Parallel tempered genetic algorithm guided by deep neural networks for inverse molecular design.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Materials Science

Background:

• Inverse molecular design aims to generate molecules with desired properties, often treated as an optimization problem.
• Population-based metaheuristic algorithms like evolutionary algorithms are common but limited by expensive property evaluations.
• High-dimensional optimization in natural sciences necessitates efficient computational methods.

Purpose of the Study:

• To introduce JANUS, a novel genetic algorithm designed to improve inverse molecular design.
• To reduce the number of expensive property evaluations required in molecular optimization.
• To enhance molecular sampling and property prediction using deep learning and active learning.

Main Methods:

• JANUS employs a parallel tempering-inspired approach with two populations for exploration and exploitation.
• A deep neural network is integrated for approximating molecular properties.
• Active learning is utilized for efficient molecular sampling, alongside SELFIES and STONED for structure generation.

Main Results:

• JANUS outperforms existing generative models on standard inverse molecular design benchmarks.
• The algorithm achieves state-of-the-art target metrics across multiple evaluation tasks.
• A significant portion of generated molecules were found to be synthetically infeasible.

Conclusions:

• JANUS represents a significant advancement in inverse molecular design, offering improved efficiency and performance.
• The study underscores the critical importance of incorporating synthesizability into molecular generation models.
• Future work should focus on developing generative models that inherently consider synthetic accessibility.