using the following formula:

[0080] in, is the interval loss, is the target protein feature mapping vector, is the active compound feature mapping vector, is the feature map vector of inactive compounds.

[0081] This application proposes a drug-target interaction prediction method that combines the Transformer architecture and contrastive learning technology to improve the accuracy and generalization ability of drug-target interaction prediction. The Transformer architecture is used to enhance the global sequence modeling capabilities of drugs and targets, and the feature representation of drugs and targets is optimized through contrastive learning, so that the prediction model can still efficiently predict new drug-target interaction combinations in low-resource scenarios. In addition, protein characterization is combined with PLM to further improve the accuracy and applicability of target features. The present invention aims to provide an efficient computational tool for target discovery, drug screening and personalized treatment.

[0082] The above-mentioned low-resource scenario refers to a situation where there are few or no corresponding target proteins for drug molecules. The prediction model trained in this application can predict new drug-target interaction combinations in low-resource scenarios.

[0083] The present invention is further described below by way of examples, which however do not limit the scope of the present invention in any way.

[0084] Example 1 Drug repositioning of esketamine to predict its potential new active targets in the human proteome. The implementation process is as follows: Step 1: Build a universal drug-target dataset.

[0085] (1) Download drug-target interaction data in batches from the BindingDB database (https: / / www.bindingdb.org / rwd / bind / chemsearch / marvin / Download.jsp), limit the "Target Source Organism" to "Homo sapiens", limit the interaction data type to "half-maximal inhibitory concentration value / IC50 value", and limit the drug molecular weight to between 200-800 Da. (2) Represent drug molecules with SMILES sequences and target proteins with amino acid sequences, and perform preprocessing such as data cleaning, deduplication, sequence normalization, deletion of entries with missing values, and deletion of entries with inconsistent annotations. (3) Screen and classify drug-target pairs based on the activity threshold (10 μM) and the inactivity threshold (30 μM). For drug-target pairs with an activity less than the set activity threshold, the label value is set to 1, and the drug molecule in the pair is an active compound; for drug-target pairs with an activity greater than the inactivity threshold, the label value is set to 0, and the drug molecule in the pair is an inactive compound. After the above preprocessing, approximately 1.62 million drug-target pair data were obtained.

[0086] Step 2: Construct a general contrastive learning dataset.

[0087] The DUDe (https: / / dude.docking.org / ) database is a standardized database for drug discovery research needs such as molecular docking. It contains 102 measured active molecules and specially constructed inactive molecules for target proteins. Inactive molecules are called decoy molecules. They are similar to active molecules in physical and chemical properties, but have significant differences in two-dimensional topological structure. (1) Target, active molecule and decoy molecule data were downloaded in batches from the DUDe database. The protein species of active molecules were limited to "human proteins" among homologous proteins. (2) Drug molecules were represented by SMILES sequences and target proteins were represented by amino acid sequences. Since it is a standard library, only sequence normalization was performed. (3) The statistics of the final universal comparative learning dataset are shown in Table 1.

[0088] Step 3: Use the constructed training dataset to establish a drug-target interaction prediction model based on contrastive learning and Transformer architecture.

[0089] Structurally, the model consists of two parts: a feature extractor and a feature mapping layer. The trained model with the best AUC or AUPR on the validation dataset is selected as the optimal trained model, and its performance on the test dataset is saved as the model performance indicator. During the training phase, the model's hyperparameters (learning rate, number of attention heads, number of Transformer layers, batch size, etc.) are manually adjusted multiple times to select the optimal hyperparameter combination. This example performs debugging from the following aspects: The candidate ranges for learning rate during model training are: 0.01, 0.001, 0.0001, 0.00001; the candidate ranges for batch size are: 32, 64, 128, 256, 512; the candidate ranges for number of iterations are: 20, 50, 100, 1000.

