Cross-family reach

The second signal. How many distinct protein families a compound's chemistry reaches in the crystallographic record. It is not the consortium vote and it is not computed the same way.

The query is fingerprinted with PharmCast and screened against the Reverse Screen index, release 2026-10-29, 33,642 indexed ligands. The proteins those matched ligands were solved against are collected, and their distinct families are counted. A matched protein contributes only if it carries a family assignment, so a compound can hold matches and still reach no family: reaching no family means that no matched protein carried an assignment, not that nothing matched. No protein structure, no docked pose and no three dimensional model of the query is needed.

The table lists the 24 endpoints measured across the full range, from reaching no family through reaching sixteen or more. The four cardiotoxicity thresholds are measured from one to three families upward and are covered by the sweeps on the endpoints page.

Two fingerprints, two cutoffs, one count

Reach is measured over both arms the Reverse Screen index carries. The pharmacophore arm keeps every indexed ligand whose PharmCast similarity to the query is at least 0.50. The Morgan arm keeps every indexed ligand whose Morgan similarity, radius two over 2,048 bits, is at least 0.25. The proteins the two sets of ligands were solved against are combined and counted once each, so a ligand found by both arms is one ligand and a protein reached through several ligands is one protein.

The two cutoffs differ because the two similarities are on different scales, and one number applied to both would mean two things at once. The Morgan cutoff was chosen so that the Morgan arm admits the same median number of ligands per compound as the pharmacophore arm admits at 0.50, which makes the two arms comparable in how much evidence each one opens.

Measured over 34,198 compounds against release 2026-09-20.

What the second arm adds

The clearest gain is coverage. 6,153 compounds reach no family through the pharmacophore arm, and 4,166 of them reach at least one once the Morgan arm is included, which is 12.2% of every compound measured, and they gain a median of seven families. They are mostly small aliphatic esters and ketones, nonyl acetate and hexyl acetate among them: flexible chains that carry few pharmacophoric features and so fall below the pharmacophore cutoff, while sharing enough substructure with an indexed ligand to clear the Morgan one. Their Morgan fingerprints carry a median of 21 set bits against 36 across all compounds measured.

The second gain is accuracy against cytochrome interaction. Carried as separate features rather than as one combined number, the two arms predict cytochrome inhibition better than the pharmacophore arm alone: the mean area under the curve across the five cytochromes rises from 0.67 to 0.71, measured with five fold cross validation so that the extra features cannot pay for themselves. Counting proteins as well as families accounts for 0.03 of that, and the Morgan arm contributes a further 0.01 on top of it.

The per compound figures quoted elsewhere on this site, and in the structure panels, are the pharmacophore arm at 0.50, which is the measurement every published figure is on. The Try it page reports both arms for any molecule.

Reaching widely does not mean more kinds of toxicity

Reaching more protein families raises the chance of a particular liability. It does not multiply the number of liabilities a compound carries. Among the 5,882 compounds measured on ten or more endpoints, the share of a compound's endpoints that come back toxic is 8.4% for those reaching no family and 10.7% for those reaching sixteen or more, and the number of distinct toxicity categories it is positive in does not rise at all, 0.95 against 0.81. The rank correlation between families reached and the share of measured endpoints that are toxic is 0.05, and against the count of distinct categories it is -0.04. The signal is specific to particular endpoints.

The index does not need to grow

Subsampling the published index uniformly at random to three quarters, a half and a quarter of its 27,910 ligands, and rescreening every compound at each size, leaves the separation where it was. Averaged over the 28 endpoints, and with families reached binned to six quantile bins at every size so that resolution is held constant, the area under the curve changes by less than 0.01 across the whole range, from three independent draws at each reduced size. Compounds do reach fewer families as the index shrinks, from 9.9 on average to 6.5, and they are ranked no less well for it. Anyone working from a smaller or an older release of the index should expect the same behavior.

The numbers above were measured on index release 2026-09-20, which carried 27,910 ligands.

The same rise on compounds from outside

Everything above was measured on compounds this work had access to. The OpenADMET consortium has since measured CYP3A4 inhibition on 2,335 compounds it did not. Some of them are themselves ligands in the index the reverse screen searches, and some carry a measurement in the endpoint tables used above. Removing every compound in either group leaves the separation where it is. A compound was counted as an inhibitor at a CYP3A4 pIC50 of 5.00 or above, equivalent to an IC50 of 10 uM or below, and that threshold was fixed before any compound was screened. On that rule 481 of the 2,335 are inhibitors, a rate of 20.6%. Grouped by the number of protein families the screen reaches, the percentage measured as an inhibitor rises the same way it does on the internal endpoints.

Comparing families reached between the inhibitors and the rest gives a Mann-Whitney p of 5.09e-22. Read as a filter, a cut at sixteen families sets aside 740 compounds, just under a third of the set, at 12.0% inhibitor against 24.6% above the cut. The lower-reach group still contains 89 of the 481 inhibitors. This covers the family count against that one endpoint. It is not a test of the triage score or of the consortium vote.

These numbers were measured on index release 2026-09-20, the release used everywhere else on this page, keeping every match at or above the cutoff with no limit on how many a compound may have.

Six endpoints run the other way

Six endpoints are no higher among compounds reaching sixteen or more families than among those reaching none: Ames Mutagenicity 52.3% to 48.0%, Carcinogenicity 52.9% to 37.3%, Eye Corrosion 82.2% to 1.9%, Eye Irritation 94.9% to 22.4%, Hepatotoxicity 47.3% to 47.1% and Respiratory Toxicity 71.3% to 56.9%. CYP2D6 takes a third shape, rising from 5.9% to 16.5% at four to seven families and easing back to 13.7%. All are in the table above and none is hidden: a signal that holds on most endpoints and not on all is what the measurement found.

It is not a compound finding its own analogs

Removing every indexed ligand that shares a scaffold with the query, and then every close analog of it, leaves the bands where they were. For CYP3A4 no band moves by more than 0.4 of a percentage point, and across all 28 endpoints the largest shift in any band under either blinding is 3.2 points. The test is not a weak one: 35.2% of compounds share a scaffold with an indexed ligand and 47.0% have a match at or above 0.70, so there was a great deal to remove.

The same index answers reversescreen.ai, where a molecule can be taken further and docked into the sites it reaches.