Multi-target drug design method based on protein sequences and similarity characteristics

Abstract

The invention discloses a multi-target drug design method based on a protein sequence and similarity characteristics, and the method comprises the steps: designing a multi-target drug design model composed of a target characteristic embedding module, a target characteristic coding module and a multi-target drug generation module, and carrying out the independent coding and similarity coding of the characteristics of a plurality of targets, and according to the coded characteristic latent vector and the similarity characteristic latent vector, directly generating drug molecules aiming at a plurality of protein targets. Wherein the target spot feature embedding module is pre-trained on a large protein data set based on ProtTrans, and has relatively high protein feature embedding capability; the multi-target drug generation module is trained on a large-scale drug data set, so that a relatively high drug generation capability is obtained; then, the whole model is trained on a multi-target drug data set, and good multi-target drug design capability is obtained. The designed drug molecules can be well combined with each target protein, and the purpose of multi-target drug therapy is achieved.

Description

Technical Field

[0001] The invention belongs to the field of computer-aided drug design, and in particular relates to a multi-target drug design method based on protein sequence and similarity characteristics. Background Art

[0002] In recent years, deep learning technology has made great 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 the compound database, and targeted drug design directly designs new targeted drugs that do not exist in the compound database based on the characteristics of the target.

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

[0004] In view of the above problems, the purpose of the present invention is to provide a multi-target drug design method based on protein sequence and similarity characteristics, which can directly design targeted drugs suitable for multiple targets according to the sequence characteristics of multiple target proteins.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

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

[0007] S1. Original data set collection: collect non-targeted compound data sets and single-target drug data sets from the database;

[0008] S2, data preprocessing: preprocessing the non-targeted compound data set obtained in step S1 to obtain a drug-like data set, and preprocessing the single-target drug data set to obtain a multi-target drug data set;

[0009] S3, constructing a multi-target drug design model, including a target feature embedding module for obtaining a target feature vector, a target feature encoding module for obtaining a target feature latent vector and a target similarity feature latent vector, and a multi-target drug generation module for generating drug molecules, and the above three modules are serially connected in sequence to form the multi-target drug design model;

[0010] S4, training the multi-target drug design model: training the multi-target drug generation module with the drug-like data set described in step S2, so that it learns the structural characteristics and generation rules of drug molecules; then, fixing the parameters of the target feature embedding module, and training the entire multi-target drug design model with the multi-target drug data set, 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 a multi-target drug molecule.

[0012] Preferably, the non-targeted compound data set is from the ZINC database, and the single-target drug data set is from the ChEMBL database and the BindingDB database.

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

[0014] 1) Drug-like dataset screening: Data with oil-water partition coefficient values ​​no greater than 3 and relative molecular mass no greater 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. Store the single-target drug dataset in the form of (drug sequence, target protein sequence) binary pairs. Targeted drugs with two or more binary pairs are multi-target drugs, and the drug sequence is represented in the form of SMILES sequence;

[0017] b. Delete the binary data whose SMILES sequence length 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 by screening. Each multi-target drug is stored in the form of a triple of (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 according to the target protein sequence, comprising the following steps:

[0019] 1) Protein sequence tokenization: Tokenize the amino acid residue sequence of the target protein represented by ASCII characters into digital tokens that can be used for neural network calculations;

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

[0021] 3) The tokenized target protein sequence is input 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, which receives the feature vector of each target obtained by the target feature embedding module, and independently encodes the target protein feature to obtain the feature latent vector of each target; at the same time, protein similarity feature encoding is performed on all targets to obtain similarity feature latent vectors of all targets, including the following steps:

[0023] 1) Establish Transformer. The formula of 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 Transformer;

[0030] x represents the input target feature vector;

[0031] W {Q,K,V} Linear {Q,K,V} The weight parameter of

[0032] In formula (2):

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

[0034] SoftMax represents the activation function;

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

[0036] In formula (3):

[0037] MSA stands for multi-head attention mechanism;

[0038] Concat represents the concatenation function;

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

[0040] In formula (4):

[0041] FFN stands for Feed Forward Network;

[0042] max means taking the maximum value;

