A multi-target drug design method based on protein sequence and similarity features

Abstract

This invention discloses a multi-target drug design method based on protein sequence and similarity features. A multi-target drug design model is designed, consisting of a target feature embedding module, a target feature encoding module, and a multi-target drug generation module. By independently encoding and similarity encoding the features of multiple targets, and then using the encoded feature latent vectors and similarity feature latent vectors, drug molecules targeting multiple protein targets can be directly generated. Specifically, the target feature embedding module is pre-trained on a large protein dataset based on ProtTrans, demonstrating strong protein feature embedding capabilities; the multi-target drug generation module is trained on a large drug-like dataset, achieving strong drug generation capabilities; subsequently, the entire model is trained on a multi-target drug dataset, achieving excellent multi-target drug design capabilities. The designed drug molecules can bind well to each target protein, achieving the goal of multi-target drug therapy.

Description

Technical Field

[0001] This invention belongs to the field of computer-aided drug design, specifically relating to a multi-target drug design method based on protein sequence and similarity features. Background Technology

[0002] In recent years, deep learning technology has made significant progress in drug discovery tasks, including drug screening, de novo drug design, and targeted drug design. Drug screening aims to screen drugs from a database of known compounds, de novo drug design aims to directly design new drugs that do not exist in a compound database, and targeted drug design directly designs new targeted drugs that do not exist in a compound database based on the characteristics of the target.

[0003] Targeted drugs achieve therapeutic effects by binding to target proteins (also known as targets) and inhibiting their activity. Currently, in computer-aided drug design research, new targeted drugs can be directly designed based on single target characteristics, such as the sequence or pocket features of the target protein. However, most diseases are caused by multiple targets; for example, pneumonia caused by the SARS-CoV-2 virus includes targets such as 3CL protease (3Clpro, a 3C-like protease) and papain-like protease (PLpro). For multi-target drug design research, existing methods design multi-target drugs based on a drug targeting one of the targets; currently, there is no method to directly design multi-target drugs based on the protein sequence characteristics of multiple targets. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide a multi-target drug design method based on protein sequence and similarity features. This method can directly design targeted drugs adapted to multiple targets based on the sequence features of multiple target proteins.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] The first aspect of this invention provides a multi-target drug design method based on protein sequence and similarity features, comprising the following steps:

[0007] S1. Raw Dataset Collection: Collect non-targeted compound datasets and single-target drug datasets from the database;

[0008] S2. Data preprocessing: The untargeted compound dataset obtained in step S1 is preprocessed to obtain a drug-like dataset, and the single-target drug dataset is preprocessed to obtain a multi-target drug dataset.

[0009] S3. Construct a multi-target drug design model, including a target feature embedding module for obtaining target feature vectors, a target feature encoding module for obtaining target feature latent vectors and target similarity feature latent vectors, and a multi-target drug generation module for generating drug molecules. The above three modules are sequentially connected to form the multi-target drug design model.

[0010] S4. Training the multi-target drug design model: Train the multi-target drug generation module using the drug dataset described in step S2, so that it learns the structural features and generation rules of drug molecules; then, fix the parameters of the target feature embedding module, and train the entire multi-target drug design model using the multi-target drug dataset, so that it learns the generation rules of multi-target drugs.

[0011] S5. Input at least two target protein sequences into the multi-target drug design model trained in step S4 to generate multi-target drug molecules.

[0012] Preferably, the untargeted compound dataset comes from the ZINC database, and the single-target drug dataset comes from the ChEMBL database and the BindingDB database.

[0013] Preferably, the data preprocessing method is as follows:

[0014] 1) Drug-like dataset selection: Data with an oil-water partition coefficient of no more than 3 and a relative molecular mass of no more than 300 were selected from the non-targeted compound dataset to form a drug-like dataset;

[0015] 2) Screening of multi-target drug datasets:

[0016] a. The single-target drug dataset is stored in the form of (drug sequence, target protein sequence) tuples. Targeted drugs with two or more tuples are multi-target drugs. The drug sequences are represented in the form of SMILES sequences.

[0017] b. Delete binary data where the length of the drug's SMILES sequence is not in [34, 74] and / or the target protein sequence exceeds 1000. Finally, a multi-target drug dataset consisting of multi-target drugs targeting two targets is obtained. Each multi-target drug is stored in the form of a triple (drug sequence, target protein sequence 1, target protein sequence 2).

[0018] Preferably, the target feature embedding module is constructed based on the ProtTrans model, which obtains the feature vector of the target based on the target protein sequence, including the following steps:

[0019] 1) Protein sequence tokenization: Tokenize the amino acid residue sequence of a target protein, which is represented by ASCII characters, into a digital token that can be used for neural network computation;

[0020] 2) Load the pre-trained model: Initialize the target feature embedding module using the parameters of ProtT5-XL version of ProtTrans;

[0021] 3) Input the tokenized target protein sequence into the target feature embedding module to obtain a feature vector containing multiple target features.

