Reaction-Guided Chemical Structure Generation for Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating molecular structures for machine learning are inefficient due to the combinatorial explosion of possible structures as the number of atoms increases, leading to a high ratio of unusable molecules, which are time and economically inefficient to use as training data.
Innovation Solution
A chemical structure generating device and method that includes a generator and controller to produce and update reactant and product lists based on chemical reactions, filtering out commercially unusable structures through a prohibitive reaction list, to efficiently generate high-quality training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If exhaustive generation of molecular structure candidates is performed, then the completeness of training data is improved, but the time and computational resources required increase exponentially due to combinatorial explosion
Solution Approach 1:
The patent applies preliminary action by performing retrospective analysis on existing commercial molecules to identify feasible reaction pathways and building blocks before generating new structures. This pre-computation of reaction rules and reactant libraries enables subsequent generation to focus only on chemically feasible structures, avoiding exhaustive search of all combinatorial possibilities while maintaining reliability of generated training data
Solution Approach 2:
The patent segments the molecular generation problem into discrete chemical reactions and building blocks. By decomposing molecules into reactants and products through retrosynthetic analysis, the system generates training data through controlled chemical transformations rather than exhaustive structure enumeration, significantly reducing computational time while maintaining completeness
2Adaptability or versatility
If the number of atoms in molecules is increased, then the complexity and potential utility of molecules is improved, but the ratio of usable to possible molecular structures rapidly decreases
Solution Approach 1:
The patent performs preliminary retrosynthetic analysis on commercially available molecules to establish feasible reaction pathways and valid building blocks before generation. This pre-computation ensures that subsequent generated molecules with increased atom count and complexity adhere to chemically feasible transformations, maintaining high usability ratios even as molecular complexity increases
Solution Approach 2:
The system incorporates feedback mechanisms where generated molecular structures are evaluated against chemical feasibility rules derived from retrosynthetic analysis. This feedback loop ensures that only molecules with reasonable synthetic pathways and commercial usability are included in training data, maintaining high reliability ratios even as molecular complexity and atom count increase
3Quantity of substance
If training data includes many commercially unusable molecular structures, then the volume of training data is increased, but the economic efficiency and time consumption for machine learning increases
Solution Approach 1:
The patent performs preliminary retrosynthetic analysis to establish reaction feasibility rules and valid building blocks before generating training data. This pre-computation enables the system to generate only commercially usable molecules with feasible synthetic pathways, producing sufficient training data volume while eliminating economically inefficient unusable structures
Solution Approach 2:
The patent extracts and applies chemical feasibility rules and retrosynthetic knowledge from existing commercial molecules to filter and guide the generation process. This extraction of domain knowledge enables the system to produce training data with high commercial usability ratios, increasing productivity and economic efficiency while maintaining adequate training data volume
Data Source
AI summary
A chemical structure generating device according to the present invention includes a generator and a controller. The generator produces a product list including one or more compounds, based on a reactant list including one or more compounds and a chemical reaction list. The controller applies the product list as a new reactant list to the generator, updates a database having at least one list of the reactant list and the product list, and allows the generator to produce a new product list based on the new reactant list and the chemical reaction list.