Abstract

AI-assisted scientific discovery has produced compelling computational results — but independent laboratory validation at scale has remained limited. This paper presents a third-party benchmark study of Synth's prediction methodology across three domains: small-molecule drug candidates, heterogeneous catalysts, and protein variants.

Conducted across 12 partner institutions, the study evaluated 2,400 predictions against experimental outcomes. Our findings establish a 95% average prediction accuracy across domains, with full methodology, failure case analysis, and domain-specific performance breakdowns included.

Key Findings

01

95% average prediction accuracy across all three scientific domains — validated independently, not self-reported

02

Discovery timelines reduced from an average of 7.2 years to 18 months across participating partner programs

03

Explainability layer (SHAP attribution) confirmed to match experimental reasoning in 91% of cases reviewed by independent scientists

04

Failure analysis included — the 5% of cases where predictions were incorrect and what we learned from each

AI Prediction Benchmark Study Across Three Scientific Domains

📄 PDF · 24 pages · 2.1 MB

🏛️ ACS Central Science · 2024

📊 847 citations to date