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

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

[0045] 2) Calculate the target similarity features based on the cross-attention mechanism, and obtain the target feature latent vector and target similarity feature latent vector according to the feature vector encoding of the target protein. The calculation formula of the latent vector 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 formula (5) and formula (6):

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

[0053] FF m1 and FF m2 Represent the feature fusion matrix of two targets respectively;

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

[0055] In formula (7):

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

[0057] Concat represents matrix concatenation;

[0058] In formula (8):

[0059] O MLP Represents the calculation result 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 formula (9):

[0063] TSL represents the target similarity latent vector;

[0064] Transformer ts Transformer calculation of the latent vector representing the target similarity;

[0065] In formula (10):

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

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

[0068] In formula (11):

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

[0070] Transformer t2 Transformer calculation representing the target latent vector for target 2.

[0071] Preferably, the multi-target drug generation module is constructed based on a GRU framework model, which receives the target feature latent vector and the target similarity feature latent vector from the target feature encoding module and generates a multi-target drug sequence, comprising the following steps:

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

[0073]

[0074] In formula (12):

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

[0076] l represents the number of layers in GRU;

[0077] t represents the moment in GRU;

[0078] σ represents the sigmoid function;

[0079] W stands for weight;

[0080] h represents the hidden state;

[0081] p represents SMILES character;

[0082] In formula (13):

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

[0084] In formula (14):

[0085] Represents the candidate hidden state in GRU;

[0086] tanh represents the tanh function;

[0087] In formula (15):

[0088] b represents deviation;

[0089] In formula (16):

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

[0091] 2) The target feature latent vector and the target similarity feature latent vector are input into the multi-target drug generation module as the hidden state h, and the SMILES characters are continuously decoded and generated one by one under the control of the reset gate and update gate of GRU 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, back-propagating the deviation to the multi-target drug generation module to obtain a 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, back-propagating the deviation to the multi-target drug design model to obtain parameter gradients, and updating the parameters of the multi-target drug design model through the Adam algorithm.

[0094] Preferably, the multi-target drug design method further comprises step S6, wherein the designed multi-target drug is docked with its corresponding target protein to detect the binding affinity between the drug molecule and each target.

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

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

[0097] A target feature encoding module based on Transformer is used to receive the feature vector of each target obtained by the target feature embedding module, and independently encode the protein feature of each target to obtain the feature latent vector of each target; at the same time, protein similarity feature encoding is performed on all targets to obtain similarity feature latent vectors of all targets;

[0098] A GRU-based multi-target drug generation module, used to receive the target feature latent vector and the target similarity feature latent vector 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 connected in series to form the multi-target drug design system.

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

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

[0102] (1) The present invention proposes a multi-target drug design method based on protein sequence and similarity features, and designs a multi-target drug design model consisting of a target feature embedding module, a target feature encoding module and a multi-target drug generation module. By independently encoding the features of multiple targets and similarity encoding them, 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, which uses large-scale protein sequence data for pre-training, so that the target feature embedding module of the present invention has a strong protein feature embedding ability, thereby improving the accuracy of the multi-target drug design model.

[0104] (3) The model of the present invention was trained twice. The first training was to train the multi-target drug generation module using a drug-like data set, so that it could learn the structural characteristics and generation rules of drug molecules; the second training was to train the entire multi-target drug design model using a multi-target drug data set composed of triplets, under the premise that the parameters of the target feature embedding module were fixed, so that it could learn the generation rules of multi-target drug molecules and obtain good multi-target drug design capabilities. The two trainings further improved the accuracy of the multi-target drug design model of the present invention. The designed drug molecules were proved to be well combined with the corresponding protein targets through molecular docking, achieving the purpose of multi-target drug treatment, and providing a practical new method for computer-aided drug design. BRIEF DESCRIPTION OF THE DRAWINGS

[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 A basic structural diagram of a target feature encoding module according to an embodiment of the present invention;

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

[0108] Figure 4 for Figure 3 Conformational diagram of the docking of multi-target drugs with 3CLpro targets;

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

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

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

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

[0113] Fig. 9 for Figure 8 Molecular 3D structure diagram of the multi-target drug when docked with the PLpro target. DETAILED DESCRIPTION

[0114] In order to make the purpose, technical scheme and advantages of the present invention more clearly understood, the multi-target drug design method based on protein sequence and similarity characteristics of the present invention is specifically described below in conjunction with the accompanying drawings and examples. It should be understood that the following implementation is only a specific case of the present invention and is not intended to limit the scope of protection of the present invention.

