A universal approach to drug design for different target proteins

Through neural networks and deep reinforcement learning, combined with SMILES representation and reward function optimization, the problem of low efficiency of traditional drug design methods is solved, and the rapid design of drug molecules targeting different target proteins is achieved, which improves the efficiency and targeting of new drug design.

CN114464270BActive Publication Date: 2024-11-22BEIJING UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210051113.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-17
Publication Date
2024-11-22
Estimated Expiration
2042-01-17

AI Technical Summary

Technical Problem

Traditional drug design methods are time-consuming and inefficient in exploring new drugs. They fail to effectively utilize most of the chemical space and make it difficult to quickly design drug molecules targeting proteins at different targets.

Method used

Using neural networks and deep reinforcement learning methods, we train molecular generative models and drug target interaction models, and use SMILES representation and reinforcement learning algorithm to generate small molecule inhibitors targeted proteins, combining gated circulating unit neural networks and volumes Accumulate neural networks and use reward functions to optimize the molecular generation process.

Benefits of technology

The rapid design of drug molecules targeting any target protein is achieved, the efficiency and targeting of new drug design is improved, and new molecules with desired drug properties can be generated, making up for the shortcomings of traditional methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114464270B_ABST
    Figure CN114464270B_ABST
Patent Text Reader

Abstract

A universal method for drug design for different target proteins belongs to the field of computer artificial intelligence and new drug design, including training molecular generation models, training drug target affinity prediction models, using the affinity model output as a reward function, and using reinforcement learning to make the molecules generated by the molecular generation neural network have better affinity with proteins. The amino acid sequence of a target protein is input into the trained drug design machine learning model to obtain a small molecule inhibitor compound for this target protein.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the interdisciplinary technical field of computer artificial intelligence and new drug design. It is a universal method for designing new drug molecules for different protein targets based on neural networks and deep reinforcement learning. It is suitable for the design and generation of candidate drug molecules in the process of new drug discovery. Background Art

[0002] Small molecules are mainly used to modulate the function of specific target proteins in biological systems. Traditional drug design methods identify target protein-specific molecules by screening compound libraries available in public and commercial repositories or by generating molecules de novo using fragment and pharmacophore models. Although the screening process is time-consuming, only a portion of the chemical space, approximately 1 billion small molecules, has been explored using traditional methods. However, the actual chemical space is estimated to be 10 63 molecules or more. Deep learning methods can bridge this gap and design a variety of new drug molecules with desired drug properties. The development of deep generative models has produced a series of promising proposals to solve the problem of molecule generation, providing a promising new direction for new drug design. The Simplified Molecular Inline Entry Specification (SMILES) representation of compounds is a string-based representation derived from a molecular graph and has been widely used in this field. For SMILES-based generative models, the one-time encoding of the string is input into a recurrent neural network (RNN) model for learning, and then the string is generated character by character based on the learned distribution. The concept of reinforcement learning is integrated into molecule generation. Fine-tuning a pre-trained RNN-based generator can generate ideal molecules. Based on this, we integrate the drug-target interaction model into reinforcement learning and develop a method that can design small molecules suitable for any target protein. Developing this method can quickly design drugs against any target protein. Summary of the invention

[0003] The purpose of the present invention is to provide a universal method for drug design for different target proteins based on artificial intelligence, by inputting the amino acid sequence of a target protein into a trained drug design machine learning model to obtain a small molecule inhibitor compound library for this target protein.

[0004] In order to achieve the above object, the technical solution of the present invention is as follows:

[0005] A general method for drug design for different target proteins comprises the following steps:

[0006] The first step is to train the molecular generation model. The dataset for training the generative model comes from the CHEMBL database, where these molecules are represented in SMILES format. The molecular generation model is trained on a dataset of 960,000 SMILES strings from the CHEMBL database. The neural network architecture of the generative model uses a gated recurrent unit (GRU) neural network. This model is a pre-trained model for molecular generation.