[0090] The feature extractor converts drug / target input representations into drug / target feature vectors, including a multi-head attention module and a cross-attention module. The candidate number of neurons in the embedding layer ranges from 200, 512, and 1024; the candidate number of TransformerEncoder layers in the multi-head attention module ranges from 1 to 2; the candidate number of attention heads in the multi-head / cross-attention module ranges from 4 to 8; the candidate number of neuron expansion multiples in the feedforward neural network of the multi-head / cross-attention module ranges from 2 to 4; the candidate dropout rate in the multi-head / cross-attention module ranges from 0.1, 0.2, 0.3, and 0.5; and the candidate number of neurons in the output layer ranges from 100, 200, and 1024.

[0091] The feature mapping layer projects drug / target feature vectors into a shared high-dimensional feature space. During each training iteration, training is performed using the common drug-target dataset constructed in the first step, followed by training using the contrastive learning dataset constructed in the second step. Candidate ranges for the number of neurons in the MLP network include 100, 200, and 1024; and candidate ranges for contrastive learning batch sizes include 1024, 2048, 4096, and 8192.

[0092] Step 4: Use the optimal training model obtained from the previous step to predict the interaction probability for the drug-target pair consisting of esketamine and the human proteome.

[0093] (1) Obtaining human proteome sequences: Download the human proteome target sequences in batches from the UniProtKB database on the Uniprot website (https: / / www.uniprot.org / ), restricting "Species / Popular organisms" to "Human / " and "Verification Status / Status" to "Verified / Reviewed". Duplicate values ​​were then removed based on the Unprot ID, resulting in a total of 20,433 protein sequences. (2) The input representation of the drug-target sequence pair to be predicted was converted into a drug / target feature vector using a Transformer-based feature extractor, and then projected into a shared high-dimensional feature space using an MLP-based feature mapping layer to calculate the interaction probability of the drug-target pair. (3) The threshold was set to the lowest value among the predicted interaction values ​​of esketamine's reported active targets "P41145 / OPRK, P35372 / OPRM, P31645 / SC6A4" with esketamine. (4) Compare the model prediction value with the set threshold, output the interaction result data of the drug-target interaction prediction that is higher than the set threshold and save it.

[0094] Table 1 Statistics of the general contrastive learning dataset

[0095] Example 2 Drug repositioning of cannabidiol (CBD) involves predicting its potential new active targets in the human proteome. The implementation process is as follows: The first step is the same as that in Example 1 and will not be described again here.

[0096] Step 2: Construct a cannabidiol bias contrast learning dataset.

[0097] (1) Reported cannabidiol active targets were collected from literature and ChEMBL databases. Targets with activity values ​​less than 10 μM were classified as high-activity cannabidiol targets (as shown in Table 2), and targets with activity values ​​greater than 10 μM and less than 30 μM were classified as low-activity cannabidiol targets (as shown in Table 3). (2) All compounds with high-activity cannabidiol targets were collected from the ChEMBL database for activity testing. The raw data obtained were checked and selected according to the following standard process: 1) The compounds with IC 50 Delete compounds with “Not Determined” and “Not Active” clearly marked; Delete IC50 Compounds with uncertain values ​​and poorly defined ranges (e.g., compounds with a range less than 100 μM, compounds with a range greater than 1 μM). 2) Remove duplicate compounds. 3) Check the IC values ​​of the compounds against the original reference. 50 4) Compounds were screened and classified according to an activity threshold (10 μM) and an inactivity threshold (30 μM). Compounds with activity levels below the set activity threshold were defined as active, while those with activity levels above the set inactivity threshold were defined as inactive. 5) RDKit was used to convert the collected compounds into standard SMILES strings, and molecules that could not be processed by RDKit were removed. After the above data collection, organization, and cleaning process, the cannabidiol bias comparative learning dataset was obtained. The statistical data are shown in Tables 2 and 3.

[0098] The third step is the same as that in Example 1.

[0099] Step 4: Use the optimal training model obtained from the previous step to predict the interaction probability for the drug-target pair consisting of esketamine and the human proteome.