[0115] 1. Terminology explanation:

[0116] SMILES: is a specification that uses ASCII strings to clearly describe molecular structures;

[0117] ProtTrans: a protein language model for deep learning.

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

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

[0120] GRU: represents a framework model based on recurrent neural network 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 compound virtual screening database containing more than 750 million purchasable compounds, provided and maintained by the Irwin and Shoichet laboratories in the Department of Medicinal Chemistry at the University of California, San Francisco (UCSF), website: http: / / zinc15.docking.org / ;

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

[0128] BindingDB database: a publicly accessible database that mainly collects the interaction affinity between drug target proteins and drug-like small molecules. The website address is: 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 characteristics. For the sake of convenience, this embodiment takes the design of drugs targeting two targets as an example. It should be noted that the multi-target drug design method of the present invention includes but is not limited to drug design targeting two targets. Furthermore, this embodiment specifically selects the 3CL protease (3CLpro) target and the papain-like protease (PLpro) target of the SARS-CoV-2 virus in the COVID-19 influenza to design a multi-target drug for the SARS-CoV-2 virus. The detailed steps are as follows:

[0131] S1. Original data set collection: Obtain non-targeted compound data sets and single-target drug data sets from the database.

[0132] In this example, scripts were used to collect non-targeted compound data sets from the ZINC database, and single-target drug data sets from the ChEMBL database and the BindingDB database.

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

[0134] S201. Screening of drug-like data sets: According to Lipinski's Rule of Five (Ro5), compounds with an oil-water partition coefficient (LogP) not greater than 3 and a molecular weight not greater than 300 have strong drug-like properties. Therefore, 87,719,678 data with a LogP value not greater than 3 and a relative molecular mass not greater than 300 are selected from the untargeted compound data set of step S1 to form a drug-like data set.

[0135] S202, Multi-target drug dataset screening:

[0136] 1) Single-target drugs collected from ChEMBL and BindingDB databases are stored in the form of (drug sequence, target protein sequence) pairs. If there are two or more pairs with the same drug sequence but different target protein sequences, the drug corresponding to this drug sequence is a multi-target drug, and 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 the sequence length of more than 95% of the proteins in the database is less than 1000, and the SMILES sequence length of more than 95% of the drug molecules is between [34, 74], in order to reduce the variance of the data length in the training set and keep the model stable, the data with target protein sequence longer than 1000 were discarded, and the drug molecules with SMILES sequence length not between [34, 74] were discarded. Finally, a total of 52,896 triplet data of multi-target drugs that can target two targets were selected, each triplet was (drug sequence, target protein sequence 1, target protein sequence 2), forming a multi-target drug dataset.

[0138] S203. Preprocessing of drug molecule data: In order to enhance the boundary features of the SMILES sequence of drug molecules, better represent the beginning and end of the molecule, and make all SMILES sequences of the same length, add a 'G' character at the beginning of each SMILES sequence, add an 'E' character at the end, and fill the 'E' character with a 'Q' character to make each SMILES sequence length 76.

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

[0140] S3. Construction of a multi-target drug design model based on target protein sequences and similarity features: see Figure 1In this embodiment, the protein sequence features of target 1 and target 2 are analyzed, and 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 serially form a multi-target drug design model. The specific method is:

[0141] S301, constructing a target feature embedding module based on ProtTrans, and obtaining a feature vector of the target according to the protein sequence of the target, including the following steps:

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

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

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

[0145] S302, constructing a target feature encoding module based on Transformer: passing the feature vector of each target obtained in step S301 as input to the target feature encoding module based on Transformer. The basic structure of the target feature encoding module is as follows: Figure 2 As shown in the figure, the Transformer model will further process and encode the input feature vector to extract deeper feature representation, that is, obtain the feature latent vector of each target; while encoding the individual protein features of each target, it will encode the protein similarity features of all targets to obtain the similarity feature latent vectors of all targets. The following steps are included:

