Where the models are weakest.
Potency was fitted on 259 targets and selectivity on 271. The bundle carries 400 targets, so some it accepts were never in training. Every accuracy figure published here covers comparisons between targets in the fitted set. Nothing here measures what happens outside it.
The ranking pages separate the two groups, and a target that was not fitted cannot be scored at all. The interface refuses it and so does the API.
It returns which of two wins, not how tightly either binds. It has no opinion on whether either compound is active at all, so a confident ranking of two inactive compounds is still a confident ranking.
Potency scores 0.77 overall. Across the 196 targets carrying at least 50 held-out comparisons the median is 0.79 and none fall below 0.60. Selectivity scores 0.80 overall, and across its 104 targets at that threshold the median is 0.82, also with none below 0.60. The weakest potency target at that threshold reads 0.61 and the weakest selectivity target 0.65.
Targets with only a handful of held-out comparisons can read 0.00 or 1.00 and neither figure means anything. The counts are shown beside every accuracy on the targets page.
Each target's own accuracy is shown beside its name on the ranking pages.
This is where a selectivity answer would be most valuable and where the model is weakest. The muscarinic targets M1 through M5 sat near chance on the previous build. They now run between 0.75 and 0.78, which is a real gain and still the low end of the range. Opioid targets reach 0.90.
The strength returned with a comparison is not the chance that the particular answer is right. It is a ranking signal whose reliability has been measured across a large held-out set. A comparison in the 0.70 to 0.80 band belongs to a population that is right 92% of the time for potency and 87% for selectivity. It does not mean that answer is 70% likely to be correct.
On the previous build potency stopped improving above 0.80 and fell back. On this build both curves rise all the way to the top band, so the 0.90 cutoff is offered again.
Ten percent of compounds are held out, stratified so each target contributes about a tenth of its own. A compound appears on one side of the split only, so no held-out comparison involves a structure the model was fitted on. Each target's share ran from 0.08 to 0.11, median 0.10, and no target with 20 or more compounds was left with nothing held out. 22,513 compounds are held out of 225,676.
An earlier build split on publication date instead. That put far more of one source's chemistry in the test set than the other's, which made the two incomparable, so it was replaced.
"More potent" spans IC50, Ki, Kd, EC50 and Kb together. Comparisons are matched by endpoint, so a functional reading is never ranked directly against a binding one, but the single headline accuracy averages over both kinds of question.
An earlier build required an assay to carry at least ten compounds and a pActivity spread of at least 0.5. This one admits every assay. It adds about a third more measurements and lifts both accuracies. It also admits assays too small or too flat to order anything, and those contribute comparisons the model cannot do better than guess on.
These are the measured figures for the 12 September 2026 build, read from the build artifacts and reproduced here by rescoring the held-out set.