Overview
On September 17, 2026, Insilico Medicine (HKEX: 3696) published a cover article in *Cell* — *An open benchmark and language models for AI in aging biology* (Cell 189(19):5980–5994.e8, DOI 10.1016/j.cell.2026.08.026). The 13-author team includes first and co-corresponding author Alex Zhavoronkov (Insilico founder and co-CEO), Harvard's Vadim N. Gladyshev, and Insilico AI head Fedor Galkin, with collaborators from 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 of 17 tasks across five biological data modalities (clinical, genetics, epigenetics, transcriptomics, proteomics)
- Longevity-LLMs — a family of five lightweight models from 0.6B to 9B parameters
- Longevity Claw — an agentic platform that autonomously calls tools to nominate intervention targets
- Overall ranking: L-Qwen3.5-9B scored 4.4 on LongevityBench, ranking first among 26 systems — ahead of L-LFM2-2.6B (7.6), L-Qwen3-1.7B (7.8), Gemini-3.1-Pro (8.2), and Claude Opus-4.6 (9.2). Even the smallest 0.6B model (ranked 6th) outperformed Claude Opus-4.5, Kimi-K2.5, GPT-5.2, and Grok-4.3.
- DNA methylation pairwise comparison: L-Qwen3.5-9B achieved 0.868 concordance vs. 0.685 for the best frontier model.
- Olink proteomic age prediction: L-Qwen3-0.6B reached an MAE of 5.7 years vs. 10.1 years for the best frontier model (Kimi-K2.5).
- Construct validity: On CellAge senescence phenotype tasks answerable from training knowledge, frontier models scored near ceiling while L-LLMs averaged 71.4% — showing the benchmark distinguishes genuine data-driven reasoning from knowledge recall.
- Insilico Medicine: insilico.com/news/cell170926en-ai-longevity-discovery-toolkit (2026-09-17)
- unite.ai: "Insilico Medicine Releases Open Longevity AI Toolkit in Cell Study" (2026-09-17)
- Cell: DOI 10.1016/j.cell.2026.08.026 — "An open benchmark and language models for AI in aging biology" (2026-09-17)
- Forbes China: Insilico shares up 13% on Cell cover publication (2026-09-18)
Key Results: 9B Beats 18 Frontier Models
The Benchmark Design
LongevityBench draws on NHANES, GEO (DNA methylation), GTEx (bulk RNA-seq), Olink (plasma proteomics), and SynergyAge/OpenGenes (genetics). Prompts are programmatically generated with standardized research scenarios, structured omics matrices, and covariate-balanced clinical metadata. Evaluation uses strict zero-shot, single-forward-pass inference — no iterative prompting, RAG, tool calls, or post-processing — against 18 frontier closed-source models from OpenAI, Google, Anthropic, xAI, DeepSeek, and Moonshot AI.
The Models and Training
The five models — L-Qwen3-0.6B, L-LFM2-1.2B, L-Qwen3-1.7B, L-LFM2-2.6B, and L-Qwen3.5-9B — are fine-tuned from Liquid AI's LFM2 and Alibaba's Qwen3/Qwen3.5 architectures using Insilico's proprietary MMAI Gym for Science training framework. The 9B model uses two-stage LoRA (domain-adaptive continued pretraining, then multi-task instruction tuning); the others use full-parameter supervised fine-tuning with hierarchical uniform sampling.
Ablation experiments on L-Qwen3.5-9B (1,922 structured prompts, 47 biological feature modules) showed 20 feature-ablation schemes significantly reduced prediction confidence, tracing decisions to modules like blood glucose, renal function markers, sex hormone pathway CpG sites, and immune-regulatory gene programs. However, under a "clock replication" setting, L-LLMs degraded sharply — reinforcing the paper's point that "true understanding" and "imitating known models" are different capabilities.
From Evaluation to Discovery
Longevity Claw, running L-Qwen3.5-9B under constraints from 14 aging hallmarks, generated 328 non-redundant candidate target genes via 6 sampling rounds, deduplication, and confidence-weighted integration. The candidates showed significant enrichment (up to 5.6x) against validated aging-target reference sets, and identified 2 of 9 in-vivo-validated intervention targets (CXCL12 and ADAMTS14).
KDM1A, appearing in 5 of 6 runs (83.3% frequency), was prioritized and independently validated: this histone demethylase has dual roles in aging regulation and tumorigenesis, and knockdown significantly extended mean lifespan in *C. elegans*. The authors note the worm-to-human translational path requires further work.
Implications
1. Specialized small models + high-quality benchmarks beat general large models + generic benchmarks in vertical domains — a validated paradigm for AI for Biology. 2. AI evaluation is shifting from general capability assessment to vertical reasoning assessment, offering a template for materials, chemistry, and climate domains. 3. The evaluation-to-discovery loop is closed: benchmark → model → agent → wet-lab-validated target provides a reusable engineering template for AI-driven scientific discovery.
This follows Insilico's September 7, 2026 *Nature Biotechnology* paper showing its AI-designed IPF drug rentosertib lowered subjects' biological age across six proteomic aging clocks in a Phase IIa trial. Together they form a dual narrative of clinical evidence plus open infrastructure. Insilico's Hong Kong shares rose 13.12% to HK$60.80 on September 18 (market cap HK$35.4 billion).