Overview
On September 17, 2026, Insilico Medicine (HKEX: 3696) published a Cell cover article (Cell 189(19):5980–5994.e8; DOI 10.1016/j.cell.2026.08.026), *An open benchmark and language models for AI in aging biology*, with 13 authors including first and co-corresponding author Alex Zhavoronkov (Insilico founder and co-CEO). Co-authors include Vadim N. Gladyshev (Harvard Medical School / Buck Institute) and Fedor Galkin (Insilico AI team lead, co-corresponding author). Partner institutions include Liquid AI, the Buck Institute for Research on Aging, Harvard Medical School, and Brigham and Women's Hospital.
The paper open-sources three components:
- LongevityBench — a benchmark for AI reasoning on aging data
- Longevity-LLMs — a family of lightweight models from 0.6B to 9B parameters
- Longevity Claw — an agent platform that autonomously nominates intervention targets
- L-Qwen3.5-9B ranked #1 of 26 evaluated systems (composite score 4.4), outperforming Gemini-3.1-Pro (8.2), Claude Opus-4.6 (9.2), GPT-5.2, and Kimi-K2.5. Insilico's models took the top three spots (9B, 2.6B, 1.7B).
- DNA methylation paired-comparison: L-Qwen3.5-9B achieved 0.868 agreement vs. 0.685 for the best frontier model.
- Olink proteomic age prediction: L-Qwen3-0.6B MAE = 5.7 years vs. 10.1 years for the best frontier model (Kimi-K2.5).
- NHANES 10-year mortality: L-Qwen3-0.6B balanced accuracy 0.890.
- CellAge senescence phenotype control: frontier models scored near the theoretical ceiling, while L-LLMs averaged 71.4% — evidence of the benchmark's construct validity, distinguishing static knowledge recall from genuine multi-omic causal reasoning.
- Under a 'clock replication' setting (reproducing known published aging clocks), L-LLMs degraded sharply, showing the benchmark separates 'truly understanding' from 'copying'.
- Ablation study: 20 of 47 biological feature modules significantly affected L-Qwen3.5-9B's prediction confidence across 1,922 structured prompts, showing traceable, distributed cross-modal reasoning (blood glucose, renal function, sex hormone signaling, immune gene programs).
- L-Qwen3-0.6B, L-LFM2-1.2B, L-Qwen3-1.7B, L-LFM2-2.6B, L-Qwen3.5-9B
- Cell: DOI 10.1016/j.cell.2026.08.026 — 'An open benchmark and language models for AI in aging biology' (2026-09-17)
- insilico.com — 'Insilico Medicine Opens AI Longevity Discovery Toolkit to Researchers Worldwide in Cell Cover Study' (2026-09-17)
- unite.ai — 'Insilico Medicine Releases Open Longevity AI Toolkit in Cell Study' (2026-09-17)
- forbeschina.com — coverage of the 13% share-price move (2026-09-18)
Key results
The three components
LongevityBench
17 tasks across five data modalities: clinical (NHANES), genetics (SynergyAge, OpenGenes), epigenetics (GEO DNAm), transcriptomics (GTEx), proteomics (Olink). Prompts are programmatically generated with standardized research scenarios, structured omics matrices, and covariate-balanced clinical metadata, preventing 'literature memorization'. Evaluation protocol: strict zero-shot, single-forward inference — no iterative prompting, RAG, tool use, or post-hoc calibration. 18 frontier closed-source LLMs from OpenAI, Google, Anthropic, xAI, DeepSeek, and Moonshot AI were evaluated with deterministic decoding and fixed seeds.Longevity-LLMs
Five models fine-tuned from Liquid AI's LFM2 and Alibaba's Qwen3/Qwen3.5 architectures:Trained with Insilico's MMAI Gym for Science. L-Qwen3.5-9B used two-stage LoRA (domain-adaptive continued pretraining, then multi-task instruction tuning); the other four used full-parameter supervised fine-tuning with hierarchical uniform sampling.
Longevity Claw
An agent platform calling gene set enrichment, aging clock computation, population profiling, and target assessment tools. Using L-Qwen3.5-9B under structured constraints of 14 aging hallmarks, 6 rounds of independent sampling with deduplication and confidence weighting yielded 328 non-redundant candidate targets. These showed significant enrichment (up to 5.6×) against empirically validated aging targets and DrugBank druggable targets; 2 of 9 in-vivo-validated longevity targets (CXCL12, ADAMTS14) were recovered.KDM1A validation
KDM1A appeared in 5 of 6 sampling rounds (83.3%), ranking as the top-priority candidate. Independent wet-lab validation confirmed KDM1A — a histone demethylase acting on H3K4me1/me2 — has dual roles in aging regulation and tumorigenesis, with knockdown significantly extending mean lifespan in *C. elegans*. Translation from worms to mammals remains open work.
Implications
1. Specialized small models + high-quality benchmarks beat general frontier LLMs + generic benchmarks for vertical scientific domains — validated by data quality, task design, and strict evaluation protocol. 2. Vertical-domain evaluation (aging, and by extension materials, chemistry, climate) matters more than general capability scores; construct validity is a hard quality metric. 3. A full evaluation–model–discovery loop is demonstrated end-to-end, offering a reusable engineering template for AI-driven scientific discovery.
Context: on 2026-09-07, Insilico also published a Nature Biotechnology Phase IIa paper showing rentosertib (an AI-discovered IPF drug candidate) lowered participants' biological age across six proteomic aging clocks. Insilico's HK shares rose 13.12% on 2026-09-18, closing at HK$60.8 (market cap ~HK$35.4B).