WO2019210524A1

WO2019210524A1 - Neural network-based molecular structure and chemical reaction energy function building method

Info

Publication number: WO2019210524A1
Application number: PCT/CN2018/085728
Authority: WO
Inventors: 张佩宇; 方栋; 杨明俊; 马健; 赖力鹏; 温书豪
Original assignee: 深圳晶泰科技有限公司
Priority date: 2018-05-04
Filing date: 2018-05-04
Publication date: 2019-11-07

Abstract

The invention belongs to the technical field of quantum chemistry, and particularly relates to a neural network-based molecular structure and chemical reaction energy function building method. The method comprises the steps of performing sampling on each degree of freedom of a molecular or chemical reaction; searching for a low-energy conformation structure by means of quantum chemistry calculation; performing energy calculation on the structure, and preparing a training set and a test set; selecting a proper coordinate representation structure; according to different coordinates, constructing different features to describe the structure; selecting a proper neural network; selecting a proper method to train the neural network; after the training is completed, performing error statistics on the test set, and when errors are smaller than 1.0 kcal/mol, ending the training; and if errors are greater than 1.0 kcal/mol, re-searching the model following the above procedure. The method allows for higher precision with respect to obtained conformation energy, reaction energy and the like; the method can be widely applied to quantum dynamics and molecular dynamics processes; the method allows for simulation of single molecule conformation and chemical reaction including intramolecular or intermolecular bond breaking and creation.

Description

Neural network based molecular structure and chemical reaction energy function construction method

Technical field

The invention belongs to the technical field of quantum chemistry, and particularly relates to a molecular structure and a chemical reaction energy function construction method based on a neural network, and constructs a potential energy surface through a back propagation neural network.

Background technique

The structure of a molecule plays a decisive role in chemistry (such as organic chemical reactions, conformational polymorphism), organisms (such as drug molecule active conformation, enzyme catalyzed reaction). The structure of organic molecules is not static, and has various conformational degrees of freedom including rotation, stretching, bending, and the like. During the molecular reaction process, there are distances between molecules, relative orientation, bond formation and fracture, and the like. Each structure will correspond to a different energy. The conformational changes and chemical reactions of molecules are very sensitive to energy. Molecular structural changes can be described as moving over the energy function, so a very high-precision description of the energy function is required.

At present, the structure and reaction of organic molecules are described, and most of them are methods using molecular force fields. mainly includes:

In the classical force field, in order to describe intramolecular interactions and intermolecular interactions, a more general functionality has been devised. This functional form includes interaction terms such as bond length, bond angle, and dihedral angle. It also describes the electrostatic interaction term of point charge or polarization and the VDW term describing the interaction of repulsion and dispersion. The advantage of the classical force field lies in the calculation of biomacromolecules. The conformational energy error in small molecular structures is usually 2-3 kcal/mol, and the lower precision limits the industrial application in chemistry or biology. At the same time, the classic force field does not consider the breakage and generation of the bond and cannot be used to simulate the reaction.

The reaction field uses the bond pole to describe the breaking and generation of the bond. The bond level can be obtained directly from the distance between atoms. The function of the key level consists of several exponential functions and correction factors. Usually used to carry out molecular dynamics simulation reaction process. At present, the reaction field is mainly in the simulation process of hydrocarbon reaction, energetic material, combustion and the like. The function form of the reaction force field is very complicated, and many function items have specific physical meanings, which is not conducive to further development and improvement.

Summary of the invention

In view of the above technical problems, the present invention provides a neural network based molecular structure and chemical reaction energy function construction method, which can be used to simulate molecular structure and chemical reaction. The technical solution adopted is:

A neural network based molecular structure and chemical reaction energy function construction method, comprising the following steps:

(1) Sampling the various degrees of freedom of a molecule or chemical reaction; the sampling of each degree of freedom of a molecule or chemical reaction, including: for a molecule, first performing heterogeneous analysis, looking for all isomers, and then for each A heterogeneous conformational sampling; for chemical reactions, based on molecular sampling, the distance and orientation between two molecules involved in the chemical reaction are also sampled.

(2) Finding low-energy conformation structures by quantitative calculation; for chemical reactions, it also includes obtaining possible reaction paths by quantitative calculation.

(3) Perform energy calculation on the structure, prepare training sets and test sets;

(4) Select appropriate coordinates to represent the structure; the coordinates include internal coordinates, Cartesian coordinates, and spherical coordinates.

(5) For different coordinates, different features are constructed to describe the structure; the features include interatomic distance, bond angle, dihedral angle, electrostatic interaction energy, VDW interaction energy, and bond level.

(6) Selecting a suitable neural network; the neural network includes a fully connected neural network and a convolutional neural network, and the activation functions of the neural network include sigmoid and ReLU.

(7) Select appropriate methods to train the neural network; the training strategy includes the selection of the cost function, the learning rate, and the size of the parameters participating in the training.

(8) After the training is completed, the error statistics are performed in the test set. When the error is less than 1.0 kcal/mol, the training ends; if the error is greater than 1.0 kcal/mol, the model is searched again. Re-finding the model follows the following sequence: 1) modifying the training strategy; 2) modifying the neural network model; 3) modifying the features; 4) changing the coordinate system; 4) increasing the training set.