[0022] Preferably, the target feature encoding module is constructed based on the Transformer framework model. It receives the feature vectors of each target obtained by the target feature embedding module, and independently encodes the protein features of the target to obtain the latent feature vectors of each target. Simultaneously, it performs protein similarity feature encoding on all targets to obtain the similarity latent feature vectors of all targets, including the following steps:

[0023] 1) Create a Transformer. The formula for the Transformer is as follows:

[0024] Linear {Q，K,V} (x)=xW {Q，K,V} (1)

[0025]

[0026] FFN(x)=max(0,xW1+b1)W2+b2 (4)

[0027] In formula (1):

[0028] Linear {Q,K,V} Represents three linear layers;

[0029] Q, K, and V represent the three input vectors in the Transformer;

[0030] x represents the input target feature vector;

[0031] W {Q,K,V} Represents Linear {Q,K,V} Weight parameters;

[0032] In formula (2):

[0033] Att i Represents the attention mechanism of the i-th head;

[0034] SoftMax represents the activation function;

[0035] d k It is the dimension of the input vector K;

[0036] In formula (3):

[0037] MSA stands for Multi-head Attention Mechanism;

[0038] Concat represents concatenating functions;

[0039] W O Represents the weight parameters of the output layer;

[0040] In equation (4):

[0041] FFN stands for Forward Propagation Network;

[0042] `max` represents the maximum value.

[0043] W1 and W2 represent the weight parameters of FFN;

[0044] b1 and b2 represent the deviation parameters of FFN;

[0045] 2) Target similarity features are calculated based on the cross-attention mechanism, and latent vectors of target features and target similarity features are obtained by encoding the feature vectors of target proteins. The formula for calculating the latent vectors is as follows:

[0046]

[0047] O MLP =W3(W2(W1×FF+b1+b2)+b3 (8)

[0048] TSL = Transformer ts (O MLP (9)

[0049] TFL1 = Transformer t1 (TF1) (10)

[0050] TFL2 = Transformer t2 (TF2) (11)

[0051] In equations (5) and (6):

[0052] TF1 and TF2 represent the feature vectors of target point 1 and target point 2, respectively;

[0053] FF m1 and FF m2 These represent the feature fusion matrices for the two target points respectively;

[0054] d is the dimension of the feature vector;

[0055] In equation (7):

[0056] FF represents the feature fusion vector of the two target points;

[0057] Concat represents matrix concatenation;

[0058] In equation (8):

[0059] O MLP The result represents the computation of the multilayer perceptron;

[0060] W1, W2, and W3 represent the weight parameters of the multilayer perceptron;

[0061] b1, b2, and b3 represent the bias parameters of the multilayer perceptron;

[0062] In equation (9):

[0063] TSL represents the target similarity latent vector;

[0064] Transformer ts Transformer computation representing latent vectors of target similarity;

[0065] In formula (10):

[0066] TFL1 represents the latent vector of target point 1;

[0067] Transformer t1 Transformer calculation of the latent vector of target point 1;

[0068] In equation (11):

[0069] TFL2 represents the latent vector of target point 2;

[0070] Transformer t2 Transformer calculation of the latent vector of target point 2.

[0071] Preferably, the multi-target drug generation module is constructed based on the GRU framework model. It receives target feature latent vectors and target similarity feature latent vectors from the target feature encoding module and generates multi-target drug sequences, including the following steps:

[0072] 1) The calculation formula for drug sequence generation is as follows:

[0073]

[0074] In equation (12):

[0075] r represents the value of the reset gate in the GRU;

[0076] l represents the layer number in GRU;

[0077] t represents the time in the GRU;

[0078] σ represents the sigmoid function;

[0079] W represents the weight;

[0080] h represents the hidden state;

[0081] p represents the SMILES character;

[0082] In equation (13):

[0083] z represents the value of the update gate in GRU;

[0084] In equation (14):

[0085] Represents a candidate hidden state in GRU;

[0086] tanh represents the tanh function;

[0087] In equation (15):

[0088] b represents deviation;

[0089] In equation (16):

[0090] P i This represents the SMILES character output at time i;

[0091] 2) The target feature latent vector and the target similarity feature latent vector are used as hidden states h and input to the multi-target drug generation module. Under the control of the GRU reset gate and update gate, the module continuously decodes and generates SMILES characters one by one until the SMILES form a complete drug molecule.