[0146] 1) Establish Transformer. The formula of 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 Transformer;

[0153] x represents the input target feature vector;

[0154] W {Q,K,V} Linear {Q , K,V} The weight parameter of

[0155] In formula (2):

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

[0157] SoftMax represents the activation function;

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

[0159] In formula (3):

[0160] MSA stands for multi-head attention mechanism;

[0161] Concat represents the concatenation function;

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

[0163] In formula (4):

[0164] FFN stands for Feed Forward Network;

[0165] max means taking the maximum value;

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

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

[0168] 2) Calculate the target similarity features based on the cross-attention mechanism, and obtain the target feature latent vector 1, feature latent vector 2 and target similarity feature latent vector according to the feature vector encoding of all target proteins. The calculation formula of the latent vector 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 formula (5) and formula (6):

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

[0176] FF m1 and FF m2 Represent the feature fusion matrix of two targets respectively;

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

[0178] In formula (7):

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

[0180] Concat represents matrix concatenation;

[0181] In formula (8):

[0182] O MLP Represents the calculation result 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 formula (9):

[0186] TSL represents the target similarity latent vector;

[0187] Transformer ts Transformer calculation of the latent vector representing the target similarity;

[0188] In formula (10):

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

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

[0191] In formula (11):

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

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

[0194] S303, construct a multi-target drug generation module based on GRU, and obtain a multi-target drug sequence according to the feature latent vector of a single target and the similarity feature latent vector output of all targets. The steps include:

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

[0196]

[0197]

[0198] In formula (12):

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

[0200] l represents the number of layers in GRU;

[0201] t represents the moment in GRU;

[0202] σ represents the sigmoid function;

[0203] W stands for weight;

[0204] h represents the hidden state;

[0205] p represents SMILES character;

[0206] In formula (13):

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

[0208] In formula (14):

[0209] Represents the candidate hidden state in GRU;

[0210] tanh represents the tanh function;

[0211] In formula (15):

[0212] b represents deviation;

[0213] In formula (16):

[0214] P i 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 input as hidden states h to the multi-target drug generation module, and the SMILES characters are continuously decoded and generated one by one under the control of the reset gate and update gate of the GRU until the SMILES form a complete drug molecule sequence.

[0216] S4. Training of 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 multi-target drug generation module.

[0218] The drug-like dataset containing 87,719,678 data items is used to train the multi-target drug generation module to learn the structural characteristics and generation rules of drug molecules, including the following steps:

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

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

[0221] 4) Back-propagating the deviations between SMILES sequences to the multi-target drug generation module to obtain parameter gradients, and updating the parameters of the multi-target drug generation module through the Adam algorithm;

[0222] 5) After 10 rounds of training on the drug-like dataset, the multi-target drug generation module ends training and saves the parameters of this module.

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

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

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

[0226] 2) Calculate 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;

[0227] 3) Back-propagating the deviations between SMILES sequences to the multi-target drug design model to obtain parameter gradients, and updating the parameters of the multi-target drug design model using the Adam algorithm;

[0228] 4) After 10 rounds of training on the multi-target drug dataset, the multi-target drug design model ends training and saves the model parameters.

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

[0230] The target protein sequence 1 and target protein sequence 2 targeted by the drug molecule to be designed are input into the multi-target drug design model that has been trained above, and SMILES characters are generated one by one until SMILES forms a complete drug molecule sequence. The generation process of the drug molecule is as follows: Figure 3 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 'X' character, and then inputting the protein sequences of the two targets into the multi-target drug design model to obtain drug molecules targeting the two targets.

