Drug Composition Generation With Transformer-RL Validity Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing drug discovery methods face challenges in generating valid molecules with desired properties due to the large search space and limitations of generative models like RNNs, leading to low validity rates and inefficiencies in drug design.
Innovation Solution
A synergistic approach using transformer-based architectures and reinforcement learning (RL) to generate molecules, employing a two-stage process that includes learning discrete string representations and optimizing vector spaces to produce molecules with desired properties like QED and pIC50.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If generative models like RNNs are used to generate molecules, then molecule generation capability is achieved, but validity rate deteriorates
Solution Approach 1:
The patent employs reinforcement learning where a reward function provides feedback to the transformer model based on molecule validity and desired properties. This feedback mechanism guides the model to generate valid molecules by adjusting generation strategies based on validation results, directly addressing the low validity rate issue of traditional generative models
Solution Approach 2:
The patent transforms the molecule generation problem into a sequence prediction task with modified parameters. By using transformer architecture with attention mechanisms and optimizing generation parameters through reinforcement learning, the system achieves both high validity and desired molecular properties, resolving the contradiction between generation capability and validity
2Productivity
If the search space is reduced using traditional methods, then screening efficiency is improved, but the number of potential drug-like molecules deteriorates
Solution Approach 1:
The patent replaces traditional mechanical high-throughput screening methods with an AI-based transformer model. This substitution enables the system to efficiently evaluate and generate molecules with desired properties directly, achieving high screening efficiency while maintaining access to the full chemical space of potential drug-like molecules
3Manufacturing precision
If reinforcement learning is used to optimize molecule generation, then desired properties are improved, but convergence toward invalid molecules deteriorates
Solution Approach 1:
The patent designs a reward function that provides dual feedback: one component rewards desired molecular properties (QED, pIC50) while another component penalizes invalid molecules. This balanced feedback mechanism ensures the reinforcement learning process converges toward valid molecules with optimized properties, preventing convergence to invalid structures
Solution Approach 2:
The patent pre-trains the transformer model on valid SMILES strings before applying reinforcement learning. This preliminary action establishes a foundation of valid molecular structures, ensuring that subsequent RL optimization starts from a valid base and maintains validity while optimizing desired properties
Data Source
AI summary
A system and method of designing a drug by at least one processor may include obtaining a molecule string data element, representing ad-hoc structure of a molecule. The molecule string may include at least one token, representing (i) indication of a beginning of the molecule string, (ii) one or more components of the molecule, and/or (iii) relation between components of the molecule. The at least one processor may apply an embedding algorithm on the molecule string, to obtain an embedding vector, representing the ad-hoc structure of the molecule in an embedding space, and apply a pretrained transformer-based decoder model on the embedding vector, to select a subsequent token from a predetermined set of tokens; append the predicted token to the molecule string; and, following identification of occurrence of an end condition, append a token representing end of the molecule string, to determine composition of the drug.