Limitations

It is a comparator, not an affinity predictor

It returns which of two targets a compound prefers. It has no opinion on whether the compound binds either one at all, so a confident ranking of two targets it barely touches is still a confident ranking.

A near-tie is a near-tie

Where the two measured values sit within half a log the model is right 0.59 of the time across families and 0.64 within one, which in both cases is much closer to a coin flip than the headline. It reaches 0.91 and 0.92 only when the true separation is beyond two logs. The strength returned with a comparison is the guide to which case you are in.

Some family pairings are much harder than others

Comparisons between two families, over pairings carrying at least 100 held-out comparisons:

Pairings with fewer than 30 held-out comparisons are not quoted at all, here or in the manifest.

Strength is a population statistic

The strength returned with a comparison is not the chance that the particular answer is right. It is a ranking signal whose reliability was measured across a large held-out set. A comparison at or above 0.80 belongs to a population that is right 0.94 of the time across families and 0.95 of the time within one.

The two kinds of comparison are never pooled, and neither should you

Across families and within one family are measured on separate held-out sets, 8,689 and 32,738 comparisons, and this site quotes each on its own. A single number over both would be dominated by the within-family arm, which carries 79.0% of the holdout. Do not average the two, and do not read a within-family figure as covering an across-family question.

A within-family accuracy rests on the proteins behind it

A family's figure is measured over however many distinct proteins appear in its held-out comparisons, and that count varies by more than an order of magnitude: Family A G protein-coupled receptor carries 5,412 comparisons over 160 proteins, while Toll-like and Il-1 receptors reaches 0.94 over 6. A thin roster can make a family look sharper than it is, so the target count is printed beside every one of them on the methods page.

13 further families carry within-family data but fewer than 200 held-out comparisons, so they are marked rather than scored and no accuracy is published for them: Primary active transporter (170), Writer (143), Other ion channel (112), Family C G protein-coupled receptor (85), Fatty acid binding protein family (62), Family B G protein-coupled receptor (50), Isomerase (39), Aminoacyltransferase (15), Transmembrane 1-electron transfer carriers (8), Calcium channel auxiliary subunit alpha2delta family (3), Taste family G protein-coupled receptor (3), Ligase (1), Frizzled family G protein-coupled receptor (1).

Two deep dives go further inside their own family

For two kinases the sibling Kinase Foundation Model, and for two GPCRs the GPCR Foundation Model, are built on data assembled for that family alone and see more of it than a roster-wide model can.

It covers the targets it was fitted on

2,279 targets across 34 families are servable, and a target outside that roster is refused rather than guessed at. Targets are named by UniProt accession because gene symbols are many-to-many against accessions and the failure is silent.

Breadth across families is not a toxicity claim

A compound that ranks highly against many families is not thereby promiscuous or unsafe. Cross-family breadth in this data partly records how many assay panels a compound went through. This is an off-target triage and repurposing tool, and that is the claim the measurements support.