Method for predicting drug target interaction based on protein sequence

Abstract

The application discloses a drug target interaction prediction method based on protein sequences, comprising the following steps: constructing a drug target database; constructing a target feature extractor based on a pre-trained protein language model; constructing a drug feature extractor based on a pre-trained molecular model; constructing a feature learning method by using a self-attention mechanism, constructing a drug target interaction learning method by using a cross-attention mechanism, constructing a predictor in combination with a linear layer, and building a prediction drug target interaction model; training and optimizing parameters of the drug target interaction prediction model to obtain an optimized drug target prediction model; inputting the drug target into the optimized prediction model to output a prediction result; and finally visualizing the attention score to mark potential active sites and important amino acid residues of the drug and the target. The application realizes the improvement of the prediction accuracy of the interaction between the drug and the target based on the protein sequence, and realizes the explainability of the deep learning model.

Description

Drug target interaction prediction method based on protein sequence Technical Field

[0001] This invention belongs to the field of deep learning and virtual drug screening, and relates to methods for predicting drug-target interactions, particularly to methods for predicting drug-target interactions based on protein sequences. Background Technology

[0002] Virtual drug screening is an important method for identifying novel scaffold precursor molecules and lead compounds. Protein structure-based virtual screening requires the three-dimensional structure of the drug target, while ligand-based virtual screening requires known active molecules. Given the limitations of existing methods, protein sequence-based methods can overcome these limitations and have broader application value.

[0003] Artificial intelligence is widely used in predicting drug-target interactions, treating it as a binary classification task. Traditional machine learning methods predict based on predefined descriptors, while deep learning automatically learns the features of drugs and targets to make predictions. Deep learning typically utilizes neural network architectures, such as convolutional neural networks or graph neural networks, to capture complex features and relationships, thus outperforming traditional machine learning methods in prediction accuracy.

[0004] However, deep learning still faces the following problems in predicting drug-target interactions: 1) the feature-independent representations of drugs and targets cannot effectively learn their interactions; 2) the model has insufficient generalization ability to unseen target systems; and 3) the prediction results of deep learning models lack interpretability. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting drug-target interactions based on protein sequences. This method constructs a drug-target interaction database, extracts target and drug features using pre-trained protein language models and pre-trained molecular characterization models, and employs self-attention and cross-attention mechanisms to build a predictive model of drug-target interactions. Simultaneously, by combining attention score visualization, the interpretability of the model is analyzed, and potential sites of action for drugs and targets are labeled, thereby improving the model's biological interpretability.

[0006] The main technical solution of this invention includes the following steps:

[0007] (1) Construct a drug-target interaction database to obtain relevant values ​​of target-drug pairs, including IC50. 50 EC 50 ,K i and K dAnd obtain the UniProt sequence of the corresponding protein based on the target name and the drug SMILES sequence based on the drug ID;

[0008] (2) Construct a target feature extractor based on a pre-trained protein language model to extract feature vectors of target protein sequences;

[0009] (3) Construct a drug feature extractor based on a pre-trained molecular characterization model to extract feature vectors of drug molecules;

[0010] (4) A feature learning method is constructed by using a multi-head self-attention mechanism, a learning method for drug target interaction is constructed by using a cross-attention mechanism, and a predictor is constructed by combining a linear layer to build a model for predicting drug target interaction.

[0011] 5) Using the drug target database constructed in step 1), train and optimize the parameters of the drug target interaction prediction model to obtain the optimized drug target prediction model.

[0012] 6) Input the UniProt sequence of the target protein to be tested and the SMILES of the drug molecule into the optimized drug target prediction model, output the prediction results, visualize them through attention scores, mark the possible active sites of the drug and the target, perform interpretability analysis, and mark the important amino acid residues of the target-drug interaction.

[0013] In step (1), the acquisition of the drug target interaction database specifically includes: aligning the collected target data according to UniProt, with each target corresponding to a UniProt sequence; and aligning the collected drug data according to the chemical information package, with each drug corresponding to a SMILES sequence.

[0014] In step (2), the target feature extractor based on the pre-trained protein language model includes: calling a linear layer according to the open-source pre-trained protein language model to extract features from the input target UniProt sequence and output target features in amino acid units.

