Molecular Graph Inference via Tree Decomposition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for inferring molecular structures using chemical formulas represented as strings or graphs face challenges, such as generating invalid molecules and difficulty in inferring structures according to specific purposes, particularly in the context of drug development and material science.
Innovation Solution
The use of tree decomposition with site information, where a molecular graph is converted into a tree representation that includes connection information between nodes, allowing for reversible conversion between tree and graph representations, enabling the generation of valid molecular structures through reinforcement learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If chemical formula is converted into character string for reinforcement learning, then the method is simple to implement, but invalid molecules are often generated
Solution Approach 1:
The patent introduces an intermediate representation (tree structure with site information) between the character string input and the final molecular graph output. This intermediate structure serves as a mediator that guides the generation process to ensure chemical validity while maintaining the simplicity of character string manipulation for the reinforcement learning model.
Solution Approach 2:
The molecular structure generation is segmented into distinct components: the tree structure represents the molecular skeleton, site information represents connection points, and the reinforcement learning model generates actions to build the structure. This segmentation allows each component to be optimized independently while ensuring overall validity.
2Reliability
If chemical formula is converted into graph for reinforcement learning, then valid molecules are not generated, but it is difficult to infer molecular structures according to purposes
Solution Approach 1:
The patent employs a dynamic generation process where the reinforcement learning model sequentially builds the molecular structure by making decisions at each step based on the current state. This dynamic approach allows the model to adapt to different purposes by adjusting the reward function and exploration strategy while maintaining chemical validity through the structured tree representation.
Solution Approach 2:
The patent changes the parameter representation from direct graph edges to tree-based hierarchical structures with site information. This parameter transformation enables the model to maintain validity constraints inherent in the tree structure while gaining flexibility in exploring different molecular configurations through the reinforcement learning policy.
3Reliability
If tree decomposition with site information is used, then valid molecular structures are generated, but the method complexity increases
Solution Approach 1:
The complex task of generating valid molecular structures is segmented into manageable components: tree decomposition, site information assignment, and sequential bond formation. Each segment handles a specific aspect of the problem, making the overall complex task tractable while ensuring validity at each step.
Solution Approach 2:
The tree structure with site information serves as an intermediary representation that simplifies the complexity management. Instead of directly manipulating complex graph structures, the model works with the simpler tree representation, which inherently enforces validity constraints while being easier to process sequentially.
Data Source
AI summary
An inferring device comprises one or more memories and one or more processors. The one or more processors execute decision of an action based on a tree representation including a node and an edge of a molecular graph, and a trained model trained through reinforcement learning, and execute generation of a state including information on the molecular graph based on the action, wherein the edge has connection information on the nodes.