The neural network-based molecular structure and chemical reaction energy function construction method provided by the invention has the following technical effects:

(1) High precision, compared with the conventional force field, the present invention has higher conformational energy and reaction energy, and can be widely applied in quantum dynamics and molecular dynamics processes.

(2) It is easy to expand and does not need to be constrained by existing traditional functional forms. At the same time, it can simulate either a single molecule conformation or a chemical reaction, including intramolecular or intermolecular bond breaking and generation.

DRAWINGS

Figure 1 is a flow chart of the method of the present invention;

2 is a comparison of quantum chemical energy and force field energy of an embodiment;

Figure 3 is a neural network architecture of an embodiment;

4 is an energy comparison of a quantum chemical energy and a model trained with a neural network on a training set of an embodiment;

Figure 5 is an energy comparison of the quantum chemical energy and the model trained with neural networks on the test set of the embodiment.

detailed description

Specific technical solutions of the present invention will be described with reference to the embodiments.

TASELISIB is a selective inhibitor of PIK3CA with the structural formula:

This molecule contains 62 atoms with a molecular weight of 460.542g/mol. The molecule has 6 flexible single bonds that can be rotated, and a relatively large flexible ring. Quantum chemical calculations were performed on this system, and the density functional energy of 2138 constellations was obtained.

The embodiment uses the flow shown in FIG.

The force field of the molecule is extracted from the general force field parameter library. The molecular force field energy was calculated with 2138 structures, and the calculation results are shown in Fig. 2. The coefficient of determination after linear fitting was 0.2942. The definition of the coefficient of determination is 1 minus the ratio of the variance of y to the regression equation and the total variance of y:

The closer the value of the determinable coefficient is to 1, the better the correlation between the energy calculated by the model and the precise quantum chemical energy. The calculated root mean square error is 6.48 kcal/mol, which far exceeds the chemical accuracy of 1 kcal/mol, which reduces the reliability of subsequent kinetic simulation and drug design.

The neural network was trained by using 90% of the 2138 data, that is, 1925 data, as a training set. The remaining 213 structures were used as test sets to test the accuracy of the energy function obtained by the neural network.

In this embodiment, internal coordinates are employed to represent the molecular structure. Each atom establishes a feedback neural network as the input of the neural network through the atomic distances of the neighbors, the next nearest neighbors and the next nearest neighbors. As shown in Fig. 3, the network is divided into an input layer, four hidden layers and one output layer. The number of nodes in the hidden layer is 30*30*30*20, and the output value is molecular energy.

Figure 4 shows a comparison of the energy obtained by the neural network on the training set with the exact quantum chemical energy. The coefficient of determination after linear fitting is 0.95505. The root mean square error is 0.65 kcal/mol, which is less than 1 kcal/mol of chemical precision.

Using this model, simulations were performed on the test set and the results are shown in Figure 5. On the test set, the coefficient of regression after linear fit was 0.93543. The root mean square error is 0.79 kcal/mol, still less than 1 kcal/mol of chemical accuracy. Therefore, this energy can be used for subsequent conformation sampling and drug design work.

Claims

A neural network based molecular structure and chemical reaction energy function construction method, characterized in that it comprises the following steps:

(1) sampling various degrees of freedom of molecular or chemical reactions;

(2) Finding a low-energy conformation structure by quantitative calculation;

(3) Perform energy calculation on the structure, prepare training sets and test sets;

(4) Select a suitable coordinate representation structure;

(5) construct different features to describe the structure for different coordinates;

(6) selecting a suitable neural network;

(7) Select appropriate methods to train the neural network;

(8) After the training is completed, the error statistics are performed in the test set. When the error is less than 1.0 kcal/mol, the training ends; if the error is greater than 1.0 kcal/mol, the re-finding model is followed.
The neural network-based molecular structure and chemical reaction energy function construction method according to claim 1, wherein each of the degrees of freedom of the molecular or chemical reaction is sampled according to the step (1), including: for the molecule, first Perform heterogeneous analysis, find all isomers, and then perform conformation sampling for each isomer; for chemical reactions, based on molecular sampling, the distance and orientation between the two molecules involved in the chemical reaction should be sampled. .
The neural network-based molecular structure and chemical reaction energy function construction method according to claim 1, wherein in the step (2), for the chemical reaction, the possible reaction path is obtained by quantitative calculation.
The neural network-based molecular structure and chemical reaction energy function construction method according to claim 1, wherein the coordinates in the step (4) comprise internal coordinates, Cartesian coordinates, and spherical coordinates.
The neural network-based molecular structure and chemical reaction energy function construction method according to claim 1, wherein the characteristics described in the step (5) include interatomic distance, bond angle, dihedral angle, electrostatic interaction energy, VDW interaction energy, bond level.
The neural network-based molecular structure and chemical reaction energy function construction method according to claim 1, wherein the neural network according to the step (6) comprises a fully connected neural network and a convolutional neural network, and an activation function of the neural network. Includes sigmoid and ReLU.
The neural network-based molecular structure and chemical reaction energy function construction method according to claim 1, wherein the training strategy described in the step (7) comprises a selection of a cost function, a learning rate, and a parameter size participating in the training.
The neural network-based molecular structure and chemical reaction energy function construction method according to claim 1, wherein the step (8) re-finding the model follows the following sequence: 1) modifying the training strategy; 2) modifying the neural network model; Modify the feature; 4) change the coordinate system; 4) increase the training set.