[0007] Many reinforcement learning problems are called Markov decisions, which means that the current state contains all the necessary information to guide the choice of command, and no information is needed from previous states. For most practical problems, this is only an approximation rather than the truth. However, we can generalize this concept to a partially observable Markov decision. Intelligence can interact with an incomplete representation of the environment. Define r(a|s) as the reward that evaluates the advantage of taking an action in a particular state, then the long-term feedback is defined as the cumulative reward accumulated from time t to time T. Since usually only complete SMILES strings can produce desirable molecular forms, we only refer to the feedback of complete sequences.

[0008] Since using SMILES to represent a molecule requires the neural network to output a series of characters, each character output can be regarded as an action. For example, if a valid SMILES encoded molecule is successfully output and has certain characteristics, the reward function of reinforcement learning will give a certain reward. The reward function here can be a weighted sum of multiple intermediate rewards: medicinal chemistry and drug similarity scores, active chemistry scores, structural scores (suitable for ligand characteristics and / or binding pockets), novelty scores, and diversity scores. The prediction model of the drug-protein affinity neural network is used to predict affinity as a reward function, so that the overall model establishes a relationship with the protein. The molecules generated by the molecular generation neural network are passed to the affinity prediction model, and the affinity of the molecule and the protein is predicted by the prediction model. Affinity is fed back to the molecular generation neural network as a reward function through reinforcement learning, and the molecules generated by the molecular generation neural network have better affinity with the protein through reinforcement learning.

[0009] For drug target affinity prediction models, the drug target affinity prediction model developed based on deep learning can predict for any protein target and drug molecule. The advantage of using deep learning is that it can mine hidden interactions between drugs and targets, and the drug target affinity prediction model currently developed based on deep learning can predict for any protein target and drug molecule. We use the SMILES code of the drug and the amino acid sequence of the protein as input, train through a neural network, and use affinity values ​​as output. The training data is the ligand target data set in the BindingDB database. The SMILES code and amino acid sequence are converted to one-hot code, and then the high-dimensional matrix of the drug molecule and amino acid sequence is converted to the same one-dimensional matrix through convolution operation, and the matrix is ​​spliced. Then the affinity is output through a fully connected neural network to obtain a trained model as a reward function in reinforcement learning.

[0010] For reinforcement learning, both the prior molecular generation model and the agent molecular generation model use molecular generation neural networks. The role of the prior model is to serve as a reference point for the probability of sampling a given SMILES. For each batch of SMILES generated by the agent, the prior model calculates the negative log-likelihood value, denoted as logP(A) prior . P(X i =T i |X i-1 =T i-1 ...X1=x1) is the Xth i Step extraction T i The probability of a character; logP(A) prior Reflects the probability of extracting characters from the model.

[0011]

[0012] The model likelihood of sampling a given SMILES string is given by the product of the action probabilities. Let R(A) be the reward function based on the affinity prediction model. The agent calculates the prior probability by enhancing the likelihood function, as follows:

[0013] logP(A) U =logP(A) prior +δR(A)

[0014] logP(A) prior Represents the likelihood function of the prior model, δ represents the weight of the reward function is 0.1, and a balanced strategy is found between the previous strategy and the agent's strategy. The long-term benefit is written as:

[0015] r(S T )=-[logP(A)U -logP(A) A ] 2

[0016] logP(A) A Represents the likelihood function of the agent, which maximizes the reward by minimizing the objective function

[0017]

[0018] In one embodiment of the present invention, the RDKIT toolkit is used for molecular structure inspection; the deep learning framework uses Pytorch1.7.0, the CUDA version is 11.1, the operating system is Ubuntu LTS 20.04, and all computing work is completed on a 1*Geforce RTX-3090GPU server. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the method flow used in the embodiment of the present invention. DETAILED DESCRIPTION

[0020] The specific implementation modes of the present invention are described below in conjunction with the accompanying drawings to facilitate a better understanding of the present invention.

[0021] like Figure 1 As shown, the universal method for drug design for different target proteins in the present invention comprises the following steps:

[0022] The present invention provides a universal method for generating drug molecules based on drug-target interaction neural networks and deep reinforcement learning, and builds a drug molecule development process for new drug design. Figure 1 As shown, it includes a neural network model for drug molecule generation, a neural network model for predicting affinity of drug-target interaction, and a reinforcement learning module.

