scigantic-deeppk

Client for Deep-PK (small-molecule pharmacokinetic and toxicity prediction across 73 endpoints), plus a mirrored copy of its CYP450 inhibitor/substrate reference data.

import scigantic_deeppk as deeppk

df = deeppk.predict("CC(=O)OC1=CC=CC=C1C(=O)O", pred_type="metabolism")  # aspirin

Installation

$ pip install scigantic-deeppk

Live predictions

Deep-PK ships no downloadable model or pip package, only a free async job API: POST submits SMILES, GET polls by job_id until the job finishes. predict() wraps that round trip and blocks until the result is ready:

df = deeppk.predict("CC(=O)OC1=CC=CC=C1C(=O)O", pred_type="admet")

# Batch
df = deeppk.predict(["CCO", "CC(=O)OC1=CC=CC=C1C(=O)O"], pred_type="metabolism")

pred_type is one of absorption, distribution, excretion, metabolism, toxicity, admet (default, all 73 endpoints). A single molecule typically takes 30-90s; there is no synchronous endpoint, no documented rate limit, and no SLA, so this is fine for exploration but not for anything that needs to run offline or reproducibly.

The API's own docs page names the POST field SMILES_string; the field that actually works, found by testing against the live server, is lowercase smiles (smiles_file for a batch). predict() already gets this right, this is only worth knowing if you're calling the API directly.

CYP450 reference data

Deep-PK's data page also publishes ~80 static training datasets, including CYP1A2/2C9/2C19/2D6/3A4 inhibitor and substrate sets. cyp_reference() reads a small (~1.5MB, 105,695-row) parquet mirror of just those ten, useful for e.g. checking compound overlap against another CYP dataset before trusting a transfer or leakage number:

df = deeppk.cyp_reference("cyp3a4", "inhibitor")
df = deeppk.cyp_reference()  # every isoform and kind

Columns: isoform (one of cyp1a2/cyp2c9/cyp2c19/cyp2d6/cyp3a4), kind (inhibitor/substrate), split (Deep-PK's own training/val/test split), smiles, label (0/1). Row counts per isoform+kind reproduce Deep-PK's own stated totals exactly.

Command line

$ scigantic-deeppk predict "CC(=O)OC1=CC=CC=C1C(=O)O" --pred-type metabolism
$ scigantic-deeppk cyp-reference --isoform cyp3a4 --kind inhibitor

License

MIT-0. See LICENSE. This covers the code in this package only.

Data license

Deep-PK's data page states the CYP450 datasets are "Open Knowledge" but publishes no formal license file or terms beyond that. cyp_reference()'s mirror reproduces that data as-is; check with the Ascher lab (biosig.lab.uq.edu.au/deeppk) before redistributing it further if that matters for your use.