Toxicity by consortium

A compound inherits from the compounds it resembles. Draw a molecule and the compounds in the public reference set that are similar enough to it vote on how toxic it is. When nothing is similar enough, the method declines, and that is the answer.

92.1%

of answers land within one log unit of the measured rat oral LD50, on the 1,336 held out compounds it answered for

1,336

of 2,873 held out compounds answered for, at a required similarity of 0.50

further toxicity endpoints, scored by the same method with nothing retuned

27,910

indexed ligands behind the cross-family reach signal

Accuracy and coverage are two numbers, always

Raising the similarity a reference compound must reach before it may vote raises accuracy and lowers the share of compounds answered for. Across the sweep, accuracy runs 87.7% to 96.6% while coverage runs 69.3% down to 9.2%. No setting escapes that trade, and this site never shows one of the pair without the other.

three sweeps: required similarity, voters required, and reference set size — Required similarity, voters required, and the size of the voting reference set. The blue series is the share of compounds answered for; it falls wherever accuracy rises.

A shared pool buys reach, not accuracy

Growing the voting reference set from 127 to 5,953 compounds carrying a measured LD50 leaves accuracy between 85.4% and 92.1% while the share of compounds answered for rises from 3.3% to 46.5%. That is the case for pooling: a larger consortium does not make a call better, it makes more calls possible. The sampled reference set is 8,599 compounds, of which 5,953 carry an exact LD50 and may vote.

The second signal: how widely a compound reaches

A compound whose structural matches reach no protein family inhibits CYP3A4 2.5% of the time. One reaching sixteen or more inhibits it 51.0%. The count comes from screening the query against the Reverse Screen index of ligands solved in protein structures, kept current with reversescreen.ai, and counting the distinct protein families their structures were solved against; it needs no protein structure and no docked pose.

Twelve compounds that reach widely

twelve compounds, each reaching sixteen or more protein families, with the endpoints each is toxic on — Twelve of the 34,197 compounds in the TOXRIC tables, every one of them reaching sixteen or more protein families. Reach is counted against the published index of 27,910 ligands solved in protein structures, at a required similarity of 0.50. Nine binding endpoints are counted, and the denominator is how many of those nine carry a measurement for that compound. Open the figure in a new tab for it at full size.

Every SMILES below is selectable, and clicking one loads it straight into the editor on the scoring page.

Compound	Families	Proteins	Toxic on	SMILES
Emodin	18	224	6 of 7	Cc1cc(O)c2c(c1)C(=O)c1cc(O)cc(O)c1C2=O
Fipexide	20	471	4 of 7	O=C(COc1ccc(Cl)cc1)N1CCN(Cc2ccc3c(c2)OCO3)CC1
Capsazepine	19	463	4 of 5	Oc1cc2c(cc1O)CN(C(=S)NCCc1ccc(Cl)cc1)CCC2
Clemizole	18	99	4 of 6	Clc1ccc(Cn2c(CN3CCCC3)nc3ccccc32)cc1
Phloretin	18	499	4 of 5	O=C(CCc1ccc(O)cc1)c1c(O)cc(O)cc1O
Dihydrocapsaicin	17	720	4 of 5	COc1cc(CNC(=O)CCCCCCC(C)C)ccc1O
Amsacrine	17	294	4 of 7	COc1cc(NS(C)(=O)=O)ccc1Nc1c2ccccc2nc2ccccc12
Nonivamide	16	292	4 of 5	CCCCCCCCC(=O)NCc1ccc(O)c(OC)c1
(R)-Felodipine	24	229	3 of 3	CCOC(=O)C1=C(C)NC(C)=C(C(=O)OC)[C@H]1c1cccc(Cl)c1Cl
Dioxybenzone	23	359	3 of 7	COc1ccc(C(=O)c2ccccc2O)c(O)c1
Nitazoxanide	22	321	3 of 4	CC(=O)Oc1ccccc1C(=O)Nc1ncc([N+](=O)[O-])s1
Mebendazole	22	345	3 of 7	COC(=O)Nc1nc2cc(C(=O)c3ccccc3)ccc2[nH]1

percent toxic against families reached, and the proteins behind it — Percent toxic by the number of families a compound reaches, and the proteins that carry the association.

How reach is computed, and its exceptions.

What it is for

Scoring a library before any of it is made, then throwing away the worst of it. The method helps most where the problem is rare, which is where a program does not see it coming until a chemical series has already been spent on it. Every endpoint and its operating point is published, including the ones where the method is weak.