[0092] Preferably, the training of the multi-target drug generation module includes calculating the deviation between the SMILES sequence of the real drug-like molecule and the SMILES sequence of the generated drug molecule through cross-entropy loss, backpropagating the deviation to the multi-target drug generation module to obtain the parameter gradient, and updating the parameters of the multi-target drug generation module through the Adam algorithm.

[0093] The training of the multi-target drug design model includes calculating the deviation between the SMILES sequence of the real multi-target drug and the SMILES sequence of the multi-target drug generated by the model through cross-entropy loss, backpropagating the deviation to the multi-target drug design model to obtain the parameter gradient, and updating the parameters of the multi-target drug design model through the Adam algorithm.

[0094] Preferably, the multi-target drug design method further includes step S6, in which the designed multi-target drug is docked with its corresponding target protein to detect the binding affinity of the drug molecule to each target.

[0095] A second aspect of the present invention provides a multi-target drug design system based on protein sequence and similarity features, comprising at least:

[0096] The target feature embedding module based on ProtTrans is used to obtain the feature vector of the target based on the target protein sequence.

[0097] The Transformer-based target feature encoding module is used to receive the feature vectors of each target obtained by the target feature embedding module, and independently encode the protein features of each target to obtain the latent feature vector of each target; at the same time, it encodes the protein similarity features of all targets to obtain the latent feature vectors of all targets.

[0098] A GRU-based multi-target drug generation module is used to receive target feature latent vectors and target similarity feature latent vectors from the target feature encoding module to generate multi-target drug molecules.

[0099] The target feature embedding module, the target feature encoding module, and the multi-target drug generation module are sequentially arranged to form the multi-target drug design system.

[0100] A third aspect of the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of a multi-target drug design method based on protein sequence and similarity features as described in any one of claims 1-8.

[0101] Compared with the prior art, the present invention has the following beneficial effects:

[0102] (1) This invention proposes a multi-target drug design method based on protein sequence and similarity features. A multi-target drug design model is designed, consisting of a target feature embedding module, a target feature encoding module and a multi-target drug generation module. By independently encoding and similarity encoding the features of multiple targets, and then based on the encoded feature latent vectors and similarity feature latent vectors, drug molecules targeting multiple protein targets can be directly generated. This method is the first of its kind in the field.

[0103] (2) The target feature embedding module of the present invention is constructed based on the ProtTrans model. The ProtTrans model uses large-scale protein sequence data for pre-training, thereby enabling the target feature embedding module of the present invention to have strong protein feature embedding ability and improve the accuracy of multi-target drug design model.

[0104] (3) The model of this invention underwent two training sessions. The first training session used a drug-like dataset to train the multi-target drug generation module, enabling it to learn the structural features and generation rules of drug molecules. The second training session used a multi-target drug dataset composed of triplets, with the parameters of the target feature embedding module fixed, to train the entire multi-target drug design model, enabling it to learn the generation rules of multi-target drug molecules and achieving good multi-target drug design capabilities. These two training sessions further improved the accuracy of the multi-target drug design model of this invention. The designed drug molecules were demonstrated through molecular docking to bind well to the corresponding protein targets, achieving the purpose of multi-target drug therapy and providing a practical new method for computer-aided drug design. Attached Figure Description

[0105] Figure 1 This is an overall framework diagram of a multi-target drug design model according to an embodiment of the present invention;

[0106] Figure 2 This is a basic structural diagram of a target feature encoding module according to an embodiment of the present invention;

[0107] Figure 3 This is a diagram illustrating the design process of a multi-target drug according to an embodiment of the present invention;

[0108] Figure 4 for Figure 3 Conformation diagram of multi-target drug docking with 3CLpro target;

[0109] Figure 5 for Figure 4 Diagram of docking sites between multi-target drugs and the 3CLpro target;

[0110] Figure 6 for Figure 5 3D molecular structure diagram of a multi-target drug docking with the 3CLpro target;

[0111] Figure 7 for Figure 3 Conformation diagram of multi-target drugs docking with PLpro target;

[0112] Figure 8 for Figure 7 Diagram of docking sites between multi-target drugs and PLpro target;

[0113] Figure 9 for Figure 8 3D molecular structure diagram of a multi-target drug docking with the PLpro target. Detailed Implementation

[0114] To make the objectives, technical solutions, and advantages of this invention clearer, the multi-target drug design method based on protein sequence and similarity features of this invention will be specifically described below with reference to the accompanying drawings and embodiments. It should be understood that the following embodiments are merely specific examples of this invention and are not intended to limit the scope of protection of this invention.

