scigantic-comptox

Query EPA's CompTox Chemicals Dashboard from Python. Two halves, and only one of them needs anything from you:

• ToxCast bioactivity, from a public S3 mirror over DuckDB. No API key, no download, no local database.
• Live Chemical/Hazard/Exposure lookups, over EPA's own CCTE REST API. Optional, and requires your own EPA API key.

import scigantic_comptox as comptox

df = comptox.bioactivity(dtxsid="DTXSID7020182", hitc=1.0)

That query runs against s3://scigantic-comptox over DuckDB's httpfs extension.

Installation

$ pip install scigantic-comptox

What's in the bioactivity mirror

Every row is the winning dose-response model plus hit-call for one (chemical, ToxCast assay endpoint) pair, from EPA's invitrodb v4.3 mc5-6_winning_model_fits summary tier: AC50, potency, top, confidence bounds, and hitc/hitcall (active/inactive/ambiguous). This is the same level of detail as PubChem's own concise BioAssay tables, deliberately not the raw well-level plate reads or the full set of losing candidate curve models EPA also publishes (mc4_all_model_fits) -- that table is what makes EPA's own "summary" download 7GB instead of the ~600MB actually mirrored here; almost nobody needs the losing models, they need the winner.

comptox.bioactivity(aeid=1114, limit=5)

A real data-quality issue is handled for you. EPA's export uses the literal string "NA" for missing values in several columns (including dsstox_substance_id), not a real SQL NULL -- filtering with IS NOT NULL silently misses these. bioactivity() normalizes this (NULLIF(..., 'NA')) so a real NULL check works as expected; it also casts columns EPA ships as text purely because of those "NA" strings (ac50, bmd, top, and others) to real DOUBLEs. Use bioactivity_raw() if you want EPA's exact upstream types instead.

PubChem cross-reference

EPA publishes a per-assay-endpoint PubChem CID cross-reference alongside each invitrodb release. This wraps the consolidated version of that file:

comptox.pubchem_bridge(dtxsid="DTXSID7020182")

Pairs naturally with scigantic-pubchem's own BioAssay/gene/protein coverage.

What's not mirrored (yet)

DSSTox bulk chemical structures aren't mirrored -- EPA's bulk distribution for that turned out to be a large, unstructured institutional drive rather than a clean single file. comptox.releases() reports this plainly (structures: False) rather than silently missing the data. For now, chemical structure lookup by DTXSID goes through the live REST half below, the same place EPA's own client gets it from.

The live REST half: bring your own EPA API key

EPA's CCTE API needs one, and deliberately doesn't hand out a shared one: their own docs describe it as "an individual API key" that "uniquely identifies the user," issued by emailing ccte_api@epa.gov rather than self-serve signup. EPA's own reference client, ctx-python, never embeds a key either. This package follows the same rule -- no key ships with it, and none is proxied on your behalf.

$ export COMPTOX_API_KEY=your-key-here

comptox.chemical_detail("DTXSID7020182")
comptox.hazard_toxval("DTXSID7020182")
comptox.exposure_httk("DTXSID7020182")

Calling any of these without a key raises MissingApiKeyError with a link to how to get one, not a bare 401. Requests are retried with backoff on 429/5xx, and responses are cached for 30 days by default (the opposite default from the mirror half's caching, see below) -- EPA's API has no documented rate limit, but it's still a live government service, not an unlimited public mirror, so repeatedly re-fetching the same lookup in a loop is worth avoiding by default.

Endpoint coverage here is intentionally a starting set (one real, verified path per microservice), not a full reimplementation of ctx-python -- that package already covers Chemical/Exposure/Hazard comprehensively. This package's actual differentiator is the bioactivity mirror above, which ctx-python doesn't cover at all.

Working offline (mirror half)

Off by default, since zero setup is the whole point of the mirror:

comptox.enable_cache()
df = comptox.pubchem_bridge()  # downloads the file once, then reads from disk

Downloads to ~/.cache/scigantic-comptox (override with enable_cache(cache_dir=...) or the SCIGANTIC_COMPTOX_CACHE environment variable). Concurrent callers racing the first download of the same file wait for it rather than each downloading their own copy.

connect(), query(), and bioactivity() don't participate in this (registering the underlying views eagerly would mean the first call for any release downloads the full ~600MB regardless of what the query actually touches); use comptox.cache_resolve("v4_3/derived/pubchem_bridge.parquet") to cache a specific file yourself.

The live REST half has its own separate cache, ON by default (comptox.disable_ctx_cache() to turn it off) -- see above for why the two halves use opposite defaults.

What's mirrored

comptox.releases()

This table isn't hardcoded. releases() reads a small manifest published alongside each mirror run.

Command line

$ scigantic-comptox info
$ scigantic-comptox query "SELECT count(*) FROM bioactivity" --release v4_3

License

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

Data license

EPA's invitrodb v4.3 (ToxCast bioactivity data) is CC0 -- public domain, no attribution or share-alike terms to track. A courtesy citation is appreciated, not required: Filer, D.L. et al. (2017), tcpl: the ToxCast pipeline for high-throughput screening data, Bioinformatics, doi:10.1093/bioinformatics/btw680.