[0015] In step (3), the construction of a drug feature extractor based on a pre-trained molecular model includes: calling a linear layer based on an open-source pre-trained molecular model to extract features from the input drug SMILES and output drug features in units of atoms.

[0016] In step (4), the method for constructing drug target feature learning using a multi-head attention mechanism and the method for constructing drug target interaction learning using a cross-attention mechanism specifically include:

[0017] 1) Three different linear layers are used to map the acquired target sequence features P to Q. P ,K P V P Three sets of vectors are used to map the acquired drug features D to Q using three different linear layers. D ,K D V D The attention scores are calculated using the following attention formula for the three sets of vectors:

[0018]

[0019] Where softmax is the normalization function, the obtained Attention(Q) P ,K P V P ) represents the attention score of the self-attention of the target sequence;

[0020]

[0021] Among them, the obtained Attention(Q) D ,K D V D ) represents the attention score of the drug's self-attention. To further enhance the model's representational ability, a multi-head attention mechanism was adopted, consisting of h parallel heads:

[0022] MultiHead(Q,K,V)=Concat[head1,head2,…,head h W O

[0023]

[0024] in, W is the projection matrix for each head. O It is the output projection matrix, and MultiHead is the score of multi-head attention.

[0025] 2) The target sequence and drug features extracted by the multi-head attention mechanism are projected into the query (Q), key (K), and value (V) spaces, respectively, and then linearly transformed using a learnable weight matrix. The specific formula is as follows:

[0026] For querying, key-value projection of target sequence features:

[0027]

[0028] For drug characteristic queries, key-value projection:

[0029]

[0030] in, It is a query, key, and value weight matrix of target sequence features. It is a weighted matrix of keys and values ​​representing drug features. and It is the bias vector.

[0031] 3) Further study the interaction between the drug and the target, again using the multi-head attention mechanism. For each attention head h, the following calculations are performed:

[0032]

[0033] Each attention head h independently performs attention calculations, obtaining attention outputs from drug to target and from target to drug, respectively. The outputs of all attention heads are concatenated and projected through a linear transformation to obtain the final attention features.

[0034]

[0035] in, This is the output projection matrix from drug to target and from target to drug attention. The attention matrix A, obtained through a cross-attention mechanism, is used to merge features, and then a linear layer outputs the probability of drug-target interaction.

[0036] P = σ(ReLU(AW1+b1)W2+b2)

[0037] Here, ReLU(·) is the activation function applied to the linear transformation result of the hidden layer. W1 and b1 are the weight matrices and bias vectors of the hidden layer. W2 and b2 are the weight matrices and bias vectors of the output layer. σ is the sigmoid function, which converts the output into a probability value P, representing the predicted probability of drug-target interaction.

[0038] In step (5), the drug target database is constructed using step (1), and the parameters of the drug target interaction prediction model are trained and optimized, specifically including:

[0039] 1) Input the UniProt sequence of the target into an open-source protein language model to extract features, and obtain sequence features in amino acid units;

[0040] 2) Input the drug's SMILES into an open-source molecular pre-training model to extract features, and obtain drug features at the atom level.

[0041] 3) A feature learning method is constructed using a self-attention mechanism, and a drug target interaction learning method is constructed using a cross-attention mechanism to update features;

[0042] 4) Input the updated features into a predictor built on a linear layer to obtain the probability results of drug-target interaction;

[0043] 5) The cross-loss entropy function is used to calculate the loss value. Based on minimizing the loss value, the parameters of the drug target prediction model are optimized. The early stopping method is used to obtain the drug target prediction model with the optimal parameters, which is the optimized drug target prediction model.

[0044] In step 6), the target protein sequence and drug molecule SMILES are input into the optimized drug target prediction model, and the prediction results are output. The model is visualized using attention scores, and potential interaction sites between the drug and the target are marked. Interpretability analysis is performed, and important amino acid residues involved in the target-drug interaction are labeled, specifically including:

[0045] 1) The optimized drug target prediction model in step (6) is used to predict the actual drug target effect. During the model prediction process, the attention score of the interaction with the target amino acid as the query vector is retained. The attention score contains the attention information of each amino acid to each drug. The score is plotted as a heat map to detect the key amino acids of drug interaction.

