Insilico Medicine’s MMAI Gym for Science suite delivers state-of-the-art or superior performance across chemistry and biology on more than 70 benchmark tasks evaluated by the DDD Bench, highlighting broad capabilities in drug discovery.

Frontier specialist AI models for chemistry and biology under MMAI Gym achieve top performance on over 50 benchmark tasks, underscoring rapid advances in AI-driven science.

In the first half of 2026, Insilico reported about $106 million in revenue, a 287% year-over-year increase, profitability since listing, and roughly $7.3 billion in confirmed contract value from collaborations, with cumulative major partnerships around $11 billion.

Two GPCR- and kinase-focused models reach state-of-the-art IC50 predictions, supporting workflows from virtual screening to selectivity profiling.

Chemical synthesis models specialize in single-step retrosynthesis using Liquid AI’s 2.6B-parameter architecture, outperforming leading methods and setting the stage for updates in the Microsoft Marketplace.

Rentosertib (ISM001-055) becomes the world’s first drug candidate discovered with generative AI to enter Phase III for idiopathic pulmonary fibrosis, illustrating AI-driven milestones alongside broad industry partnerships.

MMAI Gym debuted at NeurIPS 2025, with notable progress reported at ICLR 2026, ICML 2026, and EMNLP 2026, showing competitive results against frontier models and traditional chemistry methods.

Chemistry models cover synthesis, ADMET prediction, and target activity for GPCRs and kinases, with ADMET achieving strong results on 28 tasks and notable performance on the Drug Candidate Essentials benchmark.

MMAI Gym seeks to turn language models into domain-specific scientific specialists by training across related tasks and benchmarking head-to-head against established methods rather than relying on general knowledge alone.

Biology-focused MMAI Gym models extend the same category-based framework to clinical, omics, and molecular biology tasks, delivering state-of-the-art results on selected benchmarks.