Contribute

Send what you have, in whatever form you have it. Three formats are accepted and all three are useful.

Every result on this site rests on the size of the pooled reference set. Growing the voting set from 127 to 5,953 compounds raised the share of compounds answered for from 3.3% to 46.5%, with accuracy holding between 91.2% and 92.1% once past the first few hundred. A contribution buys reach, for you as much as for anyone else, and toxicity findings are the least contested data a discovery organization holds: a compound that showed toxicity was deprioritized long ago and carries no position anyone is defending.

Structures and measurements

The most useful form, and the simplest: a CSV or TSV with a smiles column and one column per endpoint you measured. Nothing to install and nothing to run. Paste it below or choose the file. A compound sent this way votes in the consortium directly, the same as any compound in the published reference set.

smiles,CYP3A4,Ames Mutagenicity,pLD50
CC(=O)Oc1ccccc1C(=O)O,0,0,2.71
Cn1cnc2c1c(=O)n(C)c(=O)n2C,0,,2.26

A blank cell is fine. A compound measured on one endpoint and not another still counts on the one it was measured on, and the check below reports the per endpoint totals it found.

Fingerprints and measurements

If the structures stay in house, a folded Morgan fingerprint carries enough for the compound to vote. The record holds the on bits and your endpoint values. The toxpred package builds one from your CSV on your own machine, which is the only reason the command line exists: it computes the fingerprints locally so that only the fingerprints travel.

git clone https://github.com/smuskal/toxpred.git
cd toxpred
conda env create -f environment.yml
conda activate toxpred
toxpred contribute --input your_data.csv --out record.json --format fingerprint

Reach counts and measurements

The most compact record: your endpoint values beside a handful of integers describing how widely each compound reaches across protein families. On their own the counts are worth 0.665 across the 28 endpoints. Added to a coarsened heavy atom count, cLogP and molecular weight they reach 0.740 against 0.724 for those size descriptors alone, a gain of 0.016 that holds on 23 of the 28 endpoints.

The whole reverse screen, which needs the fingerprint, is worth 0.755 against 0.722 for exact size descriptors alone, a gain of 0.033 on 27 of the 28 endpoints. The two baselines are not the same, coarsened against exact, so it is the gains that compare: 0.016 against 0.033. That is where about half comes from. A few integers per compound keep about half of what the full reverse screen is worth, and the fingerprint record carries the rest.

toxpred contribute --input your_data.csv --out record.json --format counts

Send it

Paste a CSV of structures, or a record.json built by the toxpred package, or choose a file. Checking it stores nothing and tells you what was found. Sending is a separate step, and a contribution is reviewed before it joins the reference set.

smiles,CYP3A4,pLD50
CC(=O)Oc1ccccc1C(=O)O,0,3.21
Cn1cnc2c1c(=O)n(C)c(=O)n2C,1,2.85

{
 "meta": {"schema": 1, "format": "fingerprint",
          "n_compounds": 2, "fingerprint": "Morgan",
          "radius": 2, "bits": 1024,
          "endpoints": ["CYP3A4", "pLD50"]},
 "records": [
  {"bits": [11, 23, 33, 64, 175, 356, 386, 389, 423,
            444, 456, 592, 650, 695, 705, 726, 751,
            807, 849, 893, 909, 946, 967, 1017],
   "endpoints": {"CYP3A4": "0", "pLD50": "3.21"}},
  {"bits": [0, 33, 121, 179, 234, 283, 314, 330, 356,
            378, 385, 400, 416, 428, 463, 493, 504,
            564, 650, 672, 771, 849, 932, 935],
   "endpoints": {"CYP3A4": "1", "pLD50": "2.85"}}
 ]
}

The package runs on your machine and writes a file you can read before you send it.

What happens to it

A contribution is reviewed, then folded into a contributed reference set that is kept separate from the published one. Scoring runs against the published set by default, exactly as the paper and the package do, and a visitor can choose to include contributed data. The consortium page reports what has arrived and when the set last changed.