Boston, September 23 — Insilico Medicine has released an open-source artificial intelligence toolkit aimed at speeding up research into ageing and longevity, publishing the work in Cell, where it was chosen as the journal’s cover feature. The clinical-stage biotechnology company conducted the study with scientists from Liquid AI, the Buck Institute for Research on Aging, Harvard Medical School and Brigham and Women’s Hospital.
At the centre of it is LongevityBench, an open benchmark built to test whether AI systems can reason across different types of biological data linked to human ageing — clinical data, genetics, epigenetics, transcriptomics and proteomics.
Why a new benchmark was needed
The researchers argue that conventional AI evaluations can reward a model for recalling information rather than for interpreting new biological measurements. Think of it like a student who has memorised the textbook but freezes when handed an unfamiliar lab result. LongevityBench was designed to test evidence-based reasoning across multiple biological domains instead.
The team put 18 leading frontier AI systems to the test, from companies including OpenAI, Google, Anthropic, xAI, DeepSeek and Moonshot AI. No single frontier model produced the strongest results across all five biological data types, and performance varied substantially depending on how questions were phrased.
So Insilico built its own. The researchers developed five compact open-source Longevity Large Language Models, ranging from 0.6 billion to nine billion parameters, trained on ageing-specific clinical and multi-omics data using Insilico’s MMAI Gym for Science framework. According to the study, the specialised models matched or exceeded the 16 frontier systems evaluated on LongevityBench.
The L-Qwen3.5-9B model recorded the highest overall score among all 26 systems tested, including outperforming Google’s Gemini 3.1-Pro.
From benchmark to autonomous discovery
Insilico also used the L-Qwen3.5-9B model to power Longevity Claw, an open-source agentic platform built to carry out multi-step research workflows. The platform pairs a specialised language model with tools for gene-set enrichment analysis, biological ageing-clock calculations, population-level profiling, evidence retrieval and synthesis, and candidate target evaluation. In its first large-scale autonomous discovery campaign, Longevity Claw analysed 14 recognised hallmarks of ageing and nominated 328 genes as potential therapeutic targets.
The researchers said the candidates showed statistically significant enrichment of up to 5.6-fold against an independently published set of experimentally supported ageing targets. One nominated gene, KDM1A, was independently identified in a separate published study as a potential ageing and cancer target; that research found modulation of KDM1A extended lifespan in C. elegans.
Dr Alex Zhavoronkov, Founder and Co-CEO of Insilico Medicine, framed the release as an attempt to widen access to the underlying tools. “Our work with rentosertib demonstrated that an AI-discovered and AI-designed drug can influence biological ageing signatures in patients,” he said.
“Now, we are opening the tools that can help the global scientific community discover the next generation of longevity therapeutics.” Zhavoronkov added that LongevityBench “establishes a rigorous standard for measuring whether AI can truly reason about ageing biology, while our specialised models and Longevity Claw turn that intelligence into an autonomous discovery engine.” Insilico is releasing the benchmark, the specialised models, training resources, evaluation code and the Longevity Claw platform so researchers can test and develop them independently.
The practical upshot for the field is that the scoring rubric and the models are now in the open, which means outside labs can try to reproduce the results — or beat them. Watch for independent groups to run their own evaluations against LongevityBench, and for follow-up work on the nominated gene targets, including KDM1A.
