Data Sources
Public data sources powering the Oncology Failure Atlas.
Production dataset currently contains 11,525 terminated and withdrawn oncology trial records.
Primary Source — ClinicalTrials.gov
ClinicalTrials.gov is the primary source for trial metadata, status, phase, sponsor, intervention, dates, indication, and reported discontinuation language. The dataset is built from publicly registered oncology trial records with status Terminated or Withdrawn, identified by their public NCT ID.
ClinicalTrials.gov stop-reason language — primarily the why_stopped field — is the primary input for MDIP (Mapped Discontinuation Intelligence Patterns) classification where available. Missing or ambiguous stop-reason language is explicitly tracked through dedicated unclear categories (UNCLEAR_NOT_SPECIFIED when no reason is reported, and UNCLEAR_WHY_STOPPED_REPORTED when a reason exists but is too ambiguous to map). Evidence layers support human review but do not prove causality or final root cause.
Scope of Data Used
The Oncology Failure Atlas uses public data only. No private, patient-level, confidential, or proprietary data is ingested, processed, or displayed.
PubMed
PubMed has been recalculated across the full 11,525-record production dataset. Each trial has a PubMed coverage row.
- DIRECT_NCT — direct NCT ID match.
- TITLE_SIMILARITY — title-similarity fallback match.
- NONE — no matching PubMed signal identified by the current process.
NONE does not prove no publication exists. It means no linked result was found by the current matching process.
OpenFDA
OpenFDA has been recalculated across the full 11,525-record production dataset. Each trial has an OpenFDA coverage row.
- EXACT — exact drug name match.
- FUZZY — fuzzy drug-name match.
- NO_MATCH — no matching OpenFDA signal identified by the current process.
NO_MATCH does not prove no safety or regulatory evidence exists. It means no linked OpenFDA result was found by the current matching process.
CrossRef
CrossRef has been expanded as a supplementary positive DOI-link layer. It identifies additional DOI-to-NCT links where available.
CrossRef is not treated as a complete publication-absence layer. Absence of a CrossRef link does not mean no publication exists.
RecurSignal™ (deterministic layer)
RecurSignal™ is a deterministic historical recurrence signal computed across all 11,525 OFA records from MDIP™ category, trial phase, and classification confidence. It is a layer over the public-source data — not an external source itself.
RecurSignal is not a prediction, probability, forecast, clinical risk score, investment signal, treatment guidance, or asset-quality score.
Evidence does not mean proof
Linked evidence in OFA — PubMed, OpenFDA, CrossRef, and RecurSignal — provides navigational signal. It does not establish:
- causality between an intervention and a discontinuation;
- completeness of the underlying public source for that trial;
- clinical validation, regulatory determination, or safety adjudication;
- investment diligence on the sponsor, asset, or program.
"No linked PubMed / OpenFDA / CrossRef signal currently available in OFA" means the current OFA matching process found no link. It does not mean no publication, safety record, or regulatory action exists.
OFA Research Brief Generator — evidence pack input
The OFA Research Brief Generator (Claude-assisted, server-side via the claude-research-hub Supabase Edge Function) does not pull from new sources. It consumes an evidence pack built entirely from the OFA-filtered data described above: active filters, record counts, top MDIP™ patterns, top harmonized mechanisms, phase distribution, linked-signal availability, a small set of representative records, and known limitations.
Source Status & Evidence Alignment
OFA uses public oncology discontinuation records as the trial spine, then cross-checks each record against compact, source-linked public evidence layers. This section shows what each source contributes, how OFA taps it, where coverage exists, and what each source does not prove.
ClinicalTrials.gov / AACT
?- Status:
- Structured registry enrichment complete
- Coverage:
- 11,525 / 11,525 OFA records matched
- How OFA uses it:
- AACT is used as structured ClinicalTrials.gov-derived registry context for OFA's 11,525 discontinued oncology trials. OFA uses selected compact fields such as enrollment, eligibility presence, outcome counts, drop/withdrawal availability, and reported-event availability.
- What it helps answer:
- Does the public registry record include structured trial-design context?
- What it does not prove:
- It does not establish the true cause of discontinuation, clinical safety conclusions, treatment advice, or asset quality.
- Known gap:
- Reported-event availability is a registry-data signal only. OFA does not expose full reported-event payloads in public UI surfaces.
PubMed
?- Status:
- Linked publication signal layer
- Coverage:
- 1,493 / 11,525 records with PubMed signal
- How OFA uses it:
- OFA stores compact PubMed linkage fields such as PMID, DOI, publication year, journal, match type, and abstract availability where available.
- What it helps answer:
- Is there a publication-linked signal associated with this trial or record?
- What it does not prove:
- Missing PubMed signal does not prove that no publication exists.
- Known gap:
- OFA is not a complete literature review.
