Loading Curie Briefs...
Getting the latest healthcare news for you
Getting the latest healthcare news for you
Insilico Medicine has unveiled a suite of AI "specialist" models trained specifically for chemistry and biology tasks in drug discovery. Built through its MMAI Gym framework, these small language models now match or beat dedicated scientific tools across more than 70 benchmark tasks — from predicting drug safety profiles to mapping chemical synthesis routes. It's a significant step forward for AI's role in pharmaceutical R&D.
Insilico Medicine has launched a new series of AI "specialist" models through its MMAI Gym for Science framework, designed to tackle some of the hardest problems in drug discovery. Unlike general-purpose AI tools, these are small language models fine-tuned for specific scientific tasks — think predicting how a drug is absorbed, metabolized, or whether it'll interact dangerously with other compounds. The headline finding? On many tasks, they're not just keeping pace with traditional computational tools — they're outperforming them.
The models cover three major chemistry domains — ADMET (drug safety/pharmacokinetics) prediction, target activity prediction for GPCRs and kinases, and chemical synthesis planning — plus biology-focused models spanning clinical, omics, and molecular data. Notably, the retrosynthesis models, built on a compact 2.6B-parameter architecture, outperform leading dedicated methods on both standard and more challenging out-of-distribution benchmarks.
By the numbers:
Why it matters: Drug discovery is notoriously slow and expensive. AI models that can reliably predict drug safety, potency, and synthesis pathways earlier in the pipeline could significantly cut costs and development timelines — ultimately helping get better treatments to patients faster.
