Molecular stable conformation generation framework of pre-training and geometric enhancement graph network

Through pre-training and geometric enhancement graph network framework, the atomic-chemical bond-bond angle-dihedral angle information is fused, combined with self-supervised pre-training tasks and variational autoencoder, the efficiency and accuracy problems of macromolecular conformation generation are solved, and the efficiency and accuracy of macromolecular conformation generation are achieved. Fast and accurate high-quality conformation generation and rotation-translation invariance.

CN118969128APending Publication Date: 2024-11-15RES INST OF CHEM DEFENSE PLA ACAD OF MILITARY SCI

Patent Information

Application Number
CN202410976671.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-21
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The prior art is difficult to quickly and accurately generate a stable three-dimensional conformation of macromolecules with high quality. The traditional methods are inefficient and costly. The existing machine learning methods perform poorly on macromolecules and are difficult to maintain rotation-translation invariance.

Method used

Pre-training and geometric enhanced graph network framework, including three-dimensional graph neural network D-GeoGNN and conformation generation model Conf-GEM, combines self-supervised pre-training tasks and variable segmentation by fusing information from atomic-chemical bond-bond angle-dihedral angle information. Autoencoder, realizes multi-scale molecular characterization and conformation generation.

Benefits of technology

It realizes rapid and accurate high-quality three-dimensional conformation generation of molecules, reduces calculation costs, can distinguish subtle differences between different conformations, and maintains rotation-translation invariance, which is suitable for conformation generation of macromolecules.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118969128A_ABST
    Figure CN118969128A_ABST
Patent Text Reader

Abstract

The invention relates to a molecular stable conformation generation framework of a pre-training and geometric enhancement graph network, which belongs to the field of chemical informatics and computer science and comprises a three-dimensional (3D) graph neural network D-GeoGNN, a pre-training task D-SSLs and a conformation generation model Conf-GEM. The D-GeoGNN has the capability of representing 3D structural information of molecules by inputting three different molecular graphs and combining with a self-supervised pre-training task of molecular geometry. On the basis, a model for rapid de novo generation of the molecular stable 3D conformation is constructed by using a variational auto-encoder and a division and treatment thought. The invention provides a brand-new molecular structure representation learning framework, and complete representation of molecular 3D structure information and rapid and accurate generation of low-energy reasonable 3D conformation are realized by fusing information of'bond length-bond angle-dihedral angle 'and combining a molecular multi-scale self-supervision pre-training task, a variational auto-encoder and a divide-and-conquer idea.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of chemical informatics and computer science, and specifically relates to a molecular stable conformation generation framework of a pre-trained and geometrically enhanced graph network, which is mainly used to quickly, accurately and batch generate high-quality molecular stable three-dimensional (3D) conformations. Background Art

[0002] The data of molecular structure is actually distributed in three-dimensional (3D) space. The 3D structure of a molecule, especially the most stable equilibrium ground state geometric structure, is often the key to determining its physical and chemical properties and biological activity. Therefore, obtaining a reasonable and stable 3D conformation of a molecule is very important for the design of new molecules and new materials. In addition, high-precision stable 3D conformations of molecules help strengthen the molecular representation learning model to better capture the structural information of the molecule at the atomic level, ensuring a more comprehensive representation of the molecular structure.

[0003] The purpose of generating stable 3D conformations of molecules is to generate 3D coordinates of each atom in the molecule to support various downstream molecular design tasks, including molecular property prediction, molecular generation, molecular dynamics, molecular docking, and structure-based virtual screening. Traditionally, stable 3D conformations of molecules are mainly obtained through experimental methods based on X-ray diffraction crystal structure analysis or theoretical calculation methods based on quantum chemistry (such as density functional theory (DFT)). However, these two methods have the defects of low efficiency and high cost, which seriously hinder the acquisition and wide application of high-precision stable 3D conformations of molecules. Although low-precision molecular force fields (such as UFF, MMFF, etc.) can be used to improve the efficiency of molecular conformation generation, the conformations generated are very rough, differ greatly from the real conformations, and are unstable. Therefore, it is necessary to use machine learning models to reconstruct stable and reasonable 3D conformations of molecules, which can significantly reduce the computational cost and accelerate the acquisition of high-quality molecular 3D structures.