[0115] 1. Terminology Explanation:

[0116] SMILES: A specification for explicitly describing molecular structures using ASCII strings;

[0117] ProtTrans: Represents a protein language model in deep learning;

[0118] Transformer: Represents a framework model in deep learning based on a multi-head self-attention mechanism;

[0119] Feature latent vector: represents the vectorized representation of the latent features used in downstream tasks in deep learning; the feature latent vector of a target protein sequence represents the vectorized representation of the latent features of the target protein sequence when it is used in targeted drug design, and different bits in the vector represent different latent features in the target protein sequence.

[0120] GRU: Represents a framework model based on recurrent neural networks in deep learning;

[0121] Cross-entropy loss: represents a loss function in deep learning;

[0122] SoftMax: Represents an activation function in deep learning;

[0123] Adam: Represents a parameter optimization algorithm in deep learning;

[0124] Vina: Represents a molecular docking algorithm in virtual screening.

[0125] 2. Database Description:

[0126] ZINC Database: A free, commercial virtual screening database of compounds containing over 750 million purchasable compounds, provided and maintained by the Irwin and Shoichet labs in the Department of Medicinal Chemistry at the University of California, San Francisco (UCSF). Website: http: / / zinc15.docking.org / ;

[0127] The ChEMBL database is a free online database developed by the European Institute for Bioinformatics (EBI). It collects bioactivity data of various targets and compounds from a large number of literatures. The latest version contains nearly 2.35 million different compounds and more than 150,000 targets. Website: https: / / www.ebi.ac.uk / chembl / ;

[0128] BindingDB database: A publicly accessible database that primarily collects affinity data for interactions between drug target proteins and drug-like small molecules. URL: http: / / bindingdb.org / bind / index.jsp;

[0129] PDB database: An international database that stores three-dimensional structural data of biological macromolecules such as proteins and nucleic acids. Website: https: / / www.rcsb.org / .

[0130] This embodiment provides a multi-target drug design method based on protein sequence and similarity features. For ease of explanation, this embodiment uses the design of a drug targeting two targets as an example. It should be noted that the multi-target drug design method of this invention includes, but is not limited to, drug design targeting two targets. Furthermore, this embodiment specifically selects the 3CL protease (3CLpro) and papain-like protease (PLpro) targets of the SARS-CoV-2 virus in COVID-19 influenza to design a multi-target drug against the SARS-CoV-2 virus. Detailed steps are as follows:

[0131] S1. Raw Dataset Collection: Obtain non-targeted compound datasets and single-target drug datasets from the database.

[0132] In this embodiment, a script is used to collect a dataset of untargeted compounds from the ZINC database and a dataset of single-target drugs from the ChEMBL and BindingDB databases, respectively.

[0133] S2. Data Preprocessing: The untargeted compound dataset obtained in step S1 is preprocessed to obtain a drug-like dataset, which will be used as the training dataset for the subsequent multi-target drug generation module; the single-target drug dataset obtained in step S1 is preprocessed to obtain a multi-target drug dataset, which will be used as the training dataset for the subsequent multi-target drug design model. The specific method is as follows:

[0134] S201. Drug-like dataset screening: According to Lipinski's Rule of Five (Ro5), compounds with an oil-water partition coefficient (LogP) of no more than 3 and a molecular weight of no more than 300 have strong drug-like properties. Therefore, 87,719,678 data points with LogP values ​​of no more than 3 and relative molecular masses of no more than 300 were selected from the untargeted compound dataset in step S1 to form the drug-like dataset.

[0135] S202, Screening of multi-target drug datasets:

[0136] 1) Single-target drugs collected from the ChEMBL and BindingDB databases are stored in the form of (drug sequence, target protein sequence) tuples. If there are two or more tuples with the same drug sequence but different target protein sequences, the drug corresponding to this drug sequence is a multi-target drug. The proteins corresponding to these target protein sequences are the targets of this multi-target drug. The drug sequence is represented in the form of SMILES sequence.

[0137] 2) Since more than 95% of the proteins in the database have a sequence length of less than 1000, and more than 95% of the drug molecules have SMILES sequence lengths in the range of [34, 74], in order to reduce the variance of the data length in the training set and maintain model stability, data with target protein sequences longer than 1000 and drug molecules with SMILES sequence lengths not in the range of [34, 74] were discarded. Finally, a total of 52,896 triplet data points for multi-target drugs that can target two targets were selected. Each triplet is (drug sequence, target protein sequence 1, target protein sequence 2), which constitutes the multi-target drug dataset.

[0138] S203. Drug molecule data preprocessing: In order to enhance the boundary features of the SMILES sequences of drug molecules, better represent the start and end of the molecule, and make all SMILES sequences the same length, a 'G' character is added at the beginning of each SMILES sequence and an 'E' character is added at the end. After the 'E' character, a 'Q' character is added, so that the length of each SMILES sequence is 76.

