Molecular Generation Method Based on Subgraph-Variational Autoencoder Structure
Through the molecular sub-map-autocoded structure method based on the VAE framework, the problems of low efficiency and unstable chemical characteristics of traditional drug molecular generation methods are solved, and the generation of novel drug molecules with stable chemical properties and easy synthesis are achieved, which improves drug design and Development efficiency.
Patent Information
- Application Number
- CN202011170128.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-10-28
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2040-10-28
AI Technical Summary
Traditional drug molecule generation methods are inefficient and rely on chemists’ domain knowledge. SMILES indicates that it is not possible to effectively characterize molecular similarity and chemical structure, resulting in the chemical properties of the generated molecules being unstable and difficult to synthesize.
Using the molecular sub-graph-autocoded structure method based on the VAE framework, frequent sub-graphs are mined through the gSpan algorithm, sub-graphs that meet the conditions are selected, and chemical structure is maintained using multiscale VAE (multiscale VAE) to generate new molecular structures. .
It improves the efficiency of design and development of new drugs, generates molecular structures with stable chemical properties and easy to synthesize, reduces the dependence on domain knowledge, and shortens the R&D cycle.
Smart Images

Figure CN112397157B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of artificial intelligence drug research and development, and specifically relates to the field of drug research and development based on molecular subgraph-autoencoder structure molecular generation and optimization. Background Art
[0002] In the field of drug molecule generation and optimization, due to the large number of molecular permutations and combinations, the molecular space formed is too large, and the space for screening by manually customized rules is limited. It is crucial to use existing molecular data to automatically find the molecules or molecular spaces that are most relevant to the target molecules. The traditional method is usually to analyze the chemical properties of known molecules and target molecules and manually design new molecular structures. This method is very inefficient, and the process is very dependent on the domain knowledge and experience of chemists. Therefore, some people have proposed a method based on machine learning to automatically generate new molecules; the commonly used method is to represent molecules as SMILES (Simplified Molecular Linear Input Specification) strings, and then train graph neural network models or RNN models to automatically generate new molecular structures.
[0003] At present, the common model for automatically generating new molecules mainly represents the molecules in SMILES format, then uses the commonly used time series model or natural language processing model to generate the SMILES sequence, and finally converts it into the corresponding molecule. However, the SMILES representation of molecules cannot well describe the similarity between molecules. Two molecules with highly similar chemical structures may be encoded into two strings of SMILES characters that are very different. In addition, the SMILES sequence generation model cannot well maintain the effective chemical structure.
[0004] In the latest research, some people proposed a graph-based molecular generation model, such as Figure 1 As shown in the figure, the molecule is represented as a graph, and then a new graph is generated using the graph as input data. The new graph is the newly generated molecule. This method is usually performed continuously in units of atoms, and finally a complete molecular graph is formed. Although the molecules generated by this graph-based method have better expressiveness than SMILES in some chemical properties, since the entire graph is gradually generated based on a single atom, some global chemical structures are not considered during the generation process. Therefore, the molecules obtained usually cannot meet actual needs in some chemical properties. Summary of the invention
[0005] Purpose of the invention: To maintain and transmit relevant chemical structures and properties through subgraphs, and then discover more potential molecular structures through latent variable space, and then automatically combine them to generate diversified molecules.
[0006] Technical solution: Based on the VAE framework, the present invention proposes a molecular generation and optimization method based on a molecular subgraph-self-encoding structure. The specific process is as follows Figure 1 As shown, the main process is as follows:
[0007] First, the frequent subgraphs of all molecules in the training set (without any special functions) are mined based on the gSpan algorithm. Figure 1 -b), and then screen the frequent subgraphs based on statistical counting thresholds and pharmaceutical properties to obtain the total set of molecular subgraphs that meet the conditions S*( Figure 1 -c);
[0008] Secondly, each molecule G in the training set is decomposed to obtain the corresponding molecular subgraph set S( Figure 1 -d);
[0009] The molecule set S is input into the encoder of multiscale VAE again for encoding, and the corresponding latent space vector z ( Figure 1 -3);
[0010] Finally, z is input into the decoder of multiscale VAE for decoding to obtain the reconstructed molecule G′( Figure 1 -4);
[0011] Through the above process, the encoder and decoder are jointly trained. After the training is completed, the encoder can be used to map the input molecules to the latent space, and then the required molecules can be generated by decoding the latent variables. Figure 2 As shown, it includes encoding and decoding of the atomic layer and encoding and decoding of the sublayer respectively.
[0012] Technical effect: The present invention innovatively proposes a molecule generation method based on a molecular subgraph-self-encoding structure, which overcomes the shortcomings of the traditional atom-by-atom generation method such as the lack of chemical properties and difficulty in synthesis. Through the application of substructures, certain chemical properties of molecules are guaranteed, thereby generating molecular structures with stable chemical properties and easy synthesis, which greatly facilitates the discovery and design of new drug small molecules, reduces the dependence of relevant staff on domain knowledge, improves the efficiency of new drug design and development, shortens the research and development cycle, and has high practical value. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Figure 1 It is a flow chart of the subgraph-self-encoding molecule generation model of the present invention;
[0014] Figure 2 It is the multiscale VAE encoding and decoding of the present invention. DETAILED DESCRIPTION
[0015] Principle of the invention: In order to generate new drug molecules with certain or several specified chemical properties, and which can be synthesized in practice, the present invention proposes a molecule generation method based on a subgraph-variational autoencoder structure, that is, firstly obtain all frequent subgraphs in the molecules of the training set through the gSpan algorithm (an algorithm for generating frequent subgraphs based on graph pattern search), screen the subgraphs that meet the conditions as the total set of molecular subgraphs, and then divide the molecules into rings or frequent subgraphs, and then use the multiscale variational autoencoder model (multiscale VAE (variational autoencoder)) to reconstruct the original molecular graph. We maintain and transmit related chemical structures and properties through subgraphs, and then discover more potential molecular structures through the hidden variable space, and then automatically combine them to generate diversified molecules.