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

[0233] S601, obtaining target protein structure data: obtaining the structure data of each target protein from the PDB database;

[0234] S602, Target protein data preprocessing: There are ligands and water molecules in the protein structure of the PDB database. The ligands do not belong to the protein and occupy the pocket structure of the protein. The water molecules will interfere with the binding of targeted drugs to proteins. Therefore, the ligands and water molecules in the protein structure are removed in advance;

[0235] S603, checking the torsion center and torsion bond for the multi-target drug, so that the multi-target drug forms a molecular structure that adapts to the protein pocket structure during molecular docking;

[0236] S604, adding hydrogen ions to the target protein and calculating the charge number, so that the target protein can form a stable hydrogen bond and other docking structure with the multi-target drug 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, and multi-target drugs with strong binding affinity, tight binding sites and stable binding conformations are obtained based on the docking results.

[0238] In this embodiment, Figure 4-9 The generated conformational map, site map and 3D structure map of the multi-target drug docked with the 3CLpro target and PLpro target are shown respectively. In molecular docking, the atomic contacts between the protein fragment and the ligand are interaction pairs, and the interatomic distance must be ≤ Figure 5 It showed that there were two interaction pairs between the drug molecule and the 3CLpro target, with binding distances of 2.9 and 2.0, respectively; Figure 8 It shows that there are four interaction pairs between the drug molecule and the PLpro target, with binding distances of 2.6, 2.8, 3.2 and 2.7 respectively. It can be seen that the distances between the generated multi-target drug molecules and the 3CLpro target and the PLpro target all meet the requirement that the atomic distance between a single interaction pair is less than The requirements are met, and the binding conformation between the drug molecule and each target protein is good, the binding state is stable, and it has potential multi-target drug treatment effects, which proves that the multi-target drug design method of the present invention has better multi-target drug design effects and has good application prospects.

Claims

1. A multi-target drug design method based on protein sequence and similarity characteristics, characterized in that: The following steps are involved: S1. Original data set collection: collect non-targeted compound data sets and single-target drug data sets from the database; S2, data preprocessing: preprocessing the non-targeted compound data set obtained in step S1 to obtain a drug-like data set, and preprocessing the single-target drug data set to obtain a multi-target drug data set; S3, constructing a multi-target drug design model, including a target feature embedding module for obtaining a target feature vector, a target feature encoding module for obtaining a target feature latent vector and a target similarity feature latent vector, and a multi-target drug generation module for generating drug molecules, and the above three modules are serially connected in sequence to form the multi-target drug design model; S4, training the multi-target drug design model: training the multi-target drug generation module with the drug-like data set described in step S2, so that it learns the structural characteristics and generation rules of drug molecules; then, fixing the parameters of the target feature embedding module, and training the entire multi-target drug design model with the multi-target drug data set, 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 a multi-target drug molecule.

2. The multi-target drug design method according to claim 1, characterized in that: The non-targeted compound data set is from the ZINC database, and the single-target drug data set is from the ChEMBL database and the BindingDB database.