[0023] The training data set used by the molecular generation neural network model is: 1.44 million drug molecules from the Chembl (https: / / www.ebi.ac.uk / chembl / ) database; the training data set used by the drug-target interaction neural network is: 13,392 drug molecules, 1,507 protein targets, and 70,750 drug-target pairs from the Binding DB (http: / / www.bindingdb.org / bind / index.jsp) database. The general method for drug molecule generation based on drug-target interaction neural network and deep reinforcement learning includes the following steps:

[0024] Step 1: Train the molecule to generate the neural network model

[0025] The molecular generative neural network consists of 5 layers: 1 input layer, 3 recurrent layers, and 1 output layer. After the SMILES code is converted to a one-hot code representation, the one-hot code is input to the input layer. The recurrent layer uses 128 gated recurrent units (GRU) as cells, and finally outputs the probability of each string through the softmax function.

[0026] A three-layer recursive neural network is used to construct a molecular generative neural network model, which is trained using maximum likelihood estimation. Each layer has 128 gated recursive units. The initial neural network is trained using 1.44 million valid chemical structures (SMILES format) from the Chembl database. The model is trained using stochastic gradient descent for 50,000 steps, with an initial learning rate of 0.001 and a learning rate of 0.02, which decays every 100 steps.

[0027] Step 2: Training the drug-target interaction neural network

[0028] The prediction of protein-ligand interaction is formulated as a regression problem with the goal of predicting the binding affinity score. First, the amino acid sequence and SMILES data are applied to a convolutional neural network (CNN). In the amino acid sequence input, one-hot encoding is used to represent the amino acid sequence in a distributed manner. For the SMILES input, one-hot encoding is used to represent the SMILES in a distributed manner. The feature vectors obtained by the 1D-CNN of the amino acid sequence and the CNN of the protein-drug interaction network are connected to the final output layer through a connection layer to output the affinity score of the drug and the target protein.

[0029] The input lengths of SMILES and amino acid sequences are 1024 and 8420 respectively. When the input lengths of SMILES and amino acid sequences are insufficient, they are padded with the character A. The convolution kernels of drugs and amino acid sequences are [4,6,8] and [4,8,12] respectively. The batch size of a training run is 128, the number of training epochs is 100, the learning rate is 0.001, the optimizer is the adam optimizer, the neurons of the fully connected layer are 1024, 1024, and 512 respectively, and the mean square error (MSE) is used as the loss function, where P is the predicted value and Y corresponds to the true value. Where n represents the number of samples, the formula is as follows:

[0030]

[0031] Step 3: Basic process of molecule generation

[0032] A molecule represented by SMILES with specified desired properties is constructed by using a recurrent neural network as a partially observable Markov decision process, where the agent must decide what character to choose next given the current cell state. Using the probability distribution learned by the previous initial model as our initial prior strategy, the initial network and its strategy are modified to become the agent's network. Therefore, the agent has a neural network with the same architecture as the initial network. From the first step of the recurrent neural network to the end of sampling. The tasks encountered by the agent are all random. Action A = X1X2…X i The collection of SMILES strings is generated in order. i =T i |X i-1 =T i-1 ...X1=x1) is the Xth i Step extraction T i The probability of a character; logP(A) prior Reflects the probability of extracting characters from the model.

[0033] For reinforcement learning, both the prior molecular generation model and the agent molecular generation model use molecular generation neural networks. The role of the prior model is to serve as a reference point for the probability of sampling a given SMILES. For each batch of SMILES generated by the agent, the prior model calculates the negative log-likelihood value, expressed as the likelihood function of the prior model, denoted as logP(A) prior .

[0034]

[0035] Let R(A) be the reward function based on the affinity prediction model. The agent calculates the likelihood function of the agent through the likelihood function of the prior model, and the formula is as follows:

[0036] logP(A) U =logP(A) prior +δR(A)

[0037] δ represents the weight of the reward function is 0.1, R(A) represents the reward function, and in this invention represents the affinity of the drug-target interaction, logP(A) U Represents the enhanced likelihood function of the prior model, using the strategy of the prior model and the strategy of the agent to find a balanced strategy, and the long-term benefit is written as:

[0038] r(S T )=-[logP(A) U -logP(A) Agent ] 2

[0039] logP(A) Agent represents the likelihood function of the agent.