[0046] 2) Save the interaction attention scores with drug atoms as query vectors. The attention scores contain the interaction information of each drug atom and each amino acid. Plot the scores into a heatmap to detect the important and key pharmacophores of the drug.

[0047] Compared with existing technologies, the advantages of this invention are as follows:

[0048] (1) Effectively learn the interaction between drugs and targets: Through the feature update of self-attention and the feature fusion of cross-attention, the interaction between drugs and targets is effectively learned, thereby improving the accuracy of drug-target interaction prediction.

[0049] (2) Improve generalization ability: By introducing pre-trained protein language models and pre-trained molecular models, improve the generalization ability on unseen targets and molecules.

[0050] (3) Achieve model interpretability: By adopting the attention mechanism and combining it with the visualization of attention scores, the key action sites of drugs and targets are clearly marked, biological information is provided, and the application value of the model in actual drug development is enhanced.

[0051] (4) Wide range of applications: This method does not depend on the three-dimensional structure of the target protein and is suitable for situations where the target structure is unknown, thus expanding the application range of drug target interaction prediction. Attached Figure Description

[0052] Figure 1: Flowchart of a method for predicting drug target interactions based on protein sequences.

[0053] Figure 2: Schematic diagram of feature extractors based on pre-trained protein language models and pre-trained molecular models.

[0054] Figure 3: Schematic diagram of the drug target interaction prediction model based on self-attention and cross-attention mechanisms.

[0055] Figure 4: Results of the enrichment index of 0.5% for the DEKOIS2.0 dataset (HitScreen is the method of this invention).

[0056] Figure 5: Results of virtual screening test of disordered domain NTD sequences in androgen receptor (AR).

[0057] Figure 6: Schematic diagram of the interaction between the Pim-1 receptor structure and the inhibitor UNQ molecule and attention score heatmap. Detailed Implementation

[0058] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein.

[0059] Example 1

[0060] Referring to Figure 1, the method of the present invention includes the following steps:

[0061] Step (1) Construction of the drug target interaction database:

[0062] First, drug-target pair interaction data, including IC50, were collected from public databases BindingDB, ChEMBL, and DrugBank. 50 EC 50 ,K i and K d Numerical values. Converting the active label to pChEMBL values ​​(IC). 50 EC 50 ,K i and K dThe negative logarithm of the pChEMBL value was used to identify positive molecules (activity ≤ 100 nM) and negative molecules (activity ≥ 10,000 nM). A 100-fold activity difference was ensured between positive and negative molecules. The corresponding protein sequences were obtained from UniProt based on the target, and the corresponding SMILES were obtained from the drug ID. The SMILES were standardized using the cheminformatics tool RDKit, including removing unnecessary hydrogen atoms and adjusting isomer representation. Filtering criteria included removing molecules with more than 70 heavy atoms to avoid excessive computational complexity and removing protein sequences longer than 2000 amino acid residues to ensure the model's efficiency and effectiveness in processing protein sequences. Finally, the collected data were organized into a unified format to construct a drug-target interaction database.

[0063] Step (2) Construct a target feature extractor based on a pre-trained protein language model:

[0064] As shown in Figure 2, the open-source pre-trained protein language model ESM-2 is used to extract features from the UniProt sequence of the target. Specifically, the protein sequence is input into the model, passes through multiple Transformer encoders, and the result of the last layer is extracted, outputting a feature vector for each amino acid. Subsequently, a linear layer is called to perform a linear transformation on these feature vectors to obtain the final target feature representation.

[0065] Where, N P The number of amino acids in the protein sequence is represented by the ESM-2 model, which is ESM2 t30 150M, with an embedding vector dimension of 1280.

[0066] Step (3) Construct a drug feature extractor based on a pre-trained molecular characterization model:

[0067] As shown in Figure 2, the open-source pre-trained molecular language model Uni-Mol is used to extract features from the SMILES sequence of the drug. Specifically, the SMILES sequence is input into the model, passes through multiple Transformer encoders, and the result of the last layer is extracted, outputting the feature vector of each atom. A linear layer is then called to perform a linear transformation on these feature vectors to obtain the feature representation of the drug.