[0139] S204. Protein sequence data preprocessing: All characters representing uncommon amino acid residues in the protein sequence are replaced with the character 'X'.

[0140] S3. Construct a multi-target drug design model based on target protein sequences and similarity features: See [link / reference] Figure 1In this embodiment, the protein sequence features of target 1 and target 2 are analyzed. A target feature embedding module based on ProtTrans, a target feature encoding module based on Transformer, and a multi-target drug generation module based on GRU are used to sequentially form a multi-target drug design model. The specific method is as follows:

[0141] S301. Construct a target feature embedding module based on ProtTrans to obtain the target feature vector based on the protein sequence of the target, including the following steps:

[0142] 1) Protein sequence tokenization: Tokenizing amino acid residue sequences represented by ASCII characters into digital tokens that can be used for neural network computation;

[0143] 2) Load the pre-trained model: Initialize the target feature embedding module using the parameters of ProtT5-XL version of ProtTrans;

[0144] 3) Input the tokenized protein sequences of target 1 and target 2 into the target feature embedding module to obtain feature vector 1 and feature vector 2 containing rich target features.

[0145] S302. Construct a Transformer-based target feature encoding module: The feature vector of each target point obtained in step S301 is passed as input to the Transformer-based target feature encoding module. The basic structure of the target feature encoding module is as follows: Figure 2 As shown, the Transformer model further processes and encodes the input feature vectors to extract deeper feature representations, i.e., obtains the latent feature vector for each target. While encoding individual protein features for each target, it also encodes protein similarity features for all targets, obtaining the latent feature vectors for the similarity of all targets. This includes the following steps:

[0146] 1) Create a Transformer. The formula for the Transformer is as follows:

[0147] Linear {Q，K,V} (x)=xW {Q，K,V}( 1)

[0148]

[0149] FFN(x)=max(0,xW1+b1)W2+b2 (4)

[0150] In formula (1):

[0151] Linear {Q , K,V}Represents three linear layers;

[0152] Q, K, and V represent the three input vectors in the Transformer;

[0153] x represents the input target feature vector;

[0154] W {Q,K,V} Represents Linear {Q , K,V} Weight parameters;

[0155] In formula (2):

[0156] Att i Represents the attention mechanism of the i-th head;

[0157] SoftMax represents the activation function;

[0158] d k It is the dimension of the input vector K;

[0159] In formula (3):

[0160] MSA stands for Multi-head Attention Mechanism;

[0161] Concat represents concatenating functions;

[0162] W O Represents the weight parameters of the output layer;

[0163] In equation (4):

[0164] FFN stands for Forward Propagation Network;

[0165] `max` represents the maximum value.

[0166] W1 and W2 represent the weight parameters of FFN;

[0167] b1 and b2 represent the deviation parameters of FFN;

[0168] 2) Target similarity features are calculated based on the cross-attention mechanism, and target feature latent vector 1, feature latent vector 2, and target similarity feature latent vector are obtained by encoding the feature vectors of all target proteins. The calculation formula for the latent vectors is as follows:

[0169]

[0170] O MLP =W3(W2(W1×FF+b1)+b2)+b3 (8)

[0171] TSL = Transformer ts (O MLP (9)

[0172] TFL1 = Transformer t1 (TF1) (10)

[0173] TFL2 = Transformer t2 (TF2) (11)

[0174] In equations (5) and (6):

[0175] TF1 and TF2 represent the feature vectors of target point 1 and target point 2, respectively;

[0176] FF m1 and FF m2 These represent the feature fusion matrices for the two target points respectively;

[0177] d is the dimension of the feature vector;

[0178] In equation (7):

[0179] FF represents the feature fusion vector of the two target points;

[0180] Concat represents matrix concatenation;

[0181] In equation (8):

[0182] O MLP The result represents the computation of the multilayer perceptron;

[0183] W1, W2, and W3 represent the weight parameters of the multilayer perceptron;

[0184] b1, b2, and b3 represent the bias parameters of the multilayer perceptron;

[0185] In equation (9):

[0186] TSL represents the target similarity latent vector;

[0187] Transformer ts Transformer computation representing latent vectors of target similarity;

[0188] In formula (10):

[0189] TFL1 represents the latent vector of target point 1;

[0190] Transformer t1 Transformer calculation of the latent vector of target point 1;

[0191] In equation (11):

[0192] TFL2 represents the latent vector of target point 2;

[0193] Transformer t2 Transformer calculation of the latent vector of target point 2;

[0194] S303. Construct a GRU-based multi-target drug generation module to output multi-target drug sequences based on the latent feature vectors of individual targets and the similarity latent feature vectors of all targets. This includes the following steps:

[0195] 1) The calculation formula for drug sequence generation is as follows:

[0196]

[0197]

[0198] In equation (12):

[0199] r represents the value of the reset gate in the GRU;

[0200] l represents the layer number in GRU;

[0201] t represents the time in the GRU;

[0202] σ represents the sigmoid function;

[0203] W represents the weight;

[0204] h represents the hidden state;

[0205] p represents the SMILES character;

[0206] In equation (13):

[0207] z represents the value of the update gate in GRU;

[0208] In equation (14):

[0209] Represents a candidate hidden state in GRU;

[0210] tanh represents the tanh function;

[0211] In equation (15):

[0212] b represents deviation;

[0213] In equation (16):

[0214] P i This represents the SMILES character output at time i;

[0215] 2) The target feature latent vector 1, feature latent vector 2 and target similarity feature latent vector obtained from the target feature encoding module in step S302 are used as hidden state h and input to the multi-target drug generation module. Under the control of the reset gate and update gate of GRU, the module continuously decodes and generates SMILES characters one by one until SMILES form a complete drug molecule sequence.

[0216] S4. Training the multi-target drug design model includes first training the multi-target drug generation module, and then training the entire multi-target drug design model. The specific method is as follows:

[0217] S401, Training of the multi-target drug generation module.

[0218] The multi-target drug generation module was trained using a drug-like dataset containing 87,719,678 data points, enabling it to learn the structural features of drug molecules and generation rules. This included the following steps:

[0219] 1) Input the start character 'G' into the multi-target drug generation module, so that the module can automatically generate a SMILES sequence including the end character 'E' and the padding character 'Q';

[0220] 3) Calculate the deviation between the SMILES sequence of the real drug molecule and the SMILES sequence of the generated drug molecule using cross-entropy loss;

[0221] 4) Backpropagate the bias between SMILES sequences to the multi-target drug generation module to obtain the parameter gradient, and update the parameters of the multi-target drug generation module using the Adam algorithm;

[0222] 5) The multi-target drug generation module is trained on a drug-like dataset for 10 rounds, and the training ends and the parameters of this module are saved.

[0223] S402, Training of multi-target drug design model.

[0224] The entire multi-target drug design model was trained using a multi-target drug dataset containing 52,896 triplet data points (drug sequence, target protein sequence 1, target protein sequence 2), including the following steps:

[0225] 1) Input the target protein sequence 1 and target protein sequence 2 from each triplet into the multi-target drug design model trained in step S401 to obtain the generated multi-target drug sequences. It should be noted that the target feature embedding module in the model is built on ProtTrans, which has been pre-trained on a publicly available large protein dataset based on ProtTrans and has strong protein feature embedding capabilities. Therefore, the parameters of this module remain fixed during the training process of this model.

[0226] 2) The deviation between the real SMILES sequence of multi-target drugs and the SMILES sequence of multi-target drugs generated by the model is calculated by cross-entropy loss.

[0227] 3) Backpropagate the bias between SMILES sequences to the multi-target drug design model to obtain parameter gradients, and update the parameters of the multi-target drug design model using the Adam algorithm;

[0228] 4) The multi-target drug design model is trained for 10 rounds on the multi-target drug dataset, and the training ends and the model parameters are saved.

[0229] S5. Generate multi-target drug molecules using the multi-target drug design model trained in step S4.

[0230] The target protein sequences 1 and 2 of the drug molecule to be designed are input into the pre-trained multi-target drug design model to generate SMILES characters one by one until the SMILES form a complete drug molecule sequence. The drug molecule generation process is as follows: Figure 3 As shown.

[0231] In this embodiment, target 1 is the 3CLpro target and target 2 is the PLpro target. The target protein data is preprocessed according to the aforementioned steps, including replacing the characters of uncommon amino acid residues in the sequence with the character 'X'. The protein sequences of the two targets are then input into the multi-target drug design model to obtain drug molecules targeting the two targets.

[0232] S6. Dock the designed multi-target drug with its corresponding target protein, and detect the binding affinity, binding site, and binding conformation of the drug and the target, including the following steps:

[0233] S601. Acquisition of target protein structure data: Obtain the structure data of each target protein from the PDB database;

[0234] S602. Preprocessing of target protein data: The protein structures in the PDB database contain ligands and water molecules. Ligands do not belong to proteins and occupy the pocket structure of proteins. Water molecules will interfere with the binding of targeted drugs to proteins. Therefore, ligands and water molecules in the protein structure are removed in advance.

[0235] S603 is used to examine the torsion center and torsion bond of multi-target drugs, so that multi-target drugs can form a molecular structure that adapts to the protein pocket structure during molecular docking.

