Molecular MC Tree Search

Synthesizing molecules and doing experimental analyses can take weeks or even months of labor-intensive research. Generative AI offers a unique opportunity to speed up this process. By efficiently proposing candidate molecules, Generative AI could help experimentalists narrow down the chemical search space significantly—saving countless work-hours.

In the current state of the art, most approaches use either a Diffusion/Flow-based model on a graph-based representation or frameworks based on LLMs, for a token-based representation of molecules (SMILES). These approaches are demonstrably effective, but are very dissimilar from the chemical intuition used by chemists in the same molecular design process. As such, these methods can seem like a black box, which hinders the furtherance of GenAI in this context.

The AlphaMol project seeks to address this issue by using a completely different approach to navigating the vast chemical space. Motivated by the success of AlphaZero & Monte Carlo Tree Search (MCTS) in Reinforcement Learning problems, you will design and develop a Reinforcement Learning agent, similar to AlphaZero, on a new domain: molecule generation. In the process, you will create your own algorithm for effectively navigating the chemical space, by autoregressively ‘growing’ a molecule from scratch. This method is more interpretable than the ‘black box’ methods and is much more similar to the thought process of a chemist when designing new molecules.

Supervision:

• Daily supervision: Alec Noppe (HIMS)
• Examinor: Bernd Ensing (HIMS)