Beyond generative models: superfast traversal, optimization, novelty, exploration and discovery (STONED) algorithm...

Area of Science:

• Computational chemistry
• Machine learning in chemistry
• Molecular modeling

Background:

• Inverse design aims to generate molecules with specific properties.
• Deep generative models offer potential for molecular property optimization.
• Current deep learning approaches require substantial data and training time.

Purpose of the Study:

• To introduce STONED, a simple and efficient algorithm for chemical space exploration and interpolation.
• To bypass the data and time constraints of deep generative models in molecular inverse design.
• To provide a comparable or superior alternative to existing generative models for molecular design tasks.

Main Methods:

• Utilizes string modifications within the SELFIES molecular representation.
• Performs interpolation and exploration in chemical space without deep learning model training.
• Applies the algorithm to benchmarks typically used for generative models.

Main Results:

• Achieves non-trivial performance on generative model benchmarks without prior training.
• Demonstrates effectiveness in high-throughput virtual screening for drug and photovoltaic design.
• Enables property and structure-based interpolation within the chemical space.

Conclusions:

• STONED offers a data-efficient and time-saving alternative for molecular inverse design.
• The algorithm facilitates exploration and interpolation in chemical space for various applications.
• Results are expected to aid in developing advanced inverse design tools and benchmarking generative models.