Protein sequence-based drug target interaction prediction method

Abstract

The invention discloses a drug target interaction prediction method based on a protein sequence. The method comprises 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 adopting a self-attention mechanism, constructing a drug target interaction learning method by adopting a cross attention mechanism, constructing a predictor in combination with a linear layer, and constructing 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 a drug target into the optimized prediction model, outputting a prediction result, and finally marking potential active sites and important amino acid residues of the drug and the target through attention score visualization. According to the method, the precision of predicting the interaction between the drug and the target spot based on the protein sequence is improved, and the interpretability of the deep learning model is realized.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning and drug virtual screening, and relates to a method for predicting drug-target interaction, and in particular to a method for predicting drug-target interaction based on protein sequences. Background Art

[0002] Virtual screening of drugs is an important method for finding new skeleton lead molecules and lead compounds. Virtual screening based on protein structure requires the three-dimensional structure of the drug target, while virtual screening based on drug ligands requires known active molecules. Given the limitations of existing methods, protein sequence-based methods can overcome these limitations and have wider application value.

[0003] Artificial intelligence is widely used in drug-target interaction prediction, which is regarded as a binary classification task. Traditional machine learning methods make predictions by inputting predefined descriptors, while deep learning completes the prediction by automatically learning the characteristics of drugs and targets. Deep learning usually uses neural network structures, such as convolutional neural networks or graph neural networks, to capture complex features and relationships, thereby 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 representation of drugs and targets cannot effectively learn the interactions between them; 2) The model's generalization ability to unseen target systems is insufficient; 3) The prediction results of deep learning models lack interpretability. Summary of the invention

[0005] The purpose of the present invention is to provide a method for predicting drug-target interactions based on protein sequences. The method constructs a drug-target interaction database, extracts target and drug features using a pre-trained protein language model and a pre-trained molecular characterization model, and constructs a drug-target interaction prediction model using self-attention and cross-attention mechanisms. At the same time, combined with the visualization of attention scores, the interpretability analysis of the model is realized, the possible action sites of drugs and targets are annotated, and the biological interpretation ability of the model is improved.

[0006] The main technical solution of the present invention comprises the following steps: (1) Build a database of drug-target interactions and obtain relevant values ​​of target-drug pairs, including IC 50 ,EC 50 ,K i and K d , and obtain the corresponding protein UniProt sequence according to the target name and the drug SMILES sequence according to the drug ID; (2) construct a target feature extractor based on a pre-trained protein language model to extract feature vectors of target protein sequences; (3) Construct a drug feature extractor based on a pre-trained molecular representation model to extract feature vectors of drug molecules; (4) A multi-head self-attention mechanism is used to construct a feature learning method, a cross-attention mechanism is used to construct a drug-target interaction learning method, and a predictor is constructed in combination with a linear layer to build a drug-target interaction prediction model; 5) using the drug target database constructed in step 1), training and optimizing the parameters of the drug target interaction prediction model to obtain an optimized drug target prediction model; 6) Input the target protein UniProt sequence and drug molecule SMILES to be tested into the optimized drug target prediction model, output the prediction results, visualize the attention scores, mark the possible active sites of the drug and target, and perform interpretability analysis to mark the important amino acid residues for the interaction between the target and the drug.

[0007] In the step (1), the acquisition of the drug-target interaction database specifically includes: aligning the collected target data according to UniProt, and each target corresponds to a UniProt sequence. Aligning the collected drug data according to the chemical information package, and each drug corresponds to a SMILES sequence.

[0008] In the 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 units of amino acids.

[0009] In the step (3), the construction of a drug feature extractor based on a pre-trained molecular model includes: calling a linear layer according to an open source pre-trained molecular model to extract features from an input drug SMILES and outputting drug features in atomic units.

[0010] In the step (4), a drug target feature learning method is constructed by using a multi-head attention mechanism and a drug target interaction learning method is constructed by using a cross attention mechanism, specifically including:

[0011] 1) Call three different linear layers to map the acquired target sequence features P to Q P ,K P ,V P Three sets of vectors, calling three different linear layers to map the acquired drug features D to Q D ,K D ,V DThree sets of vectors are used to calculate the attention scores according to the following attention formula: Among them, softmax is the normalization function, and the obtained Attention(Q P ,K P ,V P ) is the self-attention score of the target sequence; Among them, the obtained Attention(Q D ,K D ,V D ) is the attention score of the drug’s self-attention. In order to further enhance the representation ability of the model, a multi-head attention mechanism is adopted, which consists of h parallel heads: MultiHead(Q,K,V)=Concat[head1,head2,…,head h ]W O in, is the projection matrix of each head, W O is the output projection matrix, and MultiHead is the score of multi-head attention.