[0004] At present, the methods for de novo generation of molecular 3D conformations based on machine learning are mainly divided into two categories: indirect prediction based on interatomic distances and direct prediction of generating atomic coordinates. Among them, the first method first predicts information such as the distance between atoms and bond angles, and then restores the molecular conformation based on topology. The model based on this type of method is essentially rotation-translation invariant, that is, the molecular conformation will not change with the rotation, translation and other operations of the coordinates, but the predicted structural information may violate mathematical geometry rules, such as the sum of the lengths of the two sides of a triangle is less than the third side. The second method is to directly predict the 3D coordinates corresponding to each atom. This method is in line with the intuition of chemists and can avoid the problem of violating mathematical geometry rules, but a challenge faced by this type of method is how to maintain the rotation-translation invariance of the molecule. Therefore, most of the current research on molecular conformation generation uses indirect prediction methods, but direct generation methods obviously have great potential. In addition, although the existing methods have good performance on smaller molecules (with no more than 9 atoms), they perform poorly on larger molecules, that is, molecules with more application value. Therefore, it is extremely necessary to design a generative model that can meet the needs of large-sized flexible molecules to quickly generate high-quality 3D conformations to promote the development of molecular design based on artificial intelligence. Summary of the invention

[0005] (I) Purpose of the invention and technical problem

[0006] The purpose of the present invention is to provide a molecular stable conformation generation framework of a pre-trained and geometrically enhanced graph network, which overcomes the defect that the current molecular representation learning model can only represent partial geometric structure information of the molecule and ignores the absolute configuration of the molecule, especially the 3D electronic structure information.

[0007] (II) Technical solution

[0008] In order to achieve the above purpose and solve the above technical problems, the solution adopted by the present invention is as follows:

[0009] A molecular stable conformation generation framework based on pre-trained and geometry-enhanced graph networks, including a three-dimensional graph neural network D-GeoGNN, a pre-training task D-SSLs, and a conformation generation model Conf-GEM.

[0010] 1. 3D Graph Neural Network D-GeoGNN

[0011] like Figure 1 As shown in the figure, the three-dimensional graph neural network D-GeoGNN is composed of three isomorphic graph neural networks connected in series, each accepting input from three graphs:

[0012] The graph representing “atom-chemical bond” G=(V,E);

[0013] The graph representing “chemical bond-bond angle” is H = (E, A);

[0014] The diagram representing the “bond angle-dihedral angle” is I = (A, D);

[0015] Where V is the set of atoms, E is the set of chemical bonds, A is the set of bond angles, and D is the set of dihedral angles. In graph G, atom u∈V is regarded as a node of the graph, and the chemical bond (u,v)∈E connecting atoms u and v is regarded as an edge of the graph; while the nodes of graph H are chemical bonds (u,v)∈E, and the bond angle (u,v,w)∈A connecting chemical bonds (u,v) and (v,w) is regarded as an edge of graph H; finally, for graph I, the bond angle (u,v,w)∈A is regarded as a node of the graph, and the dihedral angle (u,v,w,x)∈D formed by the bond angle (u,v,w) and (v,w,x) is regarded as an edge; using x u to represent the initial characteristics of atom u, x uv represents the characteristics of the chemical bond (u,v), x uvw is the initial characteristic of the bond angle (u, v, w), x uvwx The above three images and the four initial features are used as the input of the three-dimensional graph neural network D-GeoGNN.

[0016] The three-dimensional graph neural network D-GeoGNN is also an information transmission network. During the training process, the hidden state h of the chemical bond uv The hidden state h of the bond angle uvw They are used for communication between graphs G and H, and between graphs H and I respectively; at the tth iteration: first, in the bond angle-dihedral graph I, the information of neighbor bond angles and dihedral angles connecting neighbor bond angles is gathered to update the new hidden state of the bond angle

[0017]

[0018] in, represents the atoms adjacent to atom u, Represents the neighboring bond angles of the bond angle (u,v,w); In the chemical bond-angle diagram I, the function Agg I Aggregate the information to be transmitted, and then pass it through the function Comb. I renew for Then, in the "chemical bond-bond angle" graph H, the neighbor nodes of the chemical bond and the bond angle information connecting the neighbor nodes are aggregated to obtain the hidden state of the chemical bond

[0019]

[0020] in, Represents the neighboring chemical bonds of a chemical bond (u, v); through the function Agg H Aggregate information Then through the function Comb. H renew for Next, in graph G, like most other GNNs, the information of the neighbor nodes of the atoms and the chemical bonds connecting the atoms is aggregated to obtain the current atomic hidden state

[0021]

[0022] Finally, the READOUT function is used to pool the atomic hidden state set obtained in the last iteration to obtain the eigenvector h of the molecule. G :

[0023]

[0024] Where K is the number of iterations, h G is used for prediction tasks at the molecular level, while It is used for prediction tasks at the atomic level.

