Tox21 Full

This is a Tox21-like dataset created from the public PubChem BioAssay Tox21 summary assay records. The common Tox21 benchmark includes 12 assays; this generator currently includes 75 Tox21 summary assays.

The package builds a multi-task molecular activity table with one SMILES column and one binary activity column per assay.

Downloads PubChem assay data and creates a clean CSV.GZ file ready for import into pandas:

tox21full ~/Downloads/tox21full.csv.gz

You can also create it as a parquet file (more efficient):

tox21full --format parquet ~/Downloads/tox21full.parquet

Benchmark workflow

The benchmark protocol lives in benchmarks/. It creates a deterministic scaffold split, runs ECFP/logistic regression, ECFP/random forest, and RDKit-descriptor boosted-tree baselines, and prepares the same split for a Chemprop MPNN baseline.

python -m pip install -r benchmarks/requirements-benchmark.txt
python benchmarks/run_sklearn_baselines.py \
    --data tox21full.csv \
    --out-dir benchmarks/results \
    --bootstrap 1000 \
    --seed 20260506

For the Chemprop baseline:

python -m pip install -r benchmarks/requirements-chemprop.txt
python benchmarks/prepare_chemprop.py \
    --data tox21full.csv \
    --split-file benchmarks/results/scaffold_split.csv \
    --out-dir benchmarks/results/chemprop
bash benchmarks/results/chemprop/run_chemprop.sh