[0100] (1) The same process as in Example 1. (2) The same process as in Example 1. (3) The threshold was set to the lowest value among the predicted interaction values ​​of cannabidiol with the reported low-activity cannabidiol targets (as shown in Table 3). (4) The same process as in Example 1.

[0101] Table 2 Statistics of highly active cannabidiol (CBD) targets

[0102] Table 3 Statistics of low-activity cannabidiol (CBD) targets

[0103] *GPR3 activity values ​​belong to the highly active cannabidiol (CBD) target, but no active and inactive molecules meeting the criteria were reported, so it was included in the validation set. **TRPV2 activity is slightly below the inactive threshold of 30 μM, but TRPV1, a member of the same family, is active and therefore was also included. The present invention proposes a drug-target interaction prediction method based on contrastive learning and Transformer architecture, which improves the model's learning ability for drug-target interaction patterns through efficient feature extraction, optimized feature mapping and shared high-dimensional feature space construction. The Transformer structure is used to perform deep feature extraction on drug and target sequences, and the protein language model PLM is combined to extract high-quality target protein representations. The feature differentiation ability of the prediction model is enhanced through a contrastive learning strategy, so that the prediction model still has strong prediction capabilities in the case of low-resource targets or unknown targets or unknown drugs. Compared with traditional drug-target interaction prediction methods, the present invention not only improves the prediction accuracy of known drug-target pairs, but also enhances the model's generalization ability for new targets and new drugs, providing intelligent modeling tools for virtual screening, target discovery and precision medicine. An unknown target refers to a drug molecule for which no target protein has been found.

[0104] In the second aspect, the present application embodiment proposes a drug-target interaction prediction device based on contrastive learning and Transformer, see Figure 6 , Figure 6 : This is a functional module diagram of a drug-target interaction prediction device based on contrastive learning and Transformer proposed in an embodiment of the present application, the device comprising: An acquisition module 601 is used to acquire a drug-target sequence to be predicted and characterize and encode it into a drug sequence representation to be predicted and a target sequence representation to be predicted; An input module 602 is used to input the drug sequence representation to be predicted and the target sequence representation to be predicted into a pre-trained prediction model; A feature extraction module 603 is configured to extract features from the sequence representation of the drug to be predicted and the sequence representation of the target to be predicted using the Transformer in the prediction model to obtain a feature vector of the drug to be predicted and a feature vector of the target to be predicted; A feature mapping module 604 is configured to perform feature mapping on the drug feature vector to be predicted and the target feature vector to be predicted using a multi-layer perceptron (MLP) network in the prediction model to obtain a drug feature mapping vector to be predicted and a target feature mapping vector to be predicted; The calculation prediction module 605 is used to calculate the interaction probability between the drug feature mapping vector to be predicted and the target feature mapping vector to be predicted, and take the drug-target pair with an interaction probability greater than a preset threshold as the predicted interaction result.

[0105] Wherein, the input module includes a training module, and the training module includes: an acquisition submodule, configured to acquire drug-target pair information from a drug-target interaction database, and to represent and encode the drug-target pair information to obtain training data, wherein the training data includes binary training data and triple training data; A first feature extraction submodule is used to input the binary training data into the Transformer in the prediction model to be trained, perform deep feature extraction on the binary training data, and obtain a drug feature vector and a target feature vector; A first feature mapping submodule is configured to map the drug feature vector and the target feature vector to a high-dimensional feature space based on the feature mapping of the multi-layer perceptron (MLP) network in the prediction model to be trained, to obtain the drug feature mapping vector and the target feature mapping vector; a first calculation submodule, configured to calculate a drug-target interaction probability between the drug feature mapping vector and the target feature mapping vector, and calculate a drug-target loss based on the drug-target interaction probability; A first optimization submodule is used to optimize the parameters of the Transformer and multi-layer perceptron (MLP) networks in the prediction model to be trained based on the loss between the drug targets to obtain an initially optimized prediction model to be trained; A second feature extraction submodule is used to input the triplet training data into the Transformer of the initially optimized prediction model to be trained to obtain an active compound feature vector, an inactive compound feature vector, and a target protein feature vector; A second feature mapping submodule is used to map the active compound feature vector, the inactive compound feature vector, and the target protein feature vector to the high-dimensional feature space to obtain an active compound feature mapping vector, an inactive compound feature mapping vector, and a target protein feature mapping vector; A second calculation submodule is configured to calculate a contrastive learning loss between the active compound feature mapping vector, the inactive compound feature mapping vector, and the target feature mapping vector; The second optimization submodule is used to optimize the parameters of the multi-layer perceptron MLP network in the initially optimized prediction model to be trained based on the contrastive learning loss to obtain the optimized prediction model.