[0025] (II) Pre-training Task D-SSLs

[0026] Pre-training tasks D-SSLs includes four pre-training tasks for three-dimensional geometric structures, namely: (1) predicting bond length; (2) predicting bond angle; (3) predicting dihedral angle; and (4) predicting the distance between two atoms.

[0027] The pre-training task D-SSLs is carried out in two steps. First, a large-scale unlabeled dataset is used to pre-train D-GeoGNN at the three-dimensional geometry level. During this process, the molecular conformation is calculated by the MMFF94 force field. The process of the pre-training task D-SSLs is explained in detail below.

[0028] Step 1: Use a large-scale unlabeled dataset to pre-train D-GeoGNN at the 3D geometry level. During this process, the molecular conformation is calculated by the MMFF94 force field.

[0029] For the pre-training of local geometric structures, i.e. the prediction tasks of bond lengths, bond angles and dihedral angles, the present invention adopts a training strategy of masked learning: first, 15% of the atoms in the molecule are randomly selected, and for each selected atom, the first-order neighborhood of the atom is extracted, including adjacent atoms and chemical bonds as well as the bond angles and dihedral angles formed by the selected atom. Next, the features of the selected atoms, chemical bonds, bond angles and dihedral angles are masked. Finally, the hidden states of the masked chemical bonds, bond angles and dihedral angles extracted by D-GeoGNN after the last iteration are used to predict the corresponding bond lengths, bond angles and dihedral angles, respectively, and the loss functions are:

[0030]

[0031] Where L length , L angle With L dihedral represents the loss function in the prediction task of bond length, bond angle and dihedral angle respectively; ε, A and D represent the chemical bond set, bond angle set and dihedral angle set respectively; f length (·),f angle (·) and f dihedral (·) represent the predicted values ​​of the neural network for bond length, bond angle and dihedral angle respectively; and l uv ,φ uvw and θ uvw are the true values ​​of the predicted values ​​of bond length, bond angle, and dihedral angle, respectively;

[0032] For the global geometry task, i.e. predicting the distance between any two atoms in a molecule, the present invention constructs an atomic distance matrix for each molecule based on its three-dimensional coordinates, and then predicts the distance between atom-atom pairs. For a molecule, the predicted atomic distance matrix is ​​considered as a multi-classification problem, and the atomic distances are projected into 30 equidistant components. The task of predicting bond lengths can be considered as a special case of the task of predicting atomic distances. The loss function for predicting the distance between any two atoms in a molecule is:

[0033]

[0034] During the pre-training process of D-GeoGNN, the total loss is the sum of the above losses. The D-GeoGNN after pre-training is referred to as D-GEM in the present invention.

[0035] (III) Conformation generation model Conf-GEM

[0036] The conformation generation model Conf-GEM adopts a variational autoencoder architecture, which consists of a prior model, an encoder model, and a decoder model. The base model of each model structure is D-GEM.

[0037] like Figure 2As shown in the figure, the general training process is as follows: (1) The prior model accepts a two-dimensional molecule, converts it into a three-dimensional molecule based on the MMFF94 force field as input, and then outputs atoms, chemical bonds, bond angles and dihedral angles as well as a generated conformation; (2) The encoder model accepts a real conformation and outputs the mean and variance required to construct a latent vector that conforms to the Gaussian distribution; (3) The outputs of the above two steps are used as the input of the decoder, and the decoder model will eventually output a predicted conformation.

[0038] The present invention defines ρ(·) to represent the rotational translation operation of a molecule, that is, in The three-dimensional rotation group represents the rotation operation of the conformation, b∈R (1×3) represents a translation operation. In addition, let and Denote the true conformation and the generated conformation respectively. The rotation-translation-invariant loss function is defined as:

[0039]

[0040] in‖·‖ F represents the Frobenius norm, that is L RT The function has rotation and translation invariance to the two input configurations. For this function, the present invention uses the quaternion method to solve it.