[0016] The specific process of generating new drug molecules based on the molecular subgraph-variational autoencoder molecule generation method proposed in this paper is as follows:
[0017] Generate a total set of molecular subgraphs
[0018] (1) Set the frequency k and use the gSpan algorithm to generate frequent subgraphs from the molecules in the training set;
[0019] (2) Using the random forest algorithm to train a specific chemical drug property detector R i ;
[0020] (3) Filter frequent subgraphs: Use the above detector R i Detect the frequent subgraphs obtained in (1) and select the frequent subgraphs larger than a certain threshold as the total set of molecular subgraphs.
[0021] 1. Molecules are decomposed into sub-atlases
[0022] First, the correlation function of RDKit (an open source chemical toolkit) is used to find the rings in the molecule. Then, based on the total set of molecular subgraphs and the size of the subgraphs, the original molecule is split in sequence starting from the larger subgraphs according to the post-term search and matching principle.
[0023] 2. Training multiscale variational autoencoder multiscale VAE
[0024] Here we use the feedback neural network GRU to represent the node propagation in the graph network, and use the multi-layer fully connected network MLP to represent the aggregation of atoms to subgraphs.
[0025] (1) Atomic layer molecular diagram: u and v represent atoms, and {e(au)} and {e(buv)} represent the set of all vertices with atoms and the set of all edges connecting atoms, respectively.
[0026]
[0027] (2) Sub-layer molecular graph: Si represents the i-th sub-graph, fSi represents the vertex represented by the sub-graph Si, and dij represents the connection from Si to Sj.
[0028]
[0029]
[0030] (3) Hidden variables:
[0031]
[0032] (4) Predict the next subgraph St to be added:
[0033]
[0034] (5) Predict how St is added to Sk:
[0035] p M =softmax(h M ·z G )
[0036]
[0037] Among them, 1-3 is the multiscale VAE encoding process, 4-5 is the multiscale VAE decoding process, and the loss function is
[0038]
[0039] Where Q(z|G) is a normal distribution with similar mean and variance to G, and P(G|z) represents the loss of latent variable decoding. Note that in the decoding process 5, we have the original molecule graph when reconstructing the original molecule, so there is real label information during the training process.
[0040] 4. Generate new molecules
[0041] Random sampling from the latent space and multiscale VAE-based decoder to generate new molecules;
[0042] The above examples further illustrate the content of the present invention, but should not be construed as limiting the present invention. Without departing from the spirit and substance of the present invention, modifications or replacements made to the method, steps or conditions of the present invention all belong to the scope of the present invention. Unless otherwise specified, the technical means used in the examples are conventional means well known to those skilled in the art.
Claims
1. A molecule generation method based on a subgraph-variational autoencoder structure, characterized in that: Step 1: Mining the frequent subgraphs of all molecules in the training set, and then screening the frequent subgraphs by counting thresholds and pharmaceutical properties based on statistics to obtain the total set of molecular subgraphs that meet the conditions S * ; Among them, the total set of generated subgraphs S * method: (1) Set the frequency k and use the gSpan algorithm to generate frequent subgraphs from the molecules in the training set; (2) Using the random forest algorithm to train a specific chemical drug property detector R i ; (3) Filter frequent subgraphs: Use the above detector R i Detect the frequent subgraphs obtained in (1) and select the frequent subgraphs larger than a certain threshold as the total molecular subgraph set S * ; Step 2: Perform graph decomposition on each molecule G in the training set to obtain the corresponding molecular subgraph set S; The molecular subgraph set S is generated by: (1) Use the correlation function provided by the chemical toolkit to find the ring in the molecule; (2) Based on the total set of molecular subgraphs S * , according to the size of the subgraph and the principle of post-term search and matching, the original molecule is divided in sequence from the larger subgraph to obtain the set S of the smallest units with stable chemical properties; Step 3: Input the molecule set S into the variational autoencoder for encoding to obtain the corresponding latent space vector Z; The training method of the variational autoencoder is as follows: (i) Atomic layer molecular diagram: u, v represent atoms, {e(a u )},{e(b uv )} respectively represent the set with all atoms as vertices and the set with all atoms as edges; (ii) Sublayer molecular graph: S i represents the i-th subgraph, represents the i-th subgraph S i Represents the vertex, d ij Indicates S i To S j connection; (iii) Hidden variables: (iv) Predict the next subgraph S to be added t : (v) Prediction S t Add to S k The way: p M =softmax(h M ·z G ) M tk ={(u j ,v j )||u j ∈S k ,v j ∈S t } Among them, ⅰ-ⅲ is the encoding process, iv-v is the decoding process, and the loss function is Among them, Q(z|G) is a normal distribution with similar mean and variance to G, P(G|z) represents the loss when decoding the latent variable; in the decoding process v, when the original molecule is reconstructed, there is the original molecule graph; Step 4: Input Z into the variational autodecoder for decoding to obtain the reconstructed molecule G′.
Citation Information
Patent Citations
Drug molecule processing method and device based on artificial intelligence and storage medium
CN110634539A
Drug molecule generation method based on regularization variation automatic encoder
CN110970099A
Intelligent molecular design method based on auto-encoder and third-order graph convolution
CN111428848A
Molecule generation method and device, computer readable storage medium and terminal equipment
CN111508568A
Molecule generation method and computing device
CN111816265A
