What Changed at Insilico Medicine

Insilico released a set of small language models trained as scientific specialists for chemistry and biology through its MMAI Gym for Science framework. The set covers chemical synthesis, ADMET prediction and potency prediction across GPCR and kinase panels, and includes a single step retrosynthesis model built on Liquid AI's 2.6 billion parameter architecture.

Our read

Naming the base architecture and the parameter count is unusual in this category and is the reason this is graded on model transparency rather than capability alone. For discovery teams the practical read is that these are narrow specialists rather than a general assistant, which is the right shape for ADMET and potency work where a confident wrong answer is expensive.

Insilico Medicine convened the Open Consortium for Benchmark Quality in AI Driven Drug Discovery, known as O3DC, and published a live catalogue of the benchmarks the field uses to claim performance. The catalogue tracks each benchmark's metrics, maintainers and repository activity, and unusually it documents the known caveats, biases and limitations of each one rather than presenting them as neutral yardsticks. The stated premise is that a benchmark with an undisclosed bias produces model comparisons that look rigorous and are not.

Our read

Pharmaceutical evaluators comparing AI drug discovery platforms are almost always comparing benchmark scores, and this is the first centralized attempt to say which of those benchmarks can carry the weight. It is worth noting that the vendor convening a benchmark integrity consortium is also a vendor whose own models are scored on those benchmarks, so the catalogue is useful and is not disinterested. Read it as a map of where the measurement problems are rather than as an independent audit.

Insilico Medicine published a collaborative study in npj Precision Oncology demonstrating the use of its PandaOmics AI platform to identify therapeutic targets for inverted papilloma-associated sinonasal squamous cell carcinoma. The platform integrated transcriptomic data with pathway biology and protein interaction networks to map the molecular cascade of the disease.

Our read

This peer-reviewed publication provides concrete validation of the platform's ability to discover actionable targets in rare and poorly understood cancers. Buyers evaluating the software can use this evidence to assess its utility in complex multi-omic analysis and translational research.