Insilico Medicine and Liquid AI Publish an Aging Benchmark in Cell: Open Model Outperforms Closed Flagships
Insilico Medicine and Liquid AI Publish an Aging Benchmark in Cell: Open Model Outperforms Closed Flagships
The team has publicly released LongevityBench, a test of AI comprehension of aging-related biological data, along with five compact models and the Longevity Claw program. Vadim Gladyshev of Harvard Medical School is a co-author.
Aging clocks estimate biological age from a single data type, such as blood tests alone or DNA methylation alone. Until now, no benchmark existed to test whether an AI model can reason across multiple types of aging data at once.
Earlier neural networks built for this kind of analysis required a specialized architecture for each measurement type. A different approach grew out of MMAI's Gym for Science platform, which had previously shown that a language model fine-tuned on structured data performs well in drug discovery. The authors set out to test the same strategy on aging: represent any measurement as text and train a single model on all of them.
LongevityBench comprises 17 tasks spanning five types of biological data, from clinical records and genetics to DNA methylation and blood proteins. All 25,457 questions were generated programmatically from raw measurements, so the answers cannot be found ready-made in training corpora. A test of 18 models from OpenAI, Google, Anthropic, xAI, DeepSeek, and Moonshot AI showed that predicting age from molecular profiles was the hardest category. Gemini 3.1 Pro led the ranking, though it was not statistically distinguishable from four other systems.
The authors then fine-tuned five open models on aging data, ranging from 0.6 to 9 billion parameters, on Qwen architectures from Alibaba and LFM2 from Liquid AI, a developer of compact models. The 9-billion-parameter L-Qwen3.5 took first place: its confidence interval does not overlap with any commercial model. The smallest version, at 0.6 billion parameters, placed sixth out of 26, ahead of Claude Opus 4.5, Kimi K2.5, GPT-5.2, and Grok 4.3. A control test on published factual knowledge about aging produced the opposite result: large general-purpose models won. This means the compact models' advantage comes from the ability to parse raw measurements, not from broader exposure to the literature.
A sensitivity analysis confirmed reliance on real biomarkers: removing the NHANES blood-test block changed 47.4% of correct answers.
The best-performing model was integrated into the Longevity Claw program and tasked with identifying therapeutic targets across 14 hallmarks of aging. It proposed 328 candidate genes whose composition overlapped significantly with the target list from PandaOmics, another AI platform by Insilico. The gene KDM1A (the enzyme LSD1) was selected in five out of six runs. This target had been validated experimentally: its inhibition extends lifespan in roundworms by 15.8%. Two days before the Cell publication, an independent study in the Journal of Clinical Investigation showed that an inhibitor of the same enzyme reduced signs of aging in the kidneys, heart, and liver of mice.
"The longevity community is moving beyond static aging clocks toward models capable of producing measurable, actionable insights," said Insilico Medicine CEO Alex Zhavoronkov in a press release.
The paper appeared ten days after Nature Biotechnology reported that Insilico's AI-discovered drug rentosertib reduced biological age in patients across six aging clocks in a phase IIa trial. In the January preprint, Gemini 3 Pro led LongevityBench; in the September version published in Cell, a compact open model took the lead. The benchmark, weights for all five models, and Longevity Claw source code are now publicly available, so any laboratory can verify or reproduce the results.