[0012] 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 linearly transformed through a learnable weight matrix. The specific formula is as follows: For query, key, and value projection of target sequence features: For drug feature query, key, and value projection: in, is the weight matrix of the query, key and value of the target sequence feature, is the weight matrix of query, key and value of drug features, and is the bias vector.

[0013] 3) To further learn the interaction between drugs and targets, the multi-head attention mechanism is used again. For each attention head h, the following calculations are performed: Each attention head h performs attention calculation independently, and obtains the attention output from drug to target and target to drug respectively. The outputs of all attention heads are concatenated and projected through linear transformation to obtain the final attention feature: in, is the output projection matrix of drug-to-target and target-to-drug attention. The attention matrix A obtained by the cross-attention mechanism is used to merge features, and then the linear layer is used to output the probability of drug-target interaction: P = σ(ReLU(AW1+b1)W2+b2) Where ReLU(·) is the activation function, which is the result of the linear transformation applied to the hidden layer. W1 and b1 are the weight matrix and bias vector of the hidden layer. W2 and b2 are the weight matrix and bias vector of the output layer. σ is the sigmoid function, which converts the output into a probability value P, which represents the predicted probability of the interaction between the drug and the target.

[0014] In the step (5), the drug target database is constructed by using the step (1), and the parameters of the drug target interaction prediction model are trained and optimized, which specifically includes: 1) Input the UniProt sequence of the target into the open-source protein language model to extract features and obtain sequence features in amino acids; 2) Input the SMILES of the drug into the open-source molecular pre-training model to extract features and obtain the drug features in atomic units; 3) Use the self-attention mechanism to construct a feature learning method, and use the cross-attention mechanism to construct a drug-target interaction learning method to update the features; 4) Input the updated features into the predictor built based on the linear layer to obtain the probability results of the interaction between the drug and the target; 5) The cross-loss entropy function is used to calculate the loss value, and the parameters of the drug target prediction model are optimized according to the minimization of the loss value. 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.

[0015] In the step 6), the target protein sequence and drug molecule SMILES to be tested are input into the optimized drug target prediction model, and the prediction results are output. The possible action sites of the drug and the target are marked through attention score visualization, and interpretability analysis is performed to mark the important amino acid residues for the interaction between the target and the drug, specifically including: 1) The drug target prediction model optimized in step (6) is used to predict the actual drug target effect. In the process of model prediction, 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 on each drug. The score is plotted into a heat map to detect the key amino acids of drug interaction; 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. The scores are plotted as heat maps to detect the important and key pharmacophores of the drugs.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) Effectively learn the interaction between drugs and targets: Through the feature update of self-attention and the feature fusion method of cross-attention, the interaction between drugs and targets is effectively learned, and the accuracy of drug-target interaction prediction is improved; (2) Improve generalization ability: By introducing pre-trained protein language models and pre-trained molecular models, the generalization ability on unseen targets and molecules is improved; (3) Achieve model interpretability: Use the attention mechanism and combine it with the visualization of attention scores to clearly mark the key action sites of drugs and targets, provide biologically relevant information, and enhance the application value of the model in actual drug development; (4) Wide range of applications: This method does not depend on the three-dimensional structure of the target protein and is applicable to situations where the target structure is unknown, thus expanding the application scope of drug-target interaction prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 : Flowchart of a method for predicting drug-target interactions based on protein sequences.

[0018] Figure 2 : Schematic diagram of feature extractor based on pre-trained protein language model and pre-trained molecular model.

[0019] Figure 3 : Schematic diagram of the drug-target interaction prediction model structure based on self-attention and cross-attention mechanisms.

[0020] Figure 4 : The result of the enrichment index of DEKOIS2.0 data set was 0.5% (HitScreen is the method of the present invention).

[0021] Figure 5 :Results of virtual screening test of disordered domain NTD sequences in androgen receptor (AR).

[0022] Figure 6: Schematic diagram of the interaction between Pim-1 receptor structure and inhibitor UNQ molecule and attention score heat map. DETAILED DESCRIPTION

[0023] The present invention is further described in detail below in conjunction with 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.