[0106] The acquisition submodule includes an establishment submodule, and the establishment submodule includes: a data acquisition unit, configured to acquire drug-target pair data, wherein the drug-target pair data includes drug molecules and target proteins; A sequence representation unit, used to represent the drug molecule using a SMILES sequence to obtain an original drug molecule sequence, and to represent the target protein using an amino acid sequence to obtain an original target protein sequence; a preprocessing unit, configured to preprocess the original drug molecule sequence and the original target protein sequence to obtain a drug molecule sequence and a target protein sequence; An establishing unit is used to establish the drug-target interaction database using the drug molecule sequence and the target protein sequence.

[0107] Wherein, the acquisition submodule includes: A first characterization encoding unit is used to characterize and encode the drug molecule sequence using Morgan fingerprint to obtain a drug sequence characterization; A second representation encoding unit is used to represent and encode the target protein sequence using a protein language model (PLM) to obtain a target sequence representation; a first setting unit, configured to set a label value of a drug-target pair whose interaction value is less than a first preset threshold as a first label value, wherein the drug molecule in the drug-target pair corresponding to the first label value is an active compound; a second setting unit, configured to set the label value of the drug-target pair whose interaction value is greater than the first preset threshold as a second label value, wherein the drug molecule in the drug-target pair corresponding to the second label value is an inactive compound; The training data determination unit is used to use the drug sequence representation, target sequence representation and label value as the two-tuple training data, and use the active compound sequence, inactive compound sequence and target protein sequence as the three-tuple training data.

[0108] Wherein, the first feature extraction submodule includes: An embedding unit, configured to input the drug sequence representation and the target sequence representation into a Transformer, and obtain a drug embedding vector and a target embedding vector through a linear layer and a linear embedding layer; An encoding unit, configured to pass the drug embedding vector and the target embedding vector through an encoding network to obtain an initial drug feature vector and the target feature vector; An integration unit is used to integrate the drug embedding vector and the target embedding vector using a cross attention mechanism, optimize the initial drug feature vector, and obtain the drug feature vector.

[0109] The first calculation submodule includes: Probability calculation unit, used to calculate drug feature mapping vector and target feature map vector The high-dimensional space distance is calculated by cosine similarity. To characterize and predict the probability of drug-target interaction , the calculation formula is as follows: .

[0110] The first calculation submodule further includes: The loss calculation unit is used to calculate the drug-target loss using the following calculation formula :

[0111] in, is the binary cross entropy loss, is the drug-target interaction probability, is the label value of the drug-target pair.

[0112] The second calculation submodule includes: The contrastive learning loss calculation unit is used to calculate the contrastive learning loss using the following formula:

[0113] in, is the interval loss, is the target protein feature mapping vector, is the active compound feature mapping vector, is the feature map vector of inactive compounds.

[0114] Based on the same application concept, the embodiment of the present application discloses an electronic device in a third aspect. Figure 7 A schematic diagram of an electronic device disclosed in an embodiment of the present application is shown. Figure 7 As shown, the electronic device 100 includes: a memory 110 and a processor 120. The memory of the electronic device is not less than 12G, the main frequency of the processor is not less than 2.4GHz, the memory 110 and the processor 120 are connected through a bus communication, and a computer program is stored in the memory 110. The computer program can be run on the processor 120 to implement a drug-target interaction prediction method based on contrastive learning and Transformer disclosed in an embodiment of the present application.

