Reproduce “Active Learning on Synthons for Molecular Design” on OpenResearch

About this paper

Tom George Grigg, Mason Burlage, Oliver Brook Scott, Adam Taouil, Dominique Sydow, Liam Wilbraham

Recursion developed Scalable Active Learning via Synthon Acquisition (SALSA), an algorithm that efficiently navigates non-enumerable chemical spaces up to trillions of molecules for drug discovery by factoring active learning over molecular fragments. This method demonstrates superior multi-parameter objective scores and scaffold diversity compared to leading generative approaches, while inherently ensuring synthesizability.

Key insights

• Factoring the active learning problem from full molecules to individual synthons drastically reduces combinatorial complexity, enabling exploration of chemical spaces up to trillions of compounds in a practical timeframe.
• Integrating reaction-based synthon sets provides an inherent guarantee of synthesizability for all proposed molecules, addressing a major challenge faced by many de novo generative approaches.
• Explicitly optimizing within a defined, synthesizable chemical space allows for greater control, yields higher diversity of molecular scaffolds, and offers a pragmatic alternative to generative models for complex drug design tasks.

Summary by alphaXiv.