[0068] Where, N a This represents the number of atoms in a drug molecule, and the embedding vector has a dimension of 512.

[0069] Step (4) Construct a drug target interaction feature fusion and predictor using multi-head self-attention and cross-attention mechanisms:

[0070] As shown in Figure 3, the target sequence features P and drug features D are first input into a multi-head self-attention mechanism to generate query (Q), key (K), and value (V) vectors, respectively. Specifically, three different linear layers are invoked to map the target features P to Q. P、 K P、 V P, Drug characteristic D is mapped to Q D ,K D V D ,

[0071]

[0072]

[0073] To enhance feature representation capabilities, a multi-head attention mechanism is employed, using a total of 8 heads:

[0074] MultiHead(Q,K,V)=Concat[head1,head2,…,head8]W O

[0075]

[0076] in, W is the projection matrix for each head. O This is the output projection matrix, and MultiHead represents the score of multi-head attention. The target sequence and drug features extracted by the multi-head attention mechanism are then regenerated into query (Q), key (K), and value (V) vectors, respectively. A linear transformation is performed using a learnable weight matrix, where the hidden layer has a dimension of 128. The specific formula is as follows:

[0077] For querying, key-value projection of target sequence features:

[0078]

[0079] For drug characteristic queries, key-value projection:

[0080]

[0081] in, It is a weight matrix of the target features, keys, and values. It is a weighted matrix of keys and values ​​representing drug features. and It is the bias vector.

[0082] To capture diverse interactions, a multi-head attention mechanism is employed. For each attention head h, the following calculations are performed:

[0083]

[0084] Here, h represents the number of attention heads, which is 8. Each attention head h independently performs attention calculations to obtain the attention output from drug to target and from target to drug.

[0085] The outputs of all attention heads are concatenated and projected using a linear transformation to obtain the final attention features:

[0086]

[0087] in It is the output projection matrix of drug-to-target and target-to-drug attention.

[0088] Step 4.5: Model Prediction Results

[0089] As shown in Figure 3, the attention matrix A obtained through the cross-attention mechanism is used to merge features, and then a linear layer is used to output the probability of drug-target interaction:

[0090] P = σ(ReLU(AW1+b1)W2+b2)

[0091] ReLU(·) is the activation function applied to the linear transformation result of the hidden layer. W1 and b1 are the weight matrices and bias vectors of the hidden layer. W2 and b2 are the weight matrices and bias vectors of the output layer. σ is the sigmoid function, which converts the output into a probability value P, representing the predicted probability of drug-target interaction.

[0092] Step (5) Train and optimize the parameters of the drug target interaction prediction model:

[0093] The drug target interaction database constructed in step 1 was used as training data. The training and validation sets were randomly divided in an 8:2 ratio to train the prediction model. During training, the cross-entropy loss function was used to calculate the loss value, and the Adam optimization algorithm was employed. The learning rate was set to 0.0001, β1 = 0.9, and β2 = 0.999. The batch size was set to 16, and the maximum number of training epochs was 40. An early stopping method was used: training was stopped when the loss on the validation set no longer decreased within 10 consecutive epochs, resulting in the optimal drug target prediction model.

[0094] Step (6) Input the UniProt sequence of the target protein to be tested and the SMILES drug molecule, output the prediction results and perform interpretability analysis.

[0095] During model prediction, attention scores are recorded with the target amino acid as the query vector. These attention scores reflect the degree of attention each amino acid receives from the drug molecule. These scores are plotted as a heatmap to visually represent potential key interaction sites between the drug and the target. Analysis of the attention scores identifies amino acid residues that significantly influence drug-drug interactions. These key residues can serve as important reference points for drug design and optimization, providing a basis for further experimental validation.

[0096] As shown in Figure 4, the performance was evaluated using the publicly available dataset DEKOIS2.0. DEKOIS2.0 contains 81 targets, including kinases, proteases, nuclear receptors, G protein-coupled receptors, oxidoreductases, transferases, and hydrolases. Each target has an average of 40 active molecules and 1200 inactive molecules. Using enrichment of 0.5% as an indicator, the performance of our developed model HitScreen exceeded that of existing benchmark models.