[0041] In addition, in the rotation-translation invariant loss function, the conformation generation model tends to generate "standard" conformations without considering diversity. In order to make the conformations generated by the conformation generation model have diversity, the present invention adds a probability sampling step in the generation process of the conformation generation model. That is, a random hidden variable z is introduced, and then conformation generation is performed under conditional probability:

[0042]

[0043] Among them, γ is a hyperparameter; The configuration generation model is conditioned on the latent variable z and the molecular graph; the sampling of the latent variable z conforms to the standard normal distribution.

[0044] Place Figure 3 As shown in the figure, the inference of the conformation generation model only requires the use of the prior model and the decoder model, specifically: (1) for a given two-dimensional molecule, the representation of atoms, chemical bonds and molecular graphs and the preliminary predicted conformation are obtained through the encoding of the prior model; (2) a latent variable z is randomly extracted from the standard normal distribution; (3) the three representations, the conformation output by the prior model and the latent vector are input into the decoder to obtain the final generated conformation.

[0045] For the prior model, encoder model and decoder model in the conformation generation model Conf-GEM, they are all composed of three D-GEMs in series. The divide-and-conquer strategy for each model is as follows:

[0046] (1) The first GEM accepts an initial molecular conformation Generating conformations Then the bond length l is calculated based on this conformation, that is: Then update the bond length l to the "atom-chemical bond graph" G as the input of the next model. It should be noted that for the prior model is a uniformly distributed coordinate from -1 to 1, the encoder model is the true conformation, and the decoder model is the final output coordinate of the prior model;

[0047] (2) The second GEM accepts the previous step Generate conformation The bond angle θ is then calculated based on this conformation, namely: Then the bond angle θ is updated to the "chemical bond-bond angle diagram" H as the input of the next model;

[0048] (3) The third GEM accepts the previous step Generate conformation Then the dihedral angle ω is calculated based on this conformation, namely: is the final output conformation of the model.

[0049] Finally, the loss function obtained in the pre-training task D-SSLs is used to calculate the losses of bond length, bond angle, and dihedral angle, and the final loss is obtained by summing them up.

[0050] (III) Effective income

[0051] The present invention integrates the information of "bond length-bond angle-dihedral angle" and combines self-supervised pre-training tasks of different molecular scales, variational autoencoders and the idea of ​​divide and conquer to achieve multi-scale complete characterization of molecules and rapid and accurate generation of low-energy conformations.

[0052] (1) The D-GeoGNN in the present invention achieves the encoding of the complete three-dimensional structure of the molecule by fusing the information of the three images of "atom-chemical bond", "chemical bond-bond angle" and "bond angle-dihedral angle". After pre-training D-SSLs at the geometric level, D-GeoGNN has the ability to distinguish subtle differences between different conformations. The model structure of D-GeoGNN is similar to that of pre-training task D-SSLs. Figure 1 Shown

[0053] (2) Due to the complexity of molecular geometric information, directly generating conformations will lead to a huge model search space. The Conf-GEM in the present invention is based on the idea of ​​divide and conquer, gradually determining the bond length, bond angle and dihedral angle of the molecule, which greatly reduces the search space of the model and speeds up the generation of molecules. Figure 4 shown. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 Three-dimensional graph neural network D-GeoDNN structure and pre-training task D-SSLs;

[0055] In the figure: +, tensor addition operation; 1-1, atom-chemical bond encoder; 1-2, chemical bond-bond angle encoder; 1-3, bond angle-dihedral angle encoder; 2, isomorphic graph neural network; 3, layer normalization; 4, graph normalization; 5, Relu activation function; 6-1, bond length loss function; 6-2, bond angle loss function; 6-3, dihedral angle loss function; 6-4, distance matrix loss function; 7, total loss.

[0056] Figure 2 Schematic diagram of the training framework of the conformation generation model Conf-GEM;

[0057] In the figure: Two-dimensional molecules; Three-dimensional molecules; +, tensor addition operation; z, tensor that conforms to the Gaussian distribution; 8-1, prior model; 8-2, encoder model; 8-3, decoder model; 9-1, atomic, chemical bond, bond angle and dihedral angle characteristics and generated conformations; 9-2, used to generate the mean and variance that conforms to the Gaussian distribution; 10. Loss function for calculating coordinate errors.

