Reinforcement Learning for Bioretrosynthesis

Metabolic engineering aims to produce chemicals of interest
from
living organisms, to advance toward greener chemistry. Despite efforts,
the research and development process is still long and costly, and
efficient computational design tools are required to explore the chemical
biosynthetic space. Here, we propose to explore the bioretrosynthesis
space using an artificial intelligence based approach relying on the
Monte Carlo Tree Search reinforcement learning method, guided by chemical
similarity. We implement this method in RetroPath RL, an open-source
and modular command line tool. We validate it on a golden data set
of 20 manually curated experimental pathways as well as on a larger
data set of 152 successful metabolic engineering projects. Moreover,
we provide a novel feature that suggests potential media supplements
to complement the enzymatic synthesis plan.