Aggregated Evidence

 Quick Analysis Plan Copied

 Long Analysis Plan

Meta-Analysis Plan: Does AI Accelerate Drug Discovery?

Core quantitative evidence. The only systematic review in the supplied corpus (PRISMA 2020, 173 studies, 2015–2025) reports: 100% of included studies demonstrated some timeline impact; 45% (77/173) reported clinical outcomes; 97% (168/173) reported industry partnerships; 39.3% remained preclinical; only 23.1% reached Phase I and <5% reached Phase IIb or beyond PRISMA review of AI pipeline stages and outcomes. Therapeutic focus is heavily skewed: oncology 72.8%, dermatology 5.8%, gastroenterology 5.2%, neurology 5.2% Therapeutic area distribution in AI drug discovery. Geographic concentration: US 72.1%, China 8.9%, UK 5.0% Geographic distribution of AI drug discovery. Supporting narrative reviews concur on mechanisms (de novo design, target ID, ADME/tox prediction) but flag data-quality bias and animal-to-human translational gaps as unresolved Narrative review of AI in drug discovery. One rare primary-data study (n=101 heart transplant patients) shows AI (evolutionary algorithms) predicted cyclosporine levels with test correlation r=0.85 and 80–90% accuracy — a concrete, testable clinical-pharmacology example AI cyclosporine dosing prediction study.

Analysis Plan

1. Effect definition: pre-register timeline acceleration as primary outcome (months from target ID → IND), clinical success rate as secondary. 2. Data extraction: code the 173-study corpus for stage, method (ML 40.9%, molecular simulation 20.7%, DL 10.3%), therapeutic area, geography, and reported KPIs AI method frequency distribution. 3. Bias assessment: apply ROBINS-I; explicitly model publication bias (funnel plots/Egger) — author-acknowledged blindspots include reliance on company disclosures, US concentration, and overemphasis on success stories. 4. Sensitivity analysis: stratify by oncology vs non-oncology and industry-funded vs independent; the corpus shows domain skew is severe, so pooled effects without stratification are uninterpretable. 5. Falsification: if independent prospective studies show AI candidates fail to advance or underperform conventional pipelines with appropriate controls, the acceleration claim collapses Falsification criteria for AI acceleration claims. Key blindspot: 55% of studies lack clinical outcomes entirely; 'timeline impact' was reported by all 173 studies but likely conflates proprietary claims with measured acceleration. No randomized comparison of AI vs conventional pipelines exists in this evidence base.