[0058] Figure 3 Schematic diagram of the inference framework of the conformation generation model Conf-GEM;

[0059] In the figure: Two-dimensional molecules; Output conformation; +, tensor addition operation; z, tensor conforming to Gaussian distribution; 8-1, prior model; 8-2, decoder model; 9-1, atomic, chemical bond, bond angle and dihedral angle features and generated conformation.

[0060] Figure 4 Internal logic of the conformation generation model Conf-GEM

[0061] In the figure: Δ, updated coordinates; Output conformation, 11-1, atoms, chemical bonds, bond angles, dihedral angle features and conformations output by the sub-model; 11-2, atoms, chemical bonds, bond angles, dihedral angle features and conformations after the first update; 11-3, atoms, chemical bonds, bond angles, dihedral angle features and conformations after the second update; 11-4, atoms, chemical bonds, bond angles, dihedral angle features and conformations after the third update; 12-1, update bond lengths; 12-2, update bond angles; 12-3, update dihedral angles; 13, update graphs. DETAILED DESCRIPTION

[0062] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0063] The scientific basis of the present invention can be briefly expressed as follows: a framework for generating stable conformations of molecules based on pre-trained and geometrically enhanced graph networks. The representation of molecules is based on graph convolutional neural networks. The convolution of the graph consists of two steps: (1) using a function to convert the information of the current node into Information about its surrounding neighbor nodes And the characteristics of the edges connecting them as parameters to obtain the information that needs to be transmitted at the current time step (2) Use this information to update the node information at the current time step through a function to obtain the information used as input for the next iteration For each atom in the molecule, the initial input feature x v After several iterations, the final hidden state h is obtained v . This hidden state can be used for downstream tasks at the atomic level, such as atomic coordinates. Based on this representation model, a variational autoencoder is used for de novo generation of molecular conformations, which is a commonly used generative framework for learning latent representations of input data and generating new data samples. Its principle is based on two key concepts: encoder and decoder. The encoder maps the input data to a probability distribution in the latent space, which is usually assumed to be a Gaussian distribution. The decoder reconstructs the input data from samples in the latent space. During training, the variational autoencoder attempts to minimize the reconstruction error between the input data and the decoded reconstructed data, while maximizing the KL divergence between the samples in the latent space and the standard normal distribution to ensure that the learned latent representation has good continuity and sampling properties. In this way, the variational autoencoder is able to learn a good representation of the data and can generate new data samples with diversity.

[0064] Example 1

[0065] In order to test the effect of introducing dihedral information on improving the ability to distinguish molecular conformations and intuitively observing the molecular information learned by the representation model. We used the t-SNE algorithm to map the feature vectors output by the D-GeoGNN model to a two-dimensional space for visualization. In addition, we calculated the Davies Bouldin index to measure the degree of separation of clusters. The smaller the Davies Bouldin index, the better the cluster separation. We randomly selected 1,000 molecules from ZINC as the test set for this part of the experiment. For each molecule, we first used the force field function of RDKit to generate a relatively reasonable conformation, and then added a uniformly distributed noise of 0 to 0.5 to the coordinates of each atom in the molecule as a control conformation (average RMSD = 0.238). The D-GeoGNN with the introduction of dihedral information can distinguish the two conformations well, and compared with the GeoGNN model that only uses bond length and bond angle, the Davies Bouldin index of the two clusters is reduced from 1.566 to 1.404.

[0066] Example 2

[0067] GEOM-QM9 is used as the standard dataset. The training set, validation set and test set of this dataset contain 200,000, 2,500 and 22,408 molecule-conformation pairs respectively. In the training set and validation set, each molecule contains 5 conformations, and the test set contains a total of 150 molecules.

[0068] Assume that there are N molecules x in the test set x conformations, generating 2N x The model was evaluated using coverage COV and matching MAT.

[0069]

[0070] Among them, S r and S g Represent the true conformation and the generated conformation respectively. RMSD calculation is realized by RDKit software package. The average coverage and matching degree are 77.67% and 0.2761 angstroms respectively.

