Compound property analysis method, model training method, apparatuses, and storage medium
View Patent ↗A compound property analysis method is provided. The method includes obtaining, according to a molecular structure of a compound, a feature vector of the compound, the feature vector including a node vector of each node and an edge vector of each edge, processing the feature vector by using a feature map extraction model branch to obtain a graph representation vector, and processing the graph representation vector by using a classification model branch to obtain a property of the compound. Thus, in the process of compound property analysis, the graph representation vector that can accurately represent a feature of the compound is obtained based on a graph data structure of the compound, and a classification property of the compound may be obtained based on the graph representation vector, thereby improving the accuracy of determining the classification property of the compound. Apparatus and non-transitory computer-readable storage medium counterpart embodiments are also provided.
1 . A compound property analysis method performed by a computer device, the method comprising:
obtaining, according to a molecular structure of a compound, a feature vector of the compound, the feature vector comprising a node vector of each node and an edge vector of each edge, the nodes respectively corresponding to atoms in the molecular structure, and the edges respectively corresponding to chemical bonds in the molecular structure;
processing the feature vector, by using the computer device through a compound property analysis model that includes a feature map extraction model branch configured to perform message passing across a plurality of layers, to generate a graph representation vector; and
processing the graph representation vector by using a classification model branch in the compound property analysis model, the classification model branch being configured to determine a predicted property type of the compound based on the graph representation vector,
the compound property analysis model being a machine learning (ML) model trained according to a molecular structure of a compound sample and a corresponding property type of the compound sample.
2 . The method according to claim 1 , wherein the processing the feature vector further comprises:
performing, through n message passing layers in the feature map extraction model branch, n layers of message passing on state information of each node according to the edge vector of each edge, n being an integer greater than or equal to 2; and
converting a result of an n th message passing layer into the graph representation vector with a fixed length by using a conversion function in the feature map extraction model branch.
3 . The method according to claim 2 , wherein the performing further comprises:
for an i th layer in the n message passing layers,
when i=1, processing, through a first message passing layer in the feature map extraction model branch, initial state information of each node according to the edge vector of each edge, to obtain the state information corresponding to each node outputted from the first message passing layer; and
when 2≤i≤n, processing, through an i th message passing layer in the feature map extraction model branch, the state information corresponding to each node outputted from an (i−1) th message passing layer according to the edge vector of each edge, to obtain state information corresponding to each node outputted from the i th message passing layer.
4 . The method according to claim 3 , wherein a message passing layer of the plurality of layers is configured to perform message aggregation and state fusion, comprising:
obtaining message information of a first node at a moment t+1 by processing the state information of the first node at a moment t, the state information of a second node at the moment t, and an edge between the first node and the second node by using a message aggregation function; and
obtaining state information of the first node at the moment t+1 by processing the state information of the first node at the moment t and the message information of the first node at the moment t+1 by using a state fusion function.
5 . The method according to claim 2 , wherein the converting further comprises:
processing the result of the n th message passing layer by using a sum function in the feature map extraction model branch, to obtain the graph representation vector with the fixed length.
6 . The method according to claim 1 , wherein before the processing the feature vector by using the feature map extraction model branch, the method further comprises:
obtaining the property type corresponding to the compound; and
selecting the corresponding compound property analysis model according to the property type.
7 . The method according to claim 6 , wherein the property type comprises at least one of a chemical property, a physical property, and a biological property.
8 . A model training method performed by a computer device, the method comprising:
obtaining, according to a molecular structure of a compound sample, a feature vector sample of the compound sample, the feature vector sample comprising a node vector of each node and an edge vector of each edge, the nodes respectively corresponding to atoms in the molecular structure, and the edges respectively corresponding to chemical bonds in the molecular structure;
processing the feature vector sample by using the computer device through a compound property analysis model that includes a feature map extraction model branch configured to perform message passing across a plurality of layers, to generate a graph representation vector sample;
processing the graph representation vector sample by using a classification model branch in the compound property analysis model, the classification model branch being configured to determine a predicted property type of the compound sample based on the graph representation vector sample; and
adjusting parameters in the feature map extraction model branch and the classification model branch according to a property of the compound sample and the predicted property type.