[0040] like Figure 1 As shown in the figure, the amino acid sequence of the protein is input on the right side of the figure, and the neural network on the left side is used as a molecular generation model. The generated molecules are passed to the drug protein prediction model on the right side, and the affinity between the drug and the protein is predicted. The affinity value is passed to the molecular generation model, and the molecular generation model is optimized through reinforcement learning, so that the molecules generated by the molecular generation model are passed to the affinity prediction model between the drug and the protein, so that the molecules have a higher affinity with the protein. After 1000 iterations, the molecule targeting this target protein is obtained. After the above technical process method, the amino acid sequence of a target protein is input into the trained drug design machine learning model to obtain a small molecule inhibitor compound library targeting this target protein. The reward function uses the affinity model of the drug and the protein target. This model can select different amino acid sequences and drugs as inputs, calculate the affinity of different drugs and protein amino acids, and has universality.

[0041] The molecules generated by this scheme are the candidate molecules of new drugs and their structural descriptions are output. This scheme is positioned to assist in the development of new drugs and aims to improve its screening efficiency for candidate molecules. It cannot replace the various physical experiments in the development of new drugs, including biochemical experiments, animal experiments and various phases of clinical trials. New drug development institutions need to complete subsequent experiments to obtain the final new drug compound.

[0042] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A universal method for drug design for different target proteins, characterized in that: The following steps are involved: The first step is to train the molecular generation model to learn the grammar of molecules; The dataset for generative model training comes from the CHEMBL database, and these molecules are represented in SMILES format. The neural network architecture of the generative model uses a gated recurrent unit (GRU) neural network. This model generates a pre-trained model for molecules. Since using SMILES to represent a molecule requires the neural network to output a series of characters, each character output is considered an action; For the drug-target affinity prediction model, the SMILES code of the drug and the amino acid sequence of the protein are used as input, trained through a neural network, and the affinity value is used as output; the training data is the ligand target data set in the BindingDB database, the SMILES code and amino acid sequence are converted into one-hot code, and then the high-dimensional matrix of the drug molecule and amino acid sequence is converted into the same one-dimensional matrix through convolution operation, and the matrix is ​​spliced, and then the affinity is output through a fully connected neural network to obtain the trained universal model as the reward function; For reinforcement learning, both the prior molecular generation model and the agent molecular generation model use molecular generation neural networks. The role of the prior molecular generation model is to serve as a reference point to calculate the model likelihood of a given SMILES sampled; for each batch of SMILES in the agent molecular generation model, the prior molecular generation model calculates the negative log-likelihood value, denoted as logP(A) prior ; The model likelihood of sampling a given SMILES string is given by the product of the action probabilities; let R(A) be the reward function based on the affinity prediction model; the agent molecule generation model calculates the prior probability through the enhanced likelihood function, as follows: logP(A) U =logP(A) prior +δR(A) logP(A) prior represents the likelihood function, δ represents the weight of the reward function is 0.1, and the long-term return is written as: r(S T )=-[logP(A) U -logP(A) A ] 2 logP(A) A Represents the likelihood function of the agent's molecular generative model, maximizing the reward by minimizing the objective function; By predicting the affinity between the drug and the protein, the affinity value is passed to the priori molecule generation model, and the priori molecule generation model is optimized through reinforcement learning, so that the molecules generated by the priori molecule generation model are passed to the affinity prediction model between the drug and the protein, so that the molecules have a higher affinity with the protein; after more than 1,000 iterations, multiple molecules targeting this target protein are obtained; The amino acid sequence of a target protein is input into a trained drug design machine learning model to obtain a library of small molecule inhibitor compounds targeting this target protein.

2. The method according to claim 1, characterized in that: The reward function uses a drug-protein target affinity prediction model, which can select different amino acid sequences and drugs as inputs and calculate the affinity of different drugs with protein amino acids.