[0024] Example 1 See also Figure 1 , the method of the present invention comprises the following steps:

[0025] Step (1) Construction of drug-target interaction database: First, we collected drug-target interaction data from the public databases BindingDB, ChEMBL, and DrugBank, including IC 50 ,EC 50 ,K i and K d Convert activity tags to pChEMBL values ​​(IC 50 ,EC 50 ,K i and K d The negative logarithm of pChEMBL values ​​(activity ≤ 100 nM) less than 7 was marked as positive, and the pChEMBL values ​​greater than 5 (activity ≥ 10,000 nM) were marked as negative, ensuring a 100-fold difference in activity between positive and negative molecules. The corresponding protein sequence was obtained according to the UniProt of the target, and the corresponding SMILES was obtained according to the ID of the drug. The SMILES was standardized using the chemical informatics tool RDKit, including removing unnecessary hydrogen atoms and adjusting isomer representation. The filtering criteria include: deleting molecules with more than 70 heavy atoms to avoid excessive computational complexity; deleting protein sequences with a length of more than 2000 amino acid residues to ensure the efficiency and effectiveness of the model in processing protein sequences. Finally, the collected data was organized into a unified format to construct a drug-target interaction database.

[0026] Step (2) construct a target feature extractor based on the pre-trained protein language model: like Figure 2 As shown in the figure, 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, and after passing through the multi-layer Transformer encoder, the result of the last layer is extracted, and the feature vector of each amino acid is output. Subsequently, the linear layer is called to perform linear transformation on these feature vectors to obtain the final target feature representation:

[0027] Among them, N P Indicates the number of amino acids in the protein sequence. ESM-2 selects the ESM2 t30 150M model with an embedding vector dimension of 1280.

[0028] Step (3) Construct a drug feature extractor based on the pre-trained molecular characterization model: like Figure 2 As shown in the figure, 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, and after passing through the multi-layer Transformer encoder, the result of the last layer is extracted, and the feature vector of each atom is output. The linear layer is called to perform linear transformation on these feature vectors to obtain the feature representation of the drug:

[0029] Among them, N a Represents the number of atoms in the drug molecule, and the dimension of the embedding vector is 512.

[0030] Step (4) uses multi-head self-attention and cross-attention mechanisms to construct drug-target interaction feature aggregators and predictors: like Figure 3 As shown in Figure 1, the target sequence feature P and drug feature D are first input into the multi-head self-attention mechanism to generate query (Q), key (K), and value (V) vectors respectively. Specifically, three different linear layers are called to map the target feature P to Q. P、 K P、 V P, Drug feature D is mapped to Q D ,K D ,V D ,

[0031]

[0032]

[0033] In order to enhance the feature representation capability, a multi-head attention mechanism is adopted, with 8 heads in total:

[0034] MultiHead(Q,K,V)=Concat[head1,head2,…,head8]W O in, is the projection matrix of each head, W Ois the output projection matrix, and MultiHead is the score of multi-head attention. The target sequence and drug features extracted by the multi-head attention mechanism are regenerated into query (Q), key (K), and value (V) vectors respectively. A linear transformation is performed through a learnable weight matrix, where the dimension of the hidden layer is 128. The specific formula is as follows:

[0035] For query, key, and value projection of target sequence features:

[0036] For drug feature query, key, and value projection: in, is the weight matrix of the query, key and value of the target feature, is the weight matrix of query, key and value of drug features, and is the bias vector.

[0037] In order to capture diverse interactions, a multi-head attention mechanism is used. For each attention head h, the following calculations are performed: Where h is the number of attention heads, which is 8. Each attention head h performs attention calculation independently to obtain the attention output from drug to target and target to drug.

[0038] The outputs of all attention heads are concatenated and projected through linear transformation to obtain the final attention features: in is the output projection matrix of drug-to-target and target-to-drug attention.

[0039] Step 4.5: Model prediction results like Figure 3 As shown, the attention matrix A obtained by the cross-attention mechanism is used to merge features, and then the linear layer is used to output the probability of drug-target interaction:

[0040] P = σ(ReLU(AW1+b1)W2+b2) Where ReLU(·) is the activation function, which is applied to the linear transformation result of the hidden layer. W1 and b1 are the weight matrix and bias vector of the hidden layer. W2 and b2 are the weight matrix and bias vector of the output layer. σ is the sigmoid function, which converts the output into a probability value P, which represents the predicted probability of drug-target interaction.

[0041] Step (5) training and optimizing the parameters of the drug-target interaction prediction model: The drug-target interaction database constructed in step 1 was used as training data, and the training set and validation set were randomly divided in an 8:2 ratio to train the prediction model. The cross entropy loss function was used to calculate the loss value during training. The Adam optimization algorithm was used, and 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 rounds was 40. The early stopping method was used. When the loss on the validation set no longer decreased within 10 consecutive epochs, the training was stopped to obtain the drug target prediction model with the optimal parameters.

