Phase 3 Outcome Predictor

Eight feature categories

What the model looks at.

Category	Example signals
Efficacy signal	Effect size in phase 2, biomarker shift, response durability at last follow-up
Expression durability	Transgene expression slope at 6/12/24 mo, integration site analysis
Safety	SAE rate, dose-limiting toxicities, immune-mediated hepatotoxicity signal
Dose selection	MTD vs phase-3 dose, dose-response slope, headroom to toxicity
Trial design	Endpoint choice, control arm, alpha-spending plan, sample size adequacy
Capsid biology	BBB score (if CNS), tissue tropism match, seroprev-adjusted eligible pool
Immune response	Neutralizing antibody seroprev, T-cell response in dose-escalation, complement activation
Regulatory	FDA/EMA written agreements, accelerated approval eligibility, RMAT/PRIME status

Performance

AUC 0.82+ on held-out trials.

Validation set: The model is trained on the historical AAV gene-therapy trial registry (n=20+ trials with phase-1/2 → phase-3 outcomes known). Held-out validation maintains AUC 0.82+ with proper temporal split (no future-data leakage).

Calibration is the more important property than raw AUC. The output is a probability with explicit confidence intervals — not a binary "will succeed" classification. Pre-registration with SHA-256 locking applies before any phase-3 readout we predict against.

Honest limits

What it can't predict.

Manufacturing failures — not modeled. CMC issues that halt phase-3 are out-of-scope.
Competitive landscape changes — a competitor's late phase-3 readout that re-defines standard of care is not modeled.
Sponsor-specific risk — small biotech vs large pharma execution differences are not in the feature set.
Black-swan safety — an unanticipated rare-but-fatal event surfacing in phase-3 (e.g., capsid-related thrombotic microangiopathy at scale) by definition isn't in the training set.

How to use

For sponsors and investors.

The predictor outputs a probability + the top three risk-driving features for each candidate program. Sponsors use it to identify which phase-1/2 signals to strengthen before locking phase-3 design. Investors use it as a structured second-opinion on consensus-level analyst forecasts. The methodology applies equally to small-molecule combinations under Pillar 2 — though the feature set differs.

Forecasting a phase-3 readout?

Send the trial NCT or sponsor-disclosed phase-2 data. We return a calibrated probability with the three risk-driving features explicit.