[0097] Example 1 illustrates virtual screening using the NTD sequence of the disordered domain in the androgen receptor (AR).

[0098] First, the UniProt ID of the human androgen receptor was found to be P10275. The full-length protein sequence was truncated, retaining the NTD domain (amino acid residues 1 to 555), and input into a drug target model for prediction and screening. The top 10,000 molecules in the model score were clustered into 50 categories. Subsequently, 27 compounds were manually selected, purchased, and tested.

[0099] Figure 5 shows the effects of 27 compounds on the activity of the androgen receptor NTD domain. The experiment was conducted in LNCaP cells, transfected with a construct fused with the AR-NTD and Gal4 DNA-binding domain (Gal4DBD-AR1-558) and a reporter gene located downstream of the Gal4 binding site. After cell treatment, the expression (fluorescence intensity) of the reporter gene was measured, indicating that nine compounds were active against the androgen receptor. These experimental results confirm that the invented sequence-based screening model can be applied to sequence-free proteins, overcoming the limitation of structure-based screening being restricted by protein structure, and holds promise for screening other sequence-free protein precursor molecules.

[0100] Example 2 illustrates the use of a model to predict the activity of Pim-1 inhibitors.

[0101] Pim-1 is a proto-oncogene encoding a serine / threonine kinase. First, we look up the UniProt ID of Pim-1 as P11309, retain the full-length protein sequence, and input it into our invented drug target model for prediction and screening. We retain the final attention score of the model and conduct a resolvable analysis and discussion by analyzing the crystal structure of the Pim-1 inhibitor receptor (PDB ID: 3R04).

[0102] As shown in Figure 6, the left figure is a schematic diagram of the two-dimensional interaction between the inhibitor UNQ molecule and the Pim-1 receptor structure. A total of 3 important hydrogen bond interactions are formed, which are formed between atom code N(0) and ASP 128, atom code O(18) and LYS 67, and atom code O(19) and ASP 186, respectively.

[0103] The right figure is a heatmap of attention scores predicted by the model of this invention. The horizontal axis represents the amino acid sequence number of the Pim-1 receptor, and the vertical axis represents the atom number of the UNQ molecule. The results show that the model assigns very high attention scores to the three positions where hydrogen bonds are formed: 0.125 for atom code N(0) and ASP 128; 0.161 for atom code O(18) and LYS 67; and 0.179 for atom code O(19) and ASP 186. These results demonstrate that the model can capture and reasonably interpret the important interactions between drug atoms and target amino acid residues, further confirming that the model has learned the important interactions between the drug and the target.

[0104] This invention provides a method for predicting drug-target interactions based on protein sequences. By constructing a drug target database, publicly available pre-trained protein language models and molecular models are used to extract features of targets and drugs, respectively. A prediction model is constructed by combining self-attention and cross-attention mechanisms, and the interpretability of the model is analyzed by outputting attention scores, detecting potential action sites and important amino acid residues. This method not only improves the accuracy of drug-target interaction prediction based on protein sequences but also provides important biological information for drug design and target validation, showing broad application prospects.

