英矽智能发布业界首个“药物研发基准评估”
Ranking
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - Insilico Medicine (HKEX: 3696) launched "DDD Benchmarks," a standardized evaluation framework claiming to be the industry's first drug-discovery benchmark suite, plus an accompanying Benchmark-as-a-Service (BaaS) offering and public leaderboard. It targets data contamination in existing public AI evaluations, aiming to test whether frontier models can make real drug R&D decisions rather than just score well on memorized test sets.
- Two complementary suites: Drug Discovery Foundations (300+ tests drawn from Insilico's proprietary out-of-distribution (OOD) datasets and deep-cleaned public data, covering disease biology, molecular property prediction/optimization, retrosynthesis route design, structure-based molecular design, and clinical development) and Drug Candidate Essentials (end-to-end project capability from hit screening to preclinical candidate (PCC) nomination, benchmarked against Insilico's own validated internal programs).
- Any model exposing a standard Chat Completions API can be evaluated; outputs are scored against expert-level reference data, returned as a standardized scorecard with head-to-head comparisons, and — with customer consent — published to a public leaderboard.
- Grounding claims cited: 31 PCCs nominated since 2021, 13 IND approvals, and Rentosertib (ISM001-055) in Phase III for IPF; average 12–18 months to PCC nomination vs. a stated industry norm of 2.5–4 years, typically synthesizing/testing only 60–200 molecules per program.
- The framework derives from Insilico's Pharma.AI platform and MMAI Gym training ecosystem; service is live at dddbench.insilico.com. Note: all performance figures are company-reported, with no third-party validation or benchmark results disclosed in this announcement.
Sources (1)
英矽智能发布业界首个“药物研发基准评估”
TL;DR - Insilico Medicine (HKEX: 3696) launched "DDD Benchmarks," a standardized evaluation framework claiming to be the industry's first drug-discovery benchmark suite, plus an accompanying Benchmark-as-a-Service (BaaS) offering and public leaderboard. It targets data contamination in existing public AI evaluations, aiming to test whether frontier models can make real drug R&D decisions rather than just score well on memorized test sets.
- Two complementary suites: Drug Discovery Foundations (300+ tests drawn from Insilico's proprietary out-of-distribution (OOD) datasets and deep-cleaned public data, covering disease biology, molecular property prediction/optimization, retrosynthesis route design, structure-based molecular design, and clinical development) and Drug Candidate Essentials (end-to-end project capability from hit screening to preclinical candidate (PCC) nomination, benchmarked against Insilico's own validated internal programs).
- Any model exposing a standard Chat Completions API can be evaluated; outputs are scored against expert-level reference data, returned as a standardized scorecard with head-to-head comparisons, and — with customer consent — published to a public leaderboard.
- Grounding claims cited: 31 PCCs nominated since 2021, 13 IND approvals, and Rentosertib (ISM001-055) in Phase III for IPF; average 12–18 months to PCC nomination vs. a stated industry norm of 2.5–4 years, typically synthesizing/testing only 60–200 molecules per program.
- The framework derives from Insilico's Pharma.AI platform and MMAI Gym training ecosystem; service is live at dddbench.insilico.com. Note: all performance figures are company-reported, with no third-party validation or benchmark results disclosed in this announcement.