FELA · Molecular Encoder
FELA-Chem
Setup
How hard is a molecule to make? How to find the route? Conditions to try?
Pros
Dehallucinate your candidate lists in seconds on a laptop.
Usage
Did you think ibuprofen was hard to make? Find out!
What we delivered.
Paste a target structure and triage it on the spot - is it worth making?
Can you get a suggestion on a way to make it? Which conditions should one try first? Screen a
whole candidate list in the time it takes to scope one route by hand, and
do it cheaply on commodity hardware, like your laptop - instead of the ever so nebulous cloud.
Intellectual Property safety is a pretty nice benefit of efficient compute!
Under the hood, FELA-Chem is a 286M parameter chemistry foundation model, built
on the FELA architecture. One encoder drives a whole property / yield / retro / synthesis
suite. Here it does three real jobs at once, all on abundant hardware at speed and scale.
1 · Synthesizability
An SCScore-style makeability score (1 = trivial, 6 = very hard). The
head reproduces the standard SAScore at Spearman 0.92 on 6k
held out molecules, a fast learned complexity read straight from
structure.
2 · One step retrosynthesis
An 18,240-class USPTO-full template ranker (held out
top-1 ≈ 0.78 / top-10 ≈ 0.96) proposes disconnections; each top
template is executed with rdchiral to give concrete precursor molecules.
The route search built on this policy clears AiZynthFinder on our own
evaluation.
3 · Reaction conditions
Suggested starting points, not a validated recipe. Catalyst / solvent /
reagent heads trained on USPTO. Solvent is the meaningful signal
(top-1 ≈ 0.60, well above the frequency baseline).
Full card on Hugging Face ↗