[0236] S604. Add hydrogen ions to the target protein and calculate the charge number so that the target protein can form stable hydrogen bonds and other docking structures with multi-target drugs in molecular docking.

[0237] S605. Based on the Vina algorithm, multi-target drugs are docked with their corresponding target proteins to obtain binding affinity, binding sites, and binding conformations. Based on the docking results, multi-target drugs with strong binding affinity, tight binding sites, and stable binding conformations are obtained.

[0238] In this embodiment, Figure 4-9 The conformational diagrams, site diagrams, and 3D molecular structures during docking of the generated multi-target drug with 3CLpro and PLpro targets are shown. In molecular docking, the atomic contacts between the protein fragment and the ligand constitute the interaction pair, and the interatomic distance must be ≤ [missing information]. Figure 5 The results showed that the drug molecule had two interaction pairs with the 3CLpro target, with binding distances of 2.9 and 2.0, respectively. Figure 8 The results show that the drug molecule has four interaction pairs with the PLpro target, with binding distances of 2.6, 2.8, 3.2, and 2.7, respectively. This indicates that the distances between the generated multi-target drug molecule and both the 3CLpro and PLpro targets satisfy the condition that the atomic distance between individual interaction pairs is less than [amount missing]. The drug molecules meet the requirements, and the binding conformation between the drug molecules and each target protein is good and the binding state is stable, which shows that the multi-target drug design method of the present invention has good multi-target drug design effect and good application prospects.

Claims

1. A multi-target drug design method based on protein sequence and similarity features, characterized in that, Includes the following steps: S1. Raw Dataset Collection: Collect non-targeted compound datasets and single-target drug datasets from the database; S2. Data preprocessing: The untargeted compound dataset obtained in step S1 is preprocessed to obtain a drug-like dataset, and the single-target drug dataset is preprocessed to obtain a multi-target drug dataset. S3. Construct a multi-target drug design model, including a target feature embedding module for obtaining target feature vectors, a target feature encoding module for obtaining target feature latent vectors and target similarity feature latent vectors, and a multi-target drug generation module for generating drug molecules. The above three modules are sequentially connected to form the multi-target drug design model. S4. Training the multi-target drug design model: Train the multi-target drug generation module using the drug dataset described in step S2, so that it learns the structural features and generation rules of drug molecules; then, load the pre-trained model of the target feature embedding module to obtain and fix the parameters of the module, and train the entire multi-target drug design model using the multi-target drug dataset, so that it learns the generation rules of multi-target drugs. S5. Input at least two target protein sequences into the multi-target drug design model trained in step S4 to generate multi-target drug molecules.

2. The multi-target drug design method according to claim 1, characterized in that, The dataset of untargeted compounds comes from the ZINC database, and the dataset of single-target drugs comes from the ChEMBL and BindingDB databases.

3. The multi-target drug design method according to claim 1, characterized in that, The data preprocessing method is as follows: 1) Drug-like dataset selection: Data with an oil-water partition coefficient of no more than 3 and a relative molecular mass of no more than 300 were selected from the non-targeted compound dataset to form a drug-like dataset; 2) Screening of multi-target drug datasets: a. The single-target drug dataset is stored in the form of (drug sequence, target protein sequence) tuples. Targeted drugs with two or more tuples are multi-target drugs. The drug sequences are represented in the form of SMILES sequences. b. Delete binary data where the length of the drug's SMILES sequence is not in [34, 74] and / or the target protein sequence exceeds 1000. Finally, a multi-target drug dataset consisting of multi-target drugs targeting two targets is obtained. Each multi-target drug is stored in the form of a triple (drug sequence, target protein sequence 1, target protein sequence 2).

4. The multi-target drug design method according to claim 1, characterized in that, The target feature embedding module is built based on the ProtTrans model, which obtains the feature vector of the target based on the target protein sequence, including the following steps: 1) Protein sequence tokenization: Tokenize the amino acid residue sequence of a target protein, which is represented by ASCII characters, into a digital token that can be used for neural network computation; 2) Load the pre-trained model: Initialize the target feature embedding module using the parameters of ProtT5-XL version of ProtTrans; 3) Input the tokenized target protein sequence into the target feature embedding module to obtain a feature vector containing multiple target features.

