How Trialytic Works

Every study registered on ClinicalTrials.gov, standardised and linked to the published literature and to FDA post-market data. Refreshed monthly. Sources and attribution are on the About page.

• ›Each trial holds two records: the protocol (what was planned — design, eligibility, endpoints, sites) and the results (what happened — participant flow, outcomes, adverse events).
• ›Condition and intervention names are mapped to MeSH, so 'NSCLC' and 'Non-Small Cell Lung Cancer' resolve to one disease. Drug names are normalised across INN and USAN, so 'paracetamol' and 'acetaminophen' resolve to one molecule.

Meta-analysis — DerSimonian–Laird random effects

Each study's standard error is back-calculated from its reported 95% CI, studies are weighted by precision, and between-study heterogeneity (τ²) is estimated and added to the weights rather than assumed away.

SE_i = (CI_upper − CI_lower) / (2 × 1.96)     v_i = SE_i²     w_i = 1 / (v_i + τ²)

θ_RE = Σ(w_i × θ_i) / Σ(w_i)          I² = max(0, (Q − df) / Q) × 100%

I² is the share of variation that is real disagreement rather than sampling noise. Above ~50% the forest plot matters more than the pooled diamond. Publication-bias risk uses Egger's test, which needs about ten studies to mean anything.

Trial outcome predictor

Gradient-boosted trees (XGBoost) over 31 protocol and sponsor features, trained on trials that have resolved. The label is COMPLETED versus TERMINATED — not efficacy, not approval. Sponsor and indication history enter as out-of-fold, shrunk rates, so a sponsor with two trials is not treated like one with 500. Every prediction carries SHAP attributions. A second model returns expected duration as a range, because its typical error is around ten months and a single figure would imply precision it does not have.

FDA approval outlook

For a novel molecule already in Phase 3 and not yet approved, a modelled probability of eventual FDA approval against a 44.5% base rate. Every input is clipped to what was knowable at the molecule's first Phase 3 — an unclipped version scored far higher by reading the future, since approved drugs go on to run more trials and failed ones stop. Temporally validated: trained only on the past, tested on later programmes. It refuses to answer for molecules already approved, where the outcome is known.

Post-market safety signals

Adverse-event reports are counted per drug and reaction, then scored by disproportionality — whether a reaction is reported more often for this drug than across the whole reporting system.

PRR = [a / (a + b)] / [c / (c + d)]     signal when PRR ≥ 2, a ≥ 3 and χ² ≥ 4

Everything the site does is an HTTP API, and it is the same API the site itself calls. Point Claude, a notebook or a script at it and you can work through 598,000 trials, the FDA drug and device layer, sponsors and targets without opening the app. Reading these docs needs no account; issuing a key is part of Pro.

Authentication

Create a key on your account page, then send it as a bearer token. Keys start with tk_ and are shown once — only a hash is stored, so a lost key is replaced rather than recovered. A key authenticates as your account, and it can be revoked without touching your password or your browser sessions.

curl -H "Authorization: Bearer tk_your_key_here" \
  "https://trialytic.ai/api/v1/landscape/search?condition=psoriasis&phase=PHASE3&limit=100"

What you can call

Every endpoint accepts the same filters as the equivalent page in the app, and returns the same JSON that page renders — so anything you can see in the product you can fetch.

Rate limits and paging

The daily budget is a rolling 24-hour window, not a calendar reset, so a job that starts late in the evening does not lose its allowance at midnight. Over budget returns 429 with the time the window frees up.

Errors

Using it with Claude

The API was built to be driven by an assistant: predictable paths, JSON responses, no session state, and errors that say what to do next. Hand Claude a key and this page and it can answer questions across the whole corpus without any glue code.

# Every phase 3 trial for a target, then the sponsors behind them
import requests
H = {"Authorization": "Bearer tk_your_key_here"}
B = "https://trialytic.ai/api/v1"

trials = requests.get(f"{B}/landscape/search",
                      params={"target": "TIGIT", "phase": "PHASE3", "limit": 100},
                      headers=H).json()
print(trials["total"], "trials")

summary = requests.get(f"{B}/landscape/summary",
                       params={"target": "TIGIT"}, headers=H).json()
for s in summary["top_sponsors"][:5]:
    print(s["sponsor_name"], s["n_trials"])