mottledMantis/cto-drug-approval-baseline

CTO Drug Approval Prediction Baseline

A pre-trial LightGBM classifier trained on the CTO dataset to predict clinical trial success outcomes from structured trial metadata alone.

What's honest about this model

• Uses ONLY features known at trial registration time (phase, enrollment, sponsor type, DMC presence, title keywords)
• Temporal split: trained on trials registered before 2020, tested on 2020+
• No leakage from completion status or post-trial results
• No LLM reasoning — pure structured features + gradient boosting

Performance (temporal holdout)

Metric	Value
ROC-AUC	0.869
PR-AUC	0.640
Brier	0.141

Test set: 4,359 trials registered 2020+, 24.6% success rate.

Feature importance (top 10)

1. title_length — proxy for trial complexity
2. enrollment_log — trial size
3. study_year — temporal drift in success rates
4. num_arms — design complexity
5. phase_num — later phases have higher success rates
6. is_fda_regulated — regulated trials differ
7. has_dmc — data monitoring committee presence
8. is_industry — industry vs academic sponsor
9. cond_cancer — oncology trials
10. cond_alzheimer — neurodegenerative indication

Caveats

• Labels are weakly supervised (derived from trial linkage + status, not direct FDA approval tracking)
• Title-based condition classification is crude
• No molecular features, no competitive landscape, no regulatory designations
• This is a research baseline, not investment advice

Citation

Gao et al. "Automatically Labeling Clinical Trial Outcomes: A Large-Scale Benchmark" (arXiv:2406.10292)