A method for automated SAR.
Medicinal chemistry has always been analogical: this worked — what is near it, and what happens if we change this piece? We build the method that reasons over your data, and a method does not go obsolete — it compounds. Two algorithms, both the way a chemist already thinks.
Designs new molecules from the transformations your own series has already proved, so every candidate is an analogy you can trace back to a rule in your data.
Estimates what a new molecule will do by carrying your measured results across to it — a grounded what-if, not a black-box score.
Stage one · rules
The rules come out of your own library.
Each one arrives with the evidence behind it. Every rule is mined algorithmically from your own data, never carried in from a generic set.
Stage two · analogical generation
Those rules then build new candidates from molecules you choose.
Every candidate carries the analogy that produced it.
The rule behind it, the molecule it started from, the evidence for both, and whether the edit was checked by chemistry or carried by precedent alone.
A chemist can disagree with a candidate for a reason, and name it. The loop only compounds if each turn can be audited, so the reasoning travels with the molecule.
Stage three · counterfactual prediction
Each candidate is scored against a measured anchor.
We do not ask a model how potent a molecule is in a vacuum. We ask what the measured potency of a seed becomes under this change, carried from precedent and corrected by the error the model makes on the closest real analogues.
That is also what makes the hard cases tractable. A smooth model interpolates across an activity cliff and flattens it; carrying a measured change across transfers the jump from a case that already crossed it.
The method is model-agnostic: it works with any molecular encoder, so a better model makes it sharper, not obsolete — upload your own in the demo. It is peer-reviewed in JCIM and validated on the MoleculeACE activity-cliff benchmark.
Stage four · acquisition and return
An acquisition function chooses what to make next, between exploitation and exploration.
They are ranked for acquisition, not only for score: which experiments buy the most for the next turn.
The library is read again, over data it did not have before. New rules appear, and implied changes become demonstrated, carrying a number for the first time.
What we are introducing
Two new methods — both built to use the data you already have.
Analogical generation turns the transformations your own compounds demonstrate into new molecules: a new way to generate from the data you already hold.
Counterfactual prediction grounds every estimate in your measured results, and applies to any molecular encoder you bring: a new way to predict that rides your model, not ours.