[0071] The above contents are further detailed descriptions of the present invention in combination with specific implementation methods, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A framework for generating stable molecular conformations using pre-trained and geometrically enhanced graph networks, characterized in that: Including three-dimensional graph neural network D-GeoGNN, pre-training task D-SSLs and conformation generation model Conf-GEM; The three-dimensional graph neural network D-GeoGNN is composed of three isomorphic graph neural networks connected in series, which accept three graph inputs respectively: atom-chemical bond graph G = (V, E), chemical bond-bond angle graph H = (E, A), and bond angle-dihedral angle graph I = (A, D); where V is the set of atoms, E is the set of chemical bonds, A is the set of bond angles, and D is the set of dihedral angles; in graph G, atoms u∈H are regarded as nodes of the graph, and chemical bonds (u,v)∈E connecting atoms u and v are regarded as edges of the graph; while the nodes of graph H are chemical bonds (u,v)∈E, and the bond angles (u,v,w)∈A connecting chemical bonds (u,v) and (v,w) are regarded as edges of graph H; for graph I, bond angles (u,v,w)∈A are regarded as nodes of the graph, and dihedral angles (u,v,w,x)∈D composed of bond angles (u,v,w) and (v,w,x) are regarded as edges; 4 initial features: using x u to represent the initial characteristics of atom u, x uv represents the initial characteristics of the chemical bond (u,v), x uvw is the initial characteristic of the bond angle (u, v, w), x uvwx is the initial feature of the dihedral angle (u, v, w, x); the three images and the four initial features are used as the input of the three-dimensional graph neural network D-GeoGNN; Pre-training task D-SSLs: Use a large-scale unlabeled dataset to pre-train D-GeoGNN at the three-dimensional geometry level. The molecular conformation is calculated by the MMFF94 force field; specifically, it includes predicting bond lengths; predicting bond angles; predicting dihedral angles; predicting the distance between any two atoms; and predicting the loss functions for bond lengths, bond angles, dihedral angles, and the distance between any two atoms. During the pre-training process of the three-dimensional graph neural network D-GeoGNN, the total loss is the sum of the losses of the four geometric pre-training tasks. The pre-trained three-dimensional graph neural network D-GeoGNN is called D-GEM; For the prior model, encoder model and decoder model in the conformation generation model Conf-GEM, they are all composed of three D-GEMs in series. The divide-and-conquer strategy for each model is as follows: (1) The first GEM accepts an initial molecular conformation Generating conformations Then the bond length l is calculated based on this conformation, that is: Then update the bond length l to the "atom-chemical bond graph" G as the input of the next model. It should be noted that for the prior model is a uniformly distributed coordinate from -1 to 1, the encoder model is the true conformation, and the decoder model is the final output coordinate of the prior model; (2) The second GEM accepts the previous step Generating conformations The bond angle θ is then calculated based on this conformation, namely: Then update the bond angle θ to the "chemical bond-bond angle diagram" H as the input of the next model; (3) The third GEM accepts the previous step Generating conformations Then the dihedral angle ω is calculated based on this conformation, namely: is the final output conformation of the model. Finally, the loss function obtained in the pre-training task D-SSLs is used to calculate the losses of bond length, bond angle, and dihedral angle, and the final loss is obtained by summing them up.

2. A molecular stable conformation generation framework for a pre-trained and geometrically enhanced graph network according to claim 1, characterized in that: The three-dimensional graph neural network D-GeoGNN is also an information transmission network. During the training process, the hidden state h of the chemical bond uv The hidden state h of the bond angle uvw They are used for communication between graphs G and H, and between graphs H and I respectively. The steps at the tth iteration are as follows: First, the information of neighboring bond angles and dihedral angles connecting neighboring bond angles is aggregated in the bond angle-dihedral angle graph I to update the new hidden state of the bond angle in, represents the atoms adjacent to atom u, Represents the neighboring bond angles of the bond angle (u,v,w); In the chemical bond-angle diagram I, the function Agg I Aggregate the information to be transmitted; After that, through the function Comb. I renew for Then in Figure H, the neighbor nodes of the chemical bond and the bond angle information of the neighbor nodes are aggregated to obtain the hidden state of the chemical bond in, Represents the neighboring chemical bonds of a chemical bond (u, v); through the function Agg H Aggregate information Then through the function Comb. H renew for Next, in graph G, like most other GNNs, the information of the neighbor nodes of the atoms and the chemical bonds connecting the atoms is aggregated to obtain the current atomic hidden state Finally, the READOUT function is used to pool the atomic hidden state set obtained in the last iteration to obtain the eigenvector h of the molecule. G : Where K is the number of iterations, h G is used for prediction tasks at the molecular level, while It is used for prediction tasks at the atomic level.