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 screening: Data with oil-water partition coefficient values ​​no greater than 3 and relative molecular mass no greater than 300 were selected from the non-targeted compound dataset to form a drug-like dataset; 2) Screening of multi-target drug datasets: a. Store the single-target drug dataset in the form of (drug sequence, target protein sequence) binary pairs. Targeted drugs with two or more binary pairs are multi-target drugs, and the drug sequence is represented in the form of SMILES sequence; b. Delete the binary data whose SMILES sequence length 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 by screening. Each multi-target drug is stored in the form of a triple of (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 constructed based on the ProtTrans model, which obtains the feature vector of the target according to the target protein sequence, including the following steps: 1) Protein sequence tokenization: Tokenize the amino acid residue sequence of the target protein represented by ASCII characters into digital tokens that can be used for neural network calculations; 2) Load the pre-trained model: Use the parameters of ProtT5-XL version of ProtTrans to initialize the target feature embedding module; 3) The tokenized target protein sequence is input 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 constructed based on the Transformer framework model, which receives the feature vector of each target obtained by the target feature embedding module, and independently encodes the target protein feature to obtain the feature latent vector of each target; at the same time, protein similarity feature encoding is performed on all targets to obtain similarity feature latent vectors of all targets, including the following steps: 1) Establish Transformer. The formula of Transformer is as follows: Linear {Q,K,V} (x)=xW {Q,K,V} (1) FFN(x)=max(0,xW1+b1)W2+b2 (4) In formula (1): Linear {Q,K,V} represents three linear layers; Q, K, and V represent the three input vectors in Transformer; x represents the input target feature vector; W {Q,K,V} Linear {Q,K,V} The weight parameter of In formula (2): Att i Represents the i-th attention mechanism; SoftMax represents the activation function; d k is the dimension of the input vector K; In formula (3): MSA stands for multi-head attention mechanism; Concat represents the concatenation function; W O Represents the weight parameters of the output layer; In formula (4): FFN stands for Feed Forward Network; max means taking the maximum value; W1 and W2 represent the weight parameters of FFN; b1 and b2 represent the bias parameters of FFN; 2) Calculate the target similarity features based on the cross-attention mechanism, and obtain the target feature latent vector and target similarity feature latent vector according to the feature vector encoding of the target protein. The calculation formula of the latent vector is as follows: <h2 style=";text-align:left;direction:ltr">O<h2 style=";text-align:left;direction:ltr"> MLP <h2 style=";text-align:left;direction:ltr"> =W3(W2(W1×FF+b1)+b2)+b3 (8) TSL=Transformer ts (O MLP ) (9) TFL1=Transformer t1 (TFL1) (10) TFL2=Transformer t2 (TF2) (11) In formula (5) and formula (6): TF1 and TF2 represent the feature vectors of target 1 and target 2, respectively; FF m1 and FF m2 Represent the feature fusion matrix of two targets respectively; d is the dimension of the feature vector; In formula (7): FF represents the feature fusion vector of two targets; Concat represents matrix concatenation; In formula (8): O MLP Represents the calculation result 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 formula (9): TSL represents the target similarity latent vector; Transformer ts Transformer calculation of the latent vector representing the target similarity; In formula (10): TFL1 represents the target latent vector of target 1; Transformer t1 Transformer calculation of the target latent vector representing target 1; In formula (11): TFL2 represents the target latent vector of target 2; Transformer t2 Transformer calculation representing the target latent vector for target 2.

6. The multi-target drug design method according to claim 5, characterized in that: The multi-target drug generation module is constructed based on the GRU framework model, which receives the target feature latent vector and the target similarity feature latent vector from the target feature encoding module and generates a multi-target drug sequence, including the following steps: 1) The calculation formula for drug sequence generation is as follows: In formula (12): r represents the value of the reset gate in GRU; l represents the number of layers in GRU; t represents the moment in GRU; σ represents the sigmoid function; W stands for weight; h represents the hidden state; p represents SMILES character; In formula (13): z represents the value of the update gate in GRU; In formula (14): Represents the candidate hidden state in GRU; tanh represents the tanh function; In formula (15): b represents deviation; In formula (16): P i Represents the SMILES character output at time i; 2) The target feature latent vector and the target similarity feature latent vector are input into the multi-target drug generation module as the hidden state h, and the SMILES characters are continuously decoded and generated one by one under the control of the reset gate and update gate of GRU 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-like molecule and the SMILES sequence of the generated drug molecule through Cross-entropyloss, and back-propagating the deviation to the multi-target drug generation module to obtain a 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-entropyloss, back-propagating the deviation to the multi-target drug design model to obtain parameter gradients, 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: The method further comprises step S6, wherein the designed multi-target drug is docked with its corresponding target protein to detect the binding affinity between the drug molecule and each target.

9. A multi-target drug design system based on protein sequence and similarity features, characterized in that: At least: The target feature embedding module based on ProtTrans is used to obtain the target feature vector according to the target protein sequence; A target feature encoding module based on Transformer is used to receive the feature vector of each target obtained by the target feature embedding module, and independently encode the protein feature of each target to obtain the feature latent vector of each target; at the same time, protein similarity feature encoding is performed on all targets to obtain similarity feature latent vectors of all targets; A GRU-based multi-target drug generation module, used to receive the target feature latent vector and the target similarity feature latent vector 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 connected in series 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, the steps of a multi-target drug design method based on protein sequence and similarity features as described in any one of claims 1 to 8 are implemented.

Images