AI Models Predict Aging From Blood, Gene, and Protein Data in Cell
On September 17, researchers from Insilico Medicine, Liquid AI, the Buck Institute for Research on Aging, Harvard Medical School, and Brigham and Women’s Hospital published a comprehensive framework for aging biology in Cell. The initiative introduces an integrated ecosystem designed to transition large language models from literature synthesis to direct analysis of complex biological datasets, including blood biomarkers, DNA methylation, transcriptomics, proteomics, and genetic perturbation data. Central to the project is LongevityBench, a benchmark evaluating 17 distinct tasks across 25,457 prompts to test