openFDA

The FDA's own post-market record, served as a public REST API: 20.6 M adverse-event reports, every drug application since 1939, and the full text of approved labels. It answers the question that comes after the biology — what has already happened to this molecule, or to its class, in people — and it is the cheapest safety evidence available anywhere.

Base URL: https://api.fda.gov — REST, no key required. Docs: open.fda.gov/apis · field reference Checked against: the live API, August 2026; FAERS data current to 2026-07-30.

Read references/api-reference.md before writing a query by hand, references/endpoint-fields.md before trusting a field name, and references/disproportionality.md before reporting any signal — that one is judgement, not syntax.

The three scripts

Script	Answers
fda_adverse.py	What has been reported against this drug, and is any of it disproportionate?
fda_approvals.py	When was it approved, by whom, and how many indications has it gained?
fda_labels.py	What does the approved label actually say?

Zero results arrive as HTTP 404

This is the single thing to get right. A search that matches nothing returns:

HTTP 404  {"error": {"code": "NOT_FOUND", "message": "No matches found!"}}

That is a successful query with an empty result set. Any client that treats non-200 as failure turns "this drug has no reports" into a crash, and — worse — makes a typo indistinguishable from a real zero, because a misspelled field name also returns 404. get() in scripts/_common.py converts NOT_FOUND into an empty payload; when a count comes back empty, check the field name against references/endpoint-fields.md before believing it.

The other two surprises: limit above 1000 returns 403 API_KEY_MISSING (a key raises the daily quota, not the per-request cap — lower limit instead), and skip is hard-capped at 25000, so a search matching 500 000 reports has 25 000 reachable records. Partition by receivedate to go deeper.

Adverse events

Rank what is reported, using a server-side aggregation rather than paging:

python skills/openfda/scripts/fda_adverse.py reactions --drug atorvastatin --top 6

# 518912 reports, top 6 reactions
reaction         reports  share_pct
FATIGUE          29506    5.69
NAUSEA           25700    4.95
DIARRHOEA        25358    4.89
DYSPNOEA         24427    4.71
DRUG INEFFECTIVE 24372    4.7
DIZZINESS        20872    4.02

Those are the generic complaints of a widely prescribed drug — the list is dominated by how many people take it, not by what it does. To get pharmacology out of FAERS you need a comparator.

Disproportionality

signal builds a 2x2 table from four separate totals and scores the pair:

python skills/openfda/scripts/fda_adverse.py signal --drug atorvastatin --reaction RHABDOMYOLYSIS

# 2x2: a=5713 b=513199 c=35397 d=20138381
a     b       c      d         prr    ror    ror_ci_low  ror_ci_high  chi2      signal
5713  513199  35397  20138381  6.275  6.333  6.158       6.514        21854.8   true

Statin-associated rhabdomyolysis — the toxicity that withdrew cerivastatin in 2001. Compare the same drug against a background event:

# atorvastatin x ALOPECIA
5344  513568  193048  19980730  1.076  1.077  1.048  1.107  28.33  false

This pair is why the rule is a conjunction of three tests (a ≥ 3, PRR ≥ 2, chi² ≥ 4). Alopecia's chi-squared is 28 — seven times the threshold — and on that statistic alone it would be flagged. Its PRR of 1.08 says atorvastatin reports mention alopecia at the background rate. The chi-squared is large only because N is 20 million.

Each signal run costs four requests against a 1000/day anonymous quota.

Approvals

python skills/openfda/scripts/fda_approvals.py application --ingredient pembrolizumab

application_number  sponsor             brand_names    original_approval  submissions  efficacy_supplements
BLA125514           MERCK SHARP DOHME   KEYTRUDA       20140904           127          108
BLA761467           MERCK SHARP DOHME   KEYTRUDA QLEX  20250919           11           7

efficacy_supplements counts approved SUPPL submissions with class EFFICACY — roughly, how many times the label gained an indication. It is the cheapest available measure of how far a drug travelled beyond its first approval. timeline --appno BLA125514 --type ORIG shows that first approval was a PRIORITY review of a Type 1 New Molecular Entity.

Only submission_status: AP is an approval; CR is a complete response letter — a rejection.

Labels

python skills/openfda/scripts/fda_labels.py boxed --drug metformin --limit 5
python skills/openfda/scripts/fda_labels.py section --drug atorvastatin --section mechanism_of_action
python skills/openfda/scripts/fda_labels.py classes --drug apixaban

classes returns the Established Pharmacologic Class and mechanism annotations (Factor Xa Inhibitor [EPC] / Factor Xa Inhibitors [MoA] for apixaban).

Two traps. Every manufacturer files its own SPL, so a generic drug has hundreds of near-identical records and the scripts de-duplicate before showing text. And openfda.* annotations are frequently missing entirely — checked live, every atorvastatin label examined carried no pharm_class_epc at all, while pembrolizumab and apixaban carry them. Absence is a property of that SPL, not of the drug. The same applies to boxed_warning: its absence from a generic label is not evidence the drug has no boxed warning.

Four ways this API misleads quietly

1. A report is not a patient, and never an exposure. FAERS counts submissions. A report naming five drugs appears under all five. Nothing here is an incidence or a risk.
2. A drug search matches concomitant medications. drugcharacterization is 1 suspect, 2 concomitant, 3 interacting. Without filtering, you are counting reports where the drug was merely present in the patient.
3. Data lags by a quarter or more. meta.last_updated was 2026-07-30 here. A drug approved in the last two quarters has essentially no FAERS data, so absence of signal means nothing.
4. Notoriety drives reporting. Publicity, litigation, and safety communications raise report counts with no change in biology. Disproportionality measures reporting behaviour.

When to stop using the API

For anything beyond triage — shrinkage estimators (BCPNN, MGPS/EBGM), stratification by age and sex, or multi-drug adjustment — download the quarterly FAERS extract instead. Those methods estimate a prior across the whole contingency space at once and cannot be built from per-pair queries. See references/disproportionality.md.

Composing with the rest of the bundle

• chembl → here: mechanisms and indications for an approved drug, then its real-world safety.
• clinicaltrials → alongside: what is being tested now, against what has already been approved.
• open-targets → here: the known drugs for a target, checked against their post-market record.
• target-safety → alongside: human genetic evidence for the same liability, before it is a drug.
• pkpd-translation → after: an approved label's dose and exposure as a translation anchor.

Reporting results honestly

Quote a beside every ratio and give the ROR with its confidence interval — a PRR of 12 built on four reports is noise. Name the data cut. Say "reports mention this event disproportionately", never "the drug causes" or "the risk is". Nothing here is validated for clinical use, and the API's own disclaimer says so.