[0042] Step (6) Input the target protein UniProt sequence and drug molecule SMILES to be tested, output the prediction results and perform interpretability analysis During the model prediction process, the attention scores with the target amino acid as the query vector are recorded. These attention scores reflect the degree of attention of each amino acid to the drug molecule. These scores are plotted as a heat map to intuitively display the possible key interaction sites between the drug and the target. Analyze the attention scores to identify the amino acid residues that have a greater impact on drug interactions. These key residues can serve as important reference points for drug design and optimization, providing a basis for further experimental verification.

[0043] like Figure 4 As shown, the performance was evaluated using the public dataset DEKOIS2.0. DEKOIS2.0 has a total of 81 targets, including kinases, proteases, nuclear receptors, G protein-coupled receptors, oxidoreductases, transferases, hydrolases, etc. Each target has an average of 40 active molecules and 1,200 inactive molecules. Taking the enrichment of 0.5 per thousand as the indicator, the performance of the model HitScreen we developed exceeds the existing benchmark models.

[0044] Example 1: Virtual screening using the disordered domain NTD sequence in the androgen receptor (AR). First, the UniProt ID of human androgen receptor was queried as P10275, and the full-length protein sequence was truncated to retain the sequence of the NTD domain (amino acid residue numbers 1 to 555), and then input into the drug target model for prediction and screening. The top 10,000 molecules ranked by the model were retained for cluster analysis and clustered into 50 categories, which were then manually selected, 27 compounds were purchased, and tested.

[0045] Figure 5To investigate the effects of 27 compounds on the activity of the androgen receptor NTD domain, the experiment involved transfection of a construct fused with the AR-NTD and the Gal4 DNA binding domain (Gal4DBD-AR1-558) and a reporter gene located downstream of the Gal4 binding site in LNCaP cells. After treating the cells, the expression of the reporter gene (fluorescence intensity) was measured, indicating that a total of 9 compounds were active against the androgen receptor. The above experimental results confirm that the invented sequence-based screening model can be applied to unsequenced proteins, overcoming the shortcomings of structure-based virtual screening that is limited by protein structure, and is expected to be applied to the screening of other disordered protein lead molecules.

[0046] Example 2: Use the model to predict the activity of Pim-1 inhibitors. Pim-1 is a serine / threonine kinase encoded by a proto-oncogene. First, the UniProt ID of Pim-1 is queried as P11309, and the full-length protein sequence is retained. It is input into the drug target model invented by us for prediction screening, and the final attention score of the model is retained. The crystal structure of the Pim-1 inhibitor receptor (PDB ID: 3R04) is analyzed for a resolvable analysis and discussion.

[0047] like Figure 6 As shown, the left figure is a schematic diagram of the two-dimensional interaction between the inhibitor UNQ molecule and the Pim-1 receptor structure, forming three important hydrogen bond interactions, namely, hydrogen bonds between atom code N(0) and ASP 128, atom number O(18) and LYS 67, and atom number O(19) and ASP 186.

[0048] The right figure is a heat map of the attention scores predicted by the model of the present invention, with the horizontal axis being the Pim-1 receptor amino acid sequence number and the vertical axis being the UNQ molecule atom number. The results show that the model gives very high attention scores to the three positions forming hydrogen bond interactions, with the atom code N(0) and ASP 128 having an attention score of 0.125; the atom number O(18) and LYS 67 having an attention score of 0.161; and the atom number O(19) and ASP 186 having an attention score of 0.179. The above results prove that the model can capture the important interactions between drug atoms and target amino acid residues and provide reasonable explanations, which further confirms that the model has learned the important interactions between drugs and targets.

[0049] The present invention provides a method for predicting drug-target interactions based on protein sequences. By constructing a drug target database, using a public pre-trained protein language model and a molecular model to extract the characteristics of targets and drugs respectively, combining self-attention and cross-attention mechanisms to build a prediction model, and outputting attention scores to achieve interpretability analysis of the model, detect possible 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 verification, and has broad application prospects.