[0115] Based on the same application concept, an embodiment of the present application also discloses a computer-readable storage medium on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, a drug-target interaction prediction method based on contrastive learning and Transformer disclosed in an embodiment of the present application is implemented.

[0116] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0117] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods, apparatuses, electronic devices, and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0118] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0119] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0120] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they become aware of the basic inventive concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present invention.

[0121] Finally, it should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or terminal device that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or elements that are inherent to such process, method, article, or terminal device. In the absence of further restrictions, an element defined by the phrase "comprises a ..." does not exclude the presence of additional identical elements in the process, method, article, or terminal device that includes the element.

[0122] The above is a detailed introduction to the drug-target interaction prediction method, device and equipment based on contrastive learning and Transformer provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core ideas. At the same time, for general technical personnel in this field, according to the ideas of the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A drug-target interaction prediction method based on contrastive learning and Transformer, characterized in that: The method comprises: Obtaining a drug-target sequence to be predicted and encoding the sequence representation as a drug sequence representation to be predicted and a target sequence representation to be predicted; Inputting the drug sequence representation to be predicted and the target sequence representation to be predicted into a pre-trained prediction model; Performing feature extraction on the sequence representation of the drug to be predicted and the sequence representation of the target to be predicted by using the Transformer in the prediction model to obtain a feature vector of the drug to be predicted and a feature vector of the target to be predicted; Performing feature mapping on the drug feature vector to be predicted and the target feature vector to be predicted by a multi-layer perceptron (MLP) network in the prediction model to obtain a drug feature mapping vector to be predicted and a target feature mapping vector to be predicted; The interaction probability between the drug feature mapping vector to be predicted and the target feature mapping vector to be predicted is calculated, and the drug-target pair with an interaction probability greater than a preset threshold is taken as the predicted interaction result.

2. The drug-target interaction prediction method based on contrastive learning and Transformer according to claim 1, characterized in that: The training process of the pre-trained prediction model includes: Obtaining drug-target pair information from a drug-target interaction database, and characterizing and encoding the drug-target pair information to obtain training data, wherein the training data includes binary training data and triple training data; Inputting the binary training data into the Transformer in the prediction model to be trained, performing deep feature extraction on the binary training data to obtain a drug feature vector and a target feature vector; Based on the feature mapping of the multi-layer perceptron MLP network in the prediction model to be trained, mapping the drug feature vector and the target feature vector to a high-dimensional feature space to obtain the drug feature mapping vector and the target feature mapping vector; calculating a drug-target interaction probability between the drug feature mapping vector and the target feature mapping vector, and calculating a drug-target loss based on the drug-target interaction probability; Optimizing the parameters of the Transformer and Multi-Layer Perceptron (MLP) networks in the prediction model to be trained based on the loss between the drug targets to obtain an initially optimized prediction model to be trained; Inputting the triplet training data into the Transformer of the initially optimized prediction model to be trained to obtain an active compound feature vector, an inactive compound feature vector, and a target protein feature vector; Mapping the active compound feature vector, the inactive compound feature vector, and the target protein feature vector to the high-dimensional feature space to obtain an active compound feature mapping vector, an inactive compound feature mapping vector, and a target protein feature mapping vector; calculating a contrastive learning loss between the active compound feature map vector, the inactive compound feature map vector, and the target feature map vector; The parameters of the multi-layer perceptron (MLP) network in the initially optimized prediction model to be trained are optimized based on the contrastive learning loss to obtain the optimized prediction model.

3. The drug-target interaction prediction method based on contrastive learning and Transformer according to claim 2, characterized in that: The process of establishing the drug-target interaction database: Acquiring drug-target pair data, wherein the drug-target pair data includes a drug molecule and a target protein; The drug molecule is represented by a SMILES sequence to obtain the original drug molecule sequence, and the target protein is represented by an amino acid sequence to obtain the original target protein sequence; Preprocessing the original drug molecule sequence and the original target protein sequence to obtain a drug molecule sequence and a target protein sequence; The drug-target interaction database is established using the drug molecule sequence and the target protein sequence.