9 . The method according to claim 8 , wherein the adjusting parameters further comprises:
inputting the property type of the compound sample and the classification predicted property type to a loss function to obtain a loss function value; and
adjusting the parameters in the feature map extraction model branch and the classification model branch according to the loss function value.
10 . The method according to claim 9 , wherein the adjusting the parameters in the feature map extraction model branch and the classification model branch according to the loss function value further comprises:
adjusting, in response to the loss function value being greater than a preset threshold, the parameters in the feature map extraction model branch and the classification model branch according to the loss function value.
11 . A compound property analysis apparatus, the apparatus comprising:
circuitry configured to
obtain, according to a molecular structure of a compound, a feature vector of the compound, the feature vector comprising a node vector of each node and an edge vector of each edge, the nodes respectively corresponding to atoms in the molecular structure, and the edges respectively corresponding to chemical bonds in the molecular structure;
process the feature vector, by using a compound property analysis model that includes a feature map extraction model branch configured to perform message passing across a plurality of layers, to generate a graph representation vector; and
process the graph representation vector by using a classification model branch in the compound property analysis model, the classification model branch being configured to determine a predicted property type of the compound based on the graph representation vector,
the compound property analysis model being a machine learning (ML) model trained according to a molecular structure of a compound sample and a corresponding property type of the compound sample.
12 . The apparatus according to claim 11 , wherein the circuitry is further configured to:
perform, through n message passing layers in the feature map extraction model branch, n layers of message passing on state information of each node according to the edge vector of each edge, n being an integer greater than or equal to 2; and
convert a result of an n th message passing layer into the graph representation vector with a fixed length by using a conversion function in the feature map extraction model branch.
13 . The apparatus according to claim 12 , wherein the circuitry is further configured to, for an i th layer in the n message passing layers,
when i=1, process, through a first message passing layer in the feature map extraction model branch, initial state information of each node according to the edge vector of each edge, to obtain the state information corresponding to each node outputted from the first message passing layer; and
when 2≤i≤n, process, through an i th message passing layer in the feature map extraction model branch, state information corresponding to each node outputted from an (i−1) th message passing layer according to the edge vector of each edge, to obtain the state information corresponding to each node outputted from the i th message passing layer.
14 . The apparatus according to claim 12 , wherein a message passing layer of the plurality of layers is configured to perform message aggregation and state fusion, comprising:
obtaining message information of a first node at a moment t+1 by processing the state information of the first node at a moment t, the state information of a second node at the moment t, and an edge between the first node and the second node by using a message aggregation function; and
obtaining state information of the first node at the moment t+1 by processing the state information of the first node at the moment t and the message information of the first node at the moment t+1 using a state fusion function.
15 . The apparatus according to claim 12 , wherein the circuitry is further configured to process the result of the n th message passing layer by using a sum function in the feature map extraction model branch, to obtain the graph representation vector with the fixed length.
16 . The apparatus according to claim 11 , wherein the circuitry is further configured to:
obtain the property type corresponding to the compound; and
select the corresponding compound property analysis model according to the property type.
17 . The apparatus according to claim 16 , wherein the property type obtained by the circuitry comprises at least one of a chemical property, a physical property, and a biological property.
18 . A computer device, comprising processing circuitry and a memory, the memory storing at least one instruction, at least one program, a code set, or an instruction set,
the at least one instruction, the at least one program, the code set, or the instruction set being loaded and executed by the processing circuitry to implement the compound property analysis method according to claim 1 .
19 . A non-transitory computer-readable storage medium, the storage medium storing at least one instruction, at least one program, a code set, or an instruction set,
the at least one instruction, the at least one program, the code set, or the instruction set being loaded and executed by processing circuitry to implement the compound property analysis method according to claim 1 .
20 . A non-transitory computer-readable storage medium, the storage medium storing at least one instruction, at least one program, a code set, or an instruction set,
the at least one instruction, the at least one program, the code set, or the instruction set being loaded and executed by processing circuitry to implement the compound property analysis method according to claim 8 .