OpenFDA
?- Status:
- Public FDA signal availability layer
- Coverage:
- 7,794 / 11,525 records with OpenFDA signal
- How OFA uses it:
- OFA uses OpenFDA as a compact public signal-availability layer based on matched drug names and summary counts.
- What it helps answer:
- Is there a public FDA-linked signal available for the matched drug name?
- What it does not prove:
- It does not provide a regulatory conclusion, causal interpretation, safety conclusion, or treatment advice.
- Known gap:
- Drug-name matching can be imperfect, and missing signal does not prove absence of FDA evidence.
CrossRef
?- Status:
- Supplementary DOI-link layer
- Coverage:
- 1,033 / 11,525 records with CrossRef DOI link
- How OFA uses it:
- OFA aggregates CrossRef links to one row per NCT ID and uses DOI/publication metadata as supplementary linkage evidence.
- What it helps answer:
- Is there DOI-linked publication metadata associated with this trial?
- What it does not prove:
- Missing CrossRef link does not mean no publication exists.
- Known gap:
- CrossRef is a positive link layer, not an absence-of-publication test.
MDIP
?- Status:
- Discontinuation-pattern taxonomy
- Coverage:
- 11,465 / 11,525 records with any MDIP signal
- How OFA uses it:
- MDIP maps public stop-reason language into comparable discontinuation-pattern categories.
- What it helps answer:
- What type of discontinuation pattern was reported in public registry language?
- What it does not prove:
- MDIP does not prove the true root cause of trial failure.
- Known gap:
- 60 records currently have no MDIP signal and should be displayed as not yet classified or uncategorized, not hidden.
Mechanism Harmonization
?- Status:
- Product-facing mechanism grouping complete
- Coverage:
- 11,525 / 11,525 records
- How OFA uses it:
- OFA uses a separate mechanism harmonization layer to group intervention mechanisms into comparable product-facing classes.
- What it helps answer:
- Which discontinued trials cluster by broad therapeutic mechanism?
- What it does not prove:
- It is not a definitive biological classification and should not be treated as mechanistic truth.
- Known gap:
- Some records remain broad or unknown because public intervention labels can be incomplete or ambiguous.
RecurSignal
?- Status:
- Historical recurrence/context signal complete
- Coverage:
- 11,525 / 11,525 records
- How OFA uses it:
- RecurSignal is a deterministic historical recurrence/context signal calculated offline and stored as a versioned evidence layer.
- What it helps answer:
- How often similar discontinuation-pattern contexts appear historically within OFA?
- What it does not prove:
- It is not a prediction, probability, risk score, clinical score, investment signal, or asset-quality rating.
- Known gap:
- It is only as current as the latest offline recomputation.
EMA EPAR context
?- Status:
- Backend complete; public-source context planned for controlled product exposure
- Coverage:
- 2,705 / 11,525 OFA records with EMA core context
- How OFA uses it:
- OFA now includes a controlled EMA EPAR backend context layer for matched oncology intervention names. The layer resolves public EMA human medicine pages, EPAR pages, and related regulatory-document context where exact medicine/intervention matching was available. This context is used as public regulatory-document availability context only.
- What it helps answer:
- Is public EMA human medicine / EPAR context available for the matched medicine name?
- What it does not prove:
- It does not mean the specific OFA trial was assessed by EMA, caused an EMA decision, or provides a safety, efficacy, regulatory, treatment, investment, or asset-quality conclusion.
- Known gap:
- Missing EMA context does not mean no EMA material exists. It only means OFA did not resolve a match through the current exact matching logic. Veterinary records and fuzzy matches were excluded. Full EPAR text and raw EMA JSON are not exposed.
How to interpret gaps
Source gaps are shown deliberately. A missing PubMed, OpenFDA, CrossRef, AACT subfield, or MDIP signal means OFA did not resolve that specific compact source signal for the record. It does not prove the evidence does not exist elsewhere, and it should not be interpreted as a clinical, regulatory, safety, or investment conclusion.
What OFA does not claim
OFA does not contain the full AACT database, the full PubMed database, or the full FDA database. OFA uses selected public evidence fields linked to the OFA oncology discontinuation trial set. The platform is a public-evidence navigation and research-intelligence layer, not a clinical decision-support system.
Source Coverage
| Layer | Coverage |
|---|---|
| ClinicalTrials.gov | 11,525 / 11,525 |
| PubMed coverage rows | 11,525 / 11,525 |
| OpenFDA coverage rows | 11,525 / 11,525 |
| CrossRef positive DOI links | 1,070 rows · 1,003 NCT IDs · 1,062 unique DOIs |
| RecurSignal scores | 11,525 / 11,525 |
| Trials with EMA core context (global) | 2,705 / 11,525 |
Update Note
Production dataset currently contains 11,525 terminated and withdrawn oncology trial records, deduplicated by NCT ID. PubMed, OpenFDA, CrossRef, and RecurSignal have all been recalculated across the full dataset.
Oncology Failure Atlas · Public data only · No clinical or investment advice.