Molecule generation using transformers and policy gradient reinforcement learning.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• Generating novel molecules is challenging due to vast chemical space, often relying on expert intuition.
• Deep learning models are increasingly used to accelerate molecule generation and identify potential drug candidates.

Purpose of the Study:

• To introduce Taiga, a transformer-based architecture for generating molecules with specific desired properties.
• To leverage a two-stage approach combining language modeling and reinforcement learning for molecular optimization.

Main Methods:

• Taiga treats molecule generation as a language modeling task using SMILES strings to predict the next token.
• Reinforcement learning is applied in the second stage to optimize molecular properties, such as the Quantitative Estimate of Drug-likeness (QED).

Main Results:

• Taiga demonstrates performance comparable to or exceeding state-of-the-art baselines in molecule optimization.
• Improvements in QED scores ranged from 2% to over 20% across various datasets, including lead and random molecules.
• The two-stage approach significantly enhances the generation of molecules with higher biological property scores compared to models without reinforcement learning.

Conclusions:

• Taiga effectively generates novel molecules with optimized properties, showcasing the power of transformer architectures in cheminformatics.
• The combination of language modeling and reinforcement learning offers a robust strategy for accelerating drug discovery and molecular design.