Claims

1. A method for predicting drug targets based on protein sequences, characterized in that, Includes the following steps: (1) Construct a drug-target interaction database and obtain relevant values ​​of target-drug pairs, including IC50. 50 EC 50 , K i and K d In addition, the UniProt sequence of the corresponding protein is obtained based on the target name, and the drug SMILES sequence is obtained based on the chemical information package. (2) Construct a target feature extractor based on a pre-trained protein language model to extract feature vectors of target protein sequences; the target feature extractor includes: calling a linear layer based on an open-source pre-trained protein language model to extract features from the input target UniProt sequence and output target features in units of amino acids; (3) Construct a drug feature extractor based on a pre-trained molecular characterization model to extract feature vectors of drug molecules; the drug feature extractor includes: calling a linear layer based on an open-source pre-trained molecular model to extract features from the input drug SMILES and output drug features in units of atoms; (4) Construct a feature learning method using a multi-head self-attention mechanism, construct a drug target interaction learning method using a cross-attention mechanism, and combine linear layer construction. (5) Using the drug target database constructed in step (1), train and optimize the parameters of the drug target interaction prediction model to obtain the optimized drug target prediction model; (6) Input the target protein UniProt sequence and drug molecule SMILES into the optimized drug target prediction model, output the prediction results, visualize the possible active sites of the drug and the target through attention scores, and perform interpretability analysis to mark the important amino acid residues of the target-drug interaction; In step (4), the drug target feature learning method is constructed using a multi-head attention mechanism and the drug target interaction learning method is constructed using a cross-attention mechanism, including: 1) Calling three different linear layers to map the obtained target sequence feature P to Three sets of vectors are used to map the acquired drug features D to three different linear layers. The attention scores are calculated using the following attention formula for the three sets of vectors: Where softmax is the normalization function, and the obtained... The attention score represents the self-attention of the target sequence. Among them, the obtained The attention score represents the self-attention of the drug. To further enhance the model's representational ability, a multi-head attention mechanism was adopted, consisting of h parallel heads: ； ;in, 、 It is the projection matrix for each head. It is the output projection matrix. 1) The score for multi-head attention; 2) Project the target sequence and drug features extracted by the multi-head attention mechanism onto the query, key, and value spaces respectively, and perform a linear transformation through a learnable weight matrix, as shown in the following formula: For the query, key, and value projection of the target sequence features: ； ； For querying drug characteristics, key-value projection: ； ； ;in, It is a query, key, and value weight matrix of target sequence features. It is a weighted matrix of keys and values ​​representing drug features. and It is the bias vector; 3) Further study the interaction between the drug and the target, and use the multi-head attention mechanism again. For each attention head h, perform the following calculations: ； Where h is the number of attention heads, each of which independently performs attention calculations, obtaining the attention outputs from drug to target and from target to drug respectively. The outputs of all attention heads are concatenated and projected through a linear transformation to obtain the final attention features. ； ;in, It is the output projection matrix of drug-to-target and target-to-drug attention, and the attention matrix obtained through the cross-attention mechanism. This is used to merge features, and then a linear layer is used to output the probability of drug-target interaction: ； ReLU(·) is an activation function applied to the linear transformation result of the hidden layer. and These are the weight matrix and bias vector of the hidden layer. and These are the weight matrix and bias vector of the output layer. It's the sigmoid function, which converts the output into probability values. This represents the predicted probability of drug-target interaction.

2. The method according to claim 1, characterized in that, In step (1), the acquisition of the drug target interaction database specifically includes: obtaining the UniProt sequence of the corresponding protein according to the target name and the drug SMILES sequence according to the chemical information package.

3. The method according to claim 1, characterized in that, In step (5), the drug target database is constructed in step (1), and the parameters of the drug target interaction prediction model are trained and optimized, including: 1) inputting the UniProt sequence of the target into an open-source protein language model to extract features and obtain sequence features in amino acid units; 2) inputting the SMILES of the drug into an open-source molecular pre-training model to extract features and obtain drug features in atom units; 3) constructing a feature learning method using a self-attention mechanism and a drug target interaction learning method using a cross-attention mechanism to update the features; 4) inputting the updated features into a predictor based on a linear layer to obtain the probability results of drug-target interaction; 5) calculating the loss value using a cross-loss entropy function, optimizing the parameters of the drug target prediction model based on minimizing the loss value, and using the early stopping method to obtain the optimal drug target prediction model, which is the optimized drug target prediction model.

4. The method according to claim 1, characterized in that, In step (6), the target protein sequence and drug molecule SMILES are input into the optimized drug target prediction model, and the prediction results are output. The possible interaction sites between the drug and the target are marked by the attention score visualization and interpretability analysis is performed. The important amino acid residues of the interaction between the target and the drug are marked, including: 1) The optimized drug target prediction model in step (6) is used to predict the actual drug target effect. During the model prediction process, the attention score of the interaction with the target amino acid as the query vector is retained. The attention score contains the attention information of each amino acid to each drug. The score is plotted into a heat map to detect the key amino acids of the drug interaction; 2) The interaction attention score with the drug atom as the query vector is saved. The attention score contains the interaction information of each drug atom and each amino acid. The score is plotted into a heat map to detect the important and key pharmacophores of the drug.

Images