The Pulse
Insilico Opens 328-Gene AI Toolkit for Longevity Research
Insilico Medicine has released LongevityBench, five specialized language models and the Longevity Claw research platform in a Cell study. The system nominated 328 genes across 14 hallmarks of aging, while the company says its best model ran
AI.info Team ·
Longevity Claw nominated 328 genes as possible targets for aging interventions, giving Insilico Medicine a concrete test of its new open AI research system beyond model benchmarks. The company disclosed the result on September 17, 2026, as it released a three-part toolkit in a Cell study covering evaluation, specialized language models and automated research workflows.
The release combines LongevityBench, five compact Longevity Large Language Models and Longevity Claw, an open-source platform designed to connect a domain-specific model with biological analysis tools. Insilico developed the work with researchers from Liquid AI, the Buck Institute for Research on Aging, Harvard Medical School and Brigham and Women’s Hospital.
LongevityBench Tests Five Kinds of Aging Data
LongevityBench evaluates AI systems across clinical data, genetics, epigenetics, transcriptomics and proteomics. The researchers designed the benchmark to make simple recall less useful, requiring systems to interpret biological measurements and produce answers tied to aging-related evidence.
The study assessed 18 frontier AI systems from six developer groups, including OpenAI, Google, Anthropic, xAI, DeepSeek and Moonshot AI. Insilico says no single frontier model performed best across all five biological domains. Results also changed substantially when researchers rephrased questions, a sign that general-purpose systems can produce inconsistent answers when working with specialized aging data.
Predicting biological age directly from omics measurements proved particularly difficult. The finding limits what can be inferred from model size alone: the largest systems did not consistently deliver the strongest biological reasoning.
A 0.6-Billion-Parameter Model Joins the Test
Insilico and its collaborators trained five open-source Longevity-LLMs on aging-specific clinical and multi-omics data. The models range from 0.6 billion to 9 billion parameters and use open architectures from Liquid AI and Alibaba, including Liquid AI’s LFM2 and Alibaba’s Qwen3 and Qwen3.5 families.
According to the study, the specialized models matched or exceeded the 16 frontier systems included in the comparison. The strongest model, L-Qwen3.5-9B, recorded the highest overall score among the 26 systems tested, including Google’s Gemini 3.1-Pro. Insilico also says the smallest model outperformed most of the frontier systems.
The result points to a practical tradeoff for research groups. A smaller model trained on carefully selected biological data may require less computing capacity and support deployment on local infrastructure, while still handling tasks that general-purpose systems find difficult. The benchmark results do not establish that the models can make clinical decisions or replace experimental validation.
Longevity Claw Turns Model Answers Into Workflows
Insilico embedded L-Qwen3.5-9B into Longevity Claw, which combines the language model with gene-set enrichment analysis, biological aging-clock calculations, population-level profiling, evidence retrieval and candidate-target prioritization.
The platform is designed to execute multi-step research workflows rather than respond only to individual prompts. It can gather evidence, run specialized analyses, inspect intermediate findings and assemble a case for a biological target. The researchers used it across 14 recognized hallmarks of aging and reported 328 nominated genes.
Those candidates showed enrichment of up to 5.6-fold against an independently published reference set of experimentally supported aging-related targets. The result supports the biological relevance of the output, but it does not demonstrate that the genes are effective drug targets in humans. One nominated gene, KDM1A, also appeared in a separate published study in which modulation extended the lifespan of C. elegans.
Zhavoronkov Frames the Release as Open Infrastructure
Insilico is releasing the benchmark, model collection, training resources, evaluation code and Longevity Claw platform for independent testing and further development. The model collection is available through Hugging Face, while the platform’s code is published in the Longevity Claw GitHub repository. A public results page is available at LongevityBenchmarks.org.
“The longevity community is moving beyond static aging clocks toward foundation models capable of generating measurable, actionable insights.”
Alex Zhavoronkov, founder and co-CEO, Insilico Medicine
Zhavoronkov said the company is developing benchmarked, agentic systems that could eventually support personalized health monitoring. The current publication, however, describes research infrastructure and computational results rather than a consumer health product or a proven longevity treatment.
From Rentosertib to an Open Research Stack
The Cell publication follows a September 7, 2026 study in Nature Biotechnology on rentosertib, Insilico’s AI-discovered and AI-designed candidate for idiopathic pulmonary fibrosis. The company said the Phase IIa trial analysis found reduced biological age across six independent proteomic aging clocks.
Those findings and the new toolkit address different questions. Rentosertib provides clinical-trial evidence that a drug candidate can affect biological aging signatures, while LongevityBench and Longevity Claw aim to help researchers identify and rank future targets. Neither result establishes that an intervention can extend human lifespan.
The immediate test for Insilico’s release will be whether independent researchers reproduce its benchmark rankings, inspect the nominated genes and determine how often the platform’s suggestions survive laboratory experiments. For now, the company has put a 328-gene output, five specialized models and an open evaluation framework in the public record.