5. The multi-target drug design method according to claim 4, characterized in that, The target feature encoding module is built based on the Transformer framework model. It receives the feature vectors of each target obtained by the target feature embedding module, and independently encodes the protein features of the target to obtain the latent feature vectors of each target. Simultaneously, it performs protein similarity feature encoding on all targets to obtain the similarity latent feature vectors of all targets, including the following steps: 1) Create a Transformer. The formula for the Transformer is as follows: (1) (2) (3) (4) In formula (1): Linear {Q,K,V} Represents three linear layers; Q, K, and V represent the three input vectors in the Transformer; The target feature vector represents the input. W {Q,K,V} Represents Linear {Q,K,V} Weight parameters; In formula (2): Att i Represents the attention mechanism of the i-th head; SoftMax represents the activation function; d k It is the dimension of the input vector K; In formula (3): MSA stands for Multi-head Attention Mechanism; Concat stands for concatenating functions; W O Represents the weight parameters of the output layer; In equation (4): FFN stands for Forward Propagation Network; `max` represents the maximum value. W1 and W2 represent the weight parameters of FFN; b1 and b2 represent the deviation parameters of FFN; 2) Target similarity features are calculated based on the cross-attention mechanism, and latent vectors of target features and target similarity features are obtained by encoding the feature vectors of target proteins. The formula for calculating the latent vectors is as follows: (5) (6) (7) (8) (9) (10) (11) In equations (5) and (6): TF1 and TF2 represent the feature vectors of target point 1 and target point 2, respectively; FF m1 and FF m2 These represent the feature fusion matrices for the two target points respectively; d is the dimension of the feature vector; In equation (7): FF represents the feature fusion vector of the two target points; Concat represents matrix concatenation; In equation (8): O MLP The result represents the computation of the multilayer perceptron; W1, W2, and W3 represent the weight parameters of the multilayer perceptron; b1, b2, and b3 represent the bias parameters of the multilayer perceptron; In equation (9): TSL represents the target similarity latent vector; Transformer ts Transformer computation representing latent vectors of target similarity; In formula (10): TFL1 represents the latent vector of target point 1; Transformer t1 Transformer calculation of the latent vector of target point 1; In equation (11): TFL2 represents the latent vector of target point 2; Transformer t2 Transformer calculation of the latent vector of target point 2.

6. The multi-target drug design method according to claim 5, characterized in that, The multi-target drug generation module is built based on the GRU framework model. It receives target feature latent vectors and target similarity feature latent vectors from the target feature encoding module and generates multi-target drug sequences, including the following steps: 1) The calculation formula for drug sequence generation is as follows: (12) (13) (14) (15) (16) In equation (12): r represents the value of the reset gate in GRU; l represents the layer number in GRU; t represents the time in the GRU; σ represents the sigmoid function; W represents the weight; h represents the hidden state; p represents the SMILES character; In equation (13): z represents the value of the update gate in GRU; In equation (14): Represents a candidate hidden state in GRU; tanh represents the tanh function; In equation (15): b represents deviation; In equation (16): P i This represents the SMILES character output at time i; 2) The target feature latent vector and the target similarity feature latent vector are used as hidden states h and input to the multi-target drug generation module. Under the control of the GRU reset gate and update gate, the module continuously decodes and generates SMILES characters one by one until the SMILES form a complete drug molecule.

7. The multi-target drug design method according to claim 3, characterized in that, The training of the multi-target drug generation module includes calculating the deviation between the SMILES sequence of the real drug molecule and the SMILES sequence of the generated drug molecule through Cross-entropy loss, backpropagating the deviation to the multi-target drug generation module to obtain the parameter gradient, and updating the parameters of the multi-target drug generation module through the Adam algorithm. The training of the multi-target drug design model includes calculating the deviation between the SMILES sequence of the real multi-target drug and the SMILES sequence of the multi-target drug generated by the model through cross-entropy loss, backpropagating the deviation to the multi-target drug design model to obtain the parameter gradient, and updating the parameters of the multi-target drug design model through the Adam algorithm.

8. The multi-target drug design method according to claim 1, characterized in that, It also includes step S6, in which the designed multi-target drug is docked with its corresponding target protein to detect the binding affinity of the drug molecule to each target.

9. A multi-target drug design system based on protein sequence and similarity features, characterized in that, At least including: The target feature embedding module based on ProtTrans is used to obtain the feature vector of the target based on the target protein sequence. The Transformer-based target feature encoding module is used to receive the feature vectors of each target obtained by the target feature embedding module, and independently encode the protein features of each target to obtain the latent feature vector of each target; at the same time, it encodes the protein similarity features of all targets to obtain the latent feature vectors of all targets. A GRU-based multi-target drug generation module is used to receive target feature latent vectors and target similarity feature latent vectors from the target feature encoding module to generate multi-target drug molecules. The target feature embedding module, the target feature encoding module, and the multi-target drug generation module are sequentially arranged to form the multi-target drug design system.

10. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the multi-target drug design method based on protein sequence and similarity features as described in any one of claims 1-8.

Images