3. The molecular stable conformation generation framework of a pre-trained and geometrically enhanced graph network according to claim 1, characterized in that: The first step is to use a large-scale unlabeled dataset to pre-train D-GeoGNN at the 3D geometry level. For the local prediction tasks of predicting bond length, predicting bond angle, and predicting dihedral angle, a training strategy of cover learning is adopted, which specifically includes the following steps: First, 15% of the atoms in the molecule are randomly selected. For each selected atom, the first-order neighborhood of the atom is extracted, including the adjacent atoms and chemical bonds as well as the bond angles and dihedral angles formed by the selected atom. Secondly, mask out the features of selected atoms, chemical bonds, bond angles, and dihedral angles; Finally, the hidden states of the masked chemical bonds, bond angles, and dihedral angles extracted by D-GeoGNN after the last iteration are used to predict the corresponding bond lengths, bond angles, and dihedral angles, respectively. The loss functions are: Where L length , L angle With L dihedral represents the loss function in the prediction task of bond length, bond angle and dihedral angle respectively; ε, A and D represent the chemical bond set, bond angle set and dihedral angle set respectively; f length (·),f angle (·) and f dihedral (·) represent the predicted values ​​of the neural network for bond length, bond angle and dihedral angle respectively; and l uv ,φ uvw and θ uvw are the true values ​​of the predicted values ​​of bond length, bond angle, and dihedral angle, respectively; For the global geometric task of predicting the distance between any two atoms in a molecule, an atomic distance matrix is ​​constructed for each molecule based on its three-dimensional coordinates, and then the distances between atom-atom pairs are predicted. For a molecule, the prediction of the atomic distance matrix is ​​regarded as a multi-classification problem, and the atomic distances are projected into 30 components with equal distances. The task of predicting bond lengths is regarded as a special case of the task of predicting atomic distances. The loss function for predicting the distance between any two atoms in a molecule is:

4. The molecular stable conformation generation framework of a pre-trained and geometrically enhanced graph network according to claim 1, characterized in that: The training process of the conformation generation model Conf-GEM includes the following steps: (1) The prior model accepts a two-dimensional molecule and converts it into a three-dimensional molecule based on the MMFF94 force field as input, and then outputs atoms, chemical bonds, bond angles and dihedral angles as well as a generated conformation; (2) The encoder model accepts a real configuration and outputs the mean and variance required to construct a latent vector that conforms to the Gaussian distribution; (3) The output of the above two steps is used as the input of the decoder. The decoder model will eventually output a predicted conformation 5. A molecular stable conformation generation framework of a pre-trained and geometrically enhanced graph network according to claim 4, characterized in that: In order to make the conformations generated by the conformation generation model diverse, the present invention adds a probability sampling step in the generation process of the conformation generation model. That is, a random hidden variable z is introduced, and then conformation generation is performed under conditional probability: Among them, γ is a hyperparameter; The configuration generation model is conditioned on the latent variable z and the molecular graph; the sampling of the latent variable z conforms to the standard normal distribution.

6. A molecular stable conformation generation framework for pre-training and geometrically enhanced graph networks according to claim 5, characterized in that: The inference of the conformation generation model Conf-GEM uses a priori model and a decoder model, specifically: (1) For a given two-dimensional molecule, the representation of atoms, chemical bonds and molecular graphs and the preliminary predicted conformation are obtained through encoding of the prior model; (2) Randomly extract a latent vector z from the standard normal distribution; (3) The three representations, the conformation output by the prior model, and the latent vector are input into the decoder model to obtain the final generated conformation.

Citation Information

Patent Citations

  • Space structure-based homogeneous interaction prediction method and device in drug discovery field

    CN114898815A

  • Molecular property prediction method and system based on graph neural network and three-dimensional encoder

    CN116978483A

Cited By

  • Three-dimensional molecular structure generation method and model

    CN120998348A