4. The drug-target interaction prediction method based on contrastive learning and Transformer according to claim 2, characterized in that: Drug-target pair information includes drug molecule sequence and target protein sequence. Drug-target pair information is obtained from a drug-target interaction database, and the drug-target pair information is represented and encoded to obtain training data, including: Morgan fingerprints are used to characterize and encode the drug molecular sequence to obtain a drug sequence characterization; Using a protein language model (PLM) to represent and encode the target protein sequence to obtain a target sequence representation; Setting the label value of the drug-target pair whose interaction value is less than a first preset threshold as a first label value, wherein the drug molecule in the drug-target pair corresponding to the first label value is an active compound; Setting the label value of the drug-target pair whose interaction value is greater than the first preset threshold as a second label value, wherein the drug molecule in the drug-target pair corresponding to the second label value is an inactive compound; The drug sequence representation, target sequence representation and label value are used as the binary training data, and the active compound sequence, inactive compound sequence and target protein sequence are used as the triplet training data.

5. The drug-target interaction prediction method based on contrastive learning and Transformer according to claim 4, characterized in that: The binary training data is input into the Transformer in the prediction model to be trained, and deep feature extraction is performed on the binary training data to obtain drug feature vectors and target feature vectors, including: Inputting the drug sequence representation and the target sequence representation into the Transformer, passing through the linear layer and the linear embedding layer, to obtain the drug embedding vector and the target embedding vector; The drug embedding vector and the target embedding vector are passed through an encoding network to obtain an initial drug feature vector and the target feature vector; The drug embedding vector and the target embedding vector are integrated by using a cross attention mechanism, and the initial drug feature vector is optimized to obtain the drug feature vector.

6. The drug-target interaction prediction method based on contrastive learning and Transformer according to claim 2, characterized in that: Calculate the drug-target interaction probability between the drug feature map vector and the target feature map vector, including: Calculate drug feature map vector and target feature map vector The high-dimensional space distance is calculated by cosine similarity. To characterize and predict the probability of drug-target interaction , the calculation formula is as follows: 。 7. The drug-target interaction prediction method based on contrastive learning and Transformer according to claim 2, characterized in that: The drug-target loss is calculated using the following formula: : in, is the binary cross entropy loss, is the drug-target interaction probability, is the label value of the drug-target pair.

8. The drug-target interaction prediction method based on contrastive learning and Transformer according to claim 2, characterized in that: The contrastive learning loss is calculated using the following formula: in, is the interval loss, is the target protein feature mapping vector, is the active compound feature mapping vector, is the feature map vector of inactive compounds.

9. A drug-target interaction prediction device based on contrastive learning and Transformer, characterized in that: The device comprises: An acquisition module is used to acquire the drug-target sequence to be predicted and represent and encode it into a representation of the drug sequence to be predicted and a representation of the target sequence to be predicted; An input module, configured to input the drug sequence representation to be predicted and the target sequence representation to be predicted into a pre-trained prediction model; A feature extraction module is used to extract features from the sequence representation of the drug to be predicted and the sequence representation of the target to be predicted through the Transformer in the prediction model to obtain a feature vector of the drug to be predicted and a feature vector of the target to be predicted; A feature mapping module is used to perform feature mapping on the drug feature vector to be predicted and the target feature vector to be predicted through a multi-layer perceptron (MLP) network in the prediction model to obtain a drug feature mapping vector to be predicted and a target feature mapping vector to be predicted; The calculation prediction module is used to calculate the interaction probability between the feature mapping vector of the drug to be predicted and the feature mapping vector of the target to be predicted, and take the drug-target pair with an interaction probability greater than a preset threshold as the predicted interaction result.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the drug-target interaction prediction method based on contrastive learning and Transformer according to any one of claims 1 to 8.

Images