How it works

There is no training step and there is no model to fit. The reference set is the model.

The consortium vote

A query is fingerprinted with a Morgan fingerprint, radius 2, folded to 1,024 bits, and compared by Tanimoto to every compound in the reference set. Those at or above the required similarity form the query's voting consortium. For acute toxicity the call is the mean of their measured values, mapped to a light. For a categorical endpoint it is the majority of their labels. Votes are unweighted; weighting by similarity gave no gain in the original work and is not used here.

When it declines

Two cases, both deliberate. Too few compounds reach the required similarity, or the labels tie and a tie is not a majority. In both the answer is no call, with the nearest similarity found reported beside it so a reader can see how far away the closest reference compound was. A decline is never filled in and never hidden.

Cross-family reach

Computed from two dimensional structure alone, against the Reverse Screen index, using the published PharmCast checkpoint pharmcast_scp_v10.pt. Family assignments come from the Family Foundation Model target index. A target annotated with two families counts once in each, never as a third family of its own, and a matched protein contributes only if it carries a family assignment, so reaching no family means that no matched protein carried an assignment, not that nothing matched. Every reach number on this site was measured on index release 2026-09-20, which carried 27,910 ligands solved against 28,632 distinct proteins; the service scores against the release it is serving, which the provenance endpoint reports.

The reference sets

Acute toxicity is rat oral LD50 from CATMoS, experimental values only. The sampled reference set is 8,599 compounds, of which 5,953 carry an exact value and may vote, with 2,873 compounds held out. The 28 further endpoints come from TOXRIC.

Those come from the CATMoS release in steps. It distributes 50,660 records, of which 11,989 carry experimental data rather than a consensus prediction, and 7,948 of those carry an exact point value; the rest are limit tests or ranges, which cannot enter a regression but fix a hazard category. 515 of the 11,989 carry no hazard category and 2 could not be fingerprinted, so the sweeps run on 11,472 compounds, of which 7,947 carry an exact value. Those 11,472 are split into a reference set of 8,599, with 5,953 carrying an exact value and able to vote, and 2,873 held out, of which 1,994 carry an exact value. The 11,472 the service reports as its acute reference set is that fingerprinted total.

The source data carry model predictions alongside measurements. Only measurements are used here, and a predicted value is never served as a measurement.

The papers

The current work. Muskal SM, Jha SK, Kishore P, Tyagi P. Toxicity by Consortium Revisited: Structural Neighbors and Cross-Family Reach. bioRxiv 2026. doi:10.64898/2026.10.04.756566 It is the work this site serves.

The original method, 2003. Muskal SM, Jha SK, Kishore MP, Tyagi P. A simple and readily integratable approach to toxicity prediction. J Chem Inf Comput Sci 2003;43:1673-1678. doi:10.1021/ci034080c

That paper built the consortium on a proprietary reference set of 13,645 compounds and reported 94% of its answers within one log unit. This site rebuilds the same method on public data, reaching 92.1% within one log unit on the 1,336 held out compounds it answered for, which is 1,336 of the 1,994 that carry an exact value and can be scored at all, and adds a second signal the original could not compute.

The software

The method is the public toxpred package, at github.com/smuskal/toxpred. This site imports it rather than reimplementing it, so the site and the command line answer the same question the same way.