Not another literature-trained model
A hybrid, not a black box.
Most "AI for chemistry" tools learn what's likely from what's been published — and inherit every gap and bias in that record. Allchemy starts somewhere else: an expert-curated set of organic chemistry rules and heuristics that doesn't need a dataset to know a reaction is impossible.
- No training corpus, at any point. Routes come from expert-encoded reaction rules and heuristics — not from mining precedent. There is no dataset to run out of, go stale, or be biased by.
- Built-in negative data. Allchemy considers 700 incompatible functional-group combinations, encoded directly into the rule base, so the system rejects unrealistic chemistry instead of needing to learn to avoid it statistically.
- Works on day one, on your target. No customer reaction history, no proprietary dataset, no warm-up period required to get a usable, explainable route.
- AI layered on top, deliberately. Machine-learned models handle what they're genuinely good at — property prediction, protein-binding estimates, ADME/Tox scoring — while route feasibility stays governed by chemistry logic a chemist can inspect and trust.
- Your data, only if you want it in. Allchemy doesn't need your reaction history to work — but if your team has proprietary data worth structuring, we can help digitize it for your own internal use.
0
Proprietary datasets required to get a first route. Allchemy doesn't need to see your reaction history, your ELN, or a training corpus — the chemistry logic is already there, expert-curated, before your target ever comes in.
Fig. 1 — Manual retrosynthesis: one disconnection back from the target, with synthon options
live platform view
Fig. 1b — Automated retrosynthesis: a complete route tree, generated end to end