For the medicinal chemist

Active learning turns what hit-to-lead explores into what lead optimization exploits.

What carries between the two stages is a SAR rulebook: rules mined from your own compounds, each holding the measured pairs behind it and the scaffolds it has held on. Every assay round sharpens it, and it is what names the next compound to make.

Stage one · hit-to-lead

Build the SAR map across every series.

Hit-to-lead explores on two fronts at once: every chemotype series in the library, and every rule the data will support.

The first rule set comes out of data you already hold — screening output, discontinued series, literature compounds. Each rule carries how many pairs back it and where it has held. Acquisition is exploration-heavy: the bottom rule rests on a single pair and its span is the widest, so that is the compound the round buys most from.

the group a rule replaces the group it puts there a measured pair behind the rule it has held on a scaffold like this

Each round hops rules between chemotypes, so every series stays in play — and the one that opened weakest can finish best. Every assay result sharpens a rule and becomes another anchor for a counterfactual prediction. What accumulates is the SAR map and the compounds to predict from, and the map is what says which series to carry.

where a series started opened weak, finished best other chemotype series a rule carried from one series to another

Stage two · lead optimization

Commit to a series, and work it with rules that have held.

Lead optimization exploits what those rounds found: the best chemotype series, and the rules best supported on it.

Most of the compound budget goes into that series. The compounds those rules propose go to the counterfactual predictor, which carries each value across from its measured anchor and returns an interval with it. Acquisition is exploitation-heavy: highest point, tightest span — where hit-to-lead took the widest.

the group a rule replaces the group it puts there a measured pair behind the rule it has held on a scaffold like this

Later rounds acquire by exploitation.

Exploitation is simply making the best predicted compound — which comes from the best-supported rule applied to the series you chose. The candidate carries that rule and the anchor its prediction came from, so a chemist can inspect either.

the group the rule replaces the group it puts there four measured pairs back this rule and it has held on a scaffold like this one

Across all stages

Three properties hold at every stage.

Acquisition

Two steps. The rule score triages the enumeration down to what is worth predicting; the counterfactual predictor then returns a value and an interval for each survivor, and the acquisition function ranks on both.

Derivation

Each compound records the rule that produced it and the measured compound behind its prediction, so a chemist can inspect either and reject a candidate on a stated basis.

Review

Each round revises rule support and the meta-rules describing where rules hold. The revisions are reported, so the rule set stays open to inspection and direction.

neoLAB

Open it and run the engine yourself — on a loaded library, or on your own.

Three kinase libraries come already trained, so predictions appear as soon as you open one. Point it at a CSV of your own and the rules come out of that instead. Or read what it produced on all three first.

Try neoLAB See neoLAB outputs

Every experiment deepens a library of analogies nobody else holds, and that is what makes the next pass better.

Neopoly Ltd · 2026