The 28 endpoints

The same method, with only the required similarity set per endpoint. Each row is that endpoint at its operating point: the strictest match that still answers about one compound in five.

Two rates are reported rather than percent correct. The labels are lopsided on most of these endpoints, and that ratio records which compounds were tested and reported rather than anything about chemistry. Caught is the share of measured actives called active. Passed is the share of measured inactives called inactive.

Five endpoints rest on fewer than twenty measured actives; the measured active column is there so a reader can see which. Cardiotoxicity at 30 micromolar is the one endpoint that fails in the other direction, catching most of the actives and passing few of the inactives.

Eye corrosion is the one endpoint that answers perfectly, 100.0% caught and 100.0% passed at a required similarity of 0.65. It does so for 23.4% of compounds and rests on 62 measured actives, and its positive class is small reactive molecules, which are the easiest kind to match to a close neighbor. The fingerprint is folded to 1,024 bits, so distinct structures can share one. Read the score with the coverage and the count beside it.

The reference sets behind them

Across the 28 scored endpoints the table holds 149,224 measurements over 34,162 compounds, from 156,086 raw records before duplicate resolution. A measurement is one value for one compound on one endpoint. The compound total is lower than the sum of the unique compound column because a compound measured on several endpoints appears in several rows. The table lists 30 endpoints. Developmental Toxicity and Reproductive Toxicity are not scored and are not in these totals.

Where these data come from

Every endpoint on this page is public, and every measurement behind it was published by someone else. The links go to the source rather than to our copy of it.

The toxpred package fetches both reference sets for you with toxpred fetch, and toxpred provenance prints the URL, the size and the sha256 of every file it took. The download page carries the same checksums for the files this site scores against.

Two of these thirty carry too few compounds to score against and are not offered: developmental toxicity and reproductive toxicity. The 28 that are scored are the ones above that floor.

Running a library through the score

The retrieval measures, the cross-family target preferences and the size descriptors combine into one score, fitted on 70% of the compounds and measured only on the 30% held back. The consortium predictions are evaluated separately and are not inputs to it. Put a library through it and the question is what the worst scoring tenth contains.

Read a row as a pair. The tenth the score flags holds 16.7 actives in 100 at the androgen receptor ligand binding domain where the library it came from held 2.8, and that tenth carries 60.4% of every active in the set.

Discarding the worst scoring fifth rather than the tenth removes 64.1% of the androgen receptor ligand binding domain actives, 67.9% of the androgen receptor actives, 49.6% of the CYP2D6 inhibitors and 46.4% of the CYP3A4 inhibitors, against 20% for discarding at random.

What the score can buy is set by how rare the outcome is. Where half the library already carries the endpoint there is little to sort: hepatotoxicity moves from 49.9 in 100 to 83.5, and cardiotoxicity at 30 micromolar from 82.2 to 91.9. The method earns its place on the rare outcomes, which are also the ones a triage step is run to find.