Claims

1. A method for predicting drug targets based on protein sequences, characterized in that: The following steps are involved: (1) Build a database of drug-target interactions and obtain relevant values ​​of target-drug pairs, including IC 50 ,EC 50 ,K i and K d , and obtain the corresponding protein UniProt sequence according to the target name and the drug SMILES sequence according to 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; (3) Construct a drug feature extractor based on the pre-trained molecular representation model to extract feature vectors of drug molecules; (4) Use multi-head self-attention mechanism to construct feature learning method, use cross-attention mechanism to construct drug-target interaction learning method, combine linear layer to build predictor, and build a drug-target interaction prediction model; (5) using the drug target database constructed in step (1), training and optimizing the parameters of the drug target interaction prediction model to obtain an optimized drug target prediction model; (6) The UniProt sequence of the target protein and the SMILES of the drug molecule to be tested are input into the optimized drug target prediction model, and the prediction results are output. The possible active sites of the drug and the target are marked through attention score visualization, and interpretability analysis is performed to mark the important amino acid residues for the interaction between the target and the drug.

2. The method according to claim 1, characterized in that In step (1), the acquisition of the drug-target interaction database specifically includes: acquiring 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 (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 units of amino acids.

4. The method according to claim 1, characterized in that: In step (3), the construction of a drug feature extractor based on a pre-trained molecular model includes: calling a linear layer according to an open source pre-trained molecular model to extract features from an input drug SMILES and outputting drug features in atomic units.

5. The method according to claim 1, characterized in that In step (4), a drug target feature learning method is constructed by using a multi-head attention mechanism and a drug target interaction learning method is constructed by using a cross-attention mechanism, including: 1) Call three different linear layers to map the acquired target sequence features P to Q P ,K P ,V P Three sets of vectors, calling three different linear layers to map the acquired drug features D to Q D ,K D ,V D Three sets of vectors are used to calculate the attention scores according to the following attention formula: Among them, softmax is the normalization function, and the obtained Attention(Q P ,K P ,V P ) is the self-attention score of the target sequence; Among them, the obtained Attention(Q D ,K D ,V D ) is the attention score of the drug’s self-attention. In order to further enhance the representation ability of the model, a multi-head attention mechanism is adopted, which consists of h parallel heads: MultiHead(Q,K,V)=Concat[head1,head2,…,head h ]W O in, is the projection matrix of each head, W O is the output projection matrix, MultiHead is the score of multi-head attention; 2) The target sequence and drug features extracted by the multi-head attention mechanism are projected into the query, key, and value spaces respectively, and linearly transformed through a learnable weight matrix. The specific formula is as follows: For query, key, and value projection of target sequence features: For drug feature query, key, and value projection: in, is the weight matrix of the query, key and value of the target sequence feature, is the weight matrix of query, key and value of drug features, and is the bias vector; 3) To further learn the interaction between drugs and targets, the multi-head attention mechanism is used again. For each attention head h, the following calculations are performed: Among them, h is the number of attention heads, each attention head performs attention calculation independently, and obtains the attention output from drug to target and target to drug respectively. The outputs of all attention heads are spliced ​​and projected through linear transformation to obtain the final attention feature: in, is the output projection matrix of drug-to-target and target-to-drug attention. The attention matrix A obtained by the cross-attention mechanism is used to merge features, and then the linear layer is used to output the probability of drug-target interaction: P = σ(ReLU(AW1+b1)W2+b2) Where ReLU(·) is the activation function, which is the result of the linear transformation applied to the hidden layer, W1 and b1 are the weight matrix and bias vector of the hidden layer, W2 and b2 are the weight matrix and bias vector of the output layer, and σ is the sigmoid function, which converts the output into a probability value P, which represents the predicted probability of drug-target interaction.

6. The method according to claim 1, characterized in that 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, including: 1) Input the UniProt sequence of the target into the open-source protein language model to extract features and obtain sequence features in amino acids; 2) Input the SMILES of the drug into the open-source molecular pre-training model to extract features and obtain the drug features in atomic units; 3) Use the self-attention mechanism to construct a feature learning method, and use the cross-attention mechanism to construct a drug-target interaction learning method to update the features; 4) Input the updated features into the predictor built based on the linear layer to obtain the probability results of the interaction between the drug and the target; 5) The cross-loss entropy function is used to calculate the loss value, and the parameters of the drug target prediction model are optimized according to the minimization of the loss value. 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.

7. The method according to claim 1, characterized in that In step (6), the target protein sequence and drug molecule SMILES to be tested are input into the optimized drug target prediction model, and the prediction results are output. The possible action sites of the drug and the target are marked through attention score visualization, and interpretability analysis is performed to mark the important amino acid residues for the interaction between the target and the drug, including: 1) The drug target prediction model optimized in step (6) is used to predict the actual drug target effect. In the process of model prediction, 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 on each drug. The score is plotted into a heat map to detect the key amino acids of drug interaction; 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. The scores are plotted as heat maps to detect the important and key pharmacophores of the drugs.

8. The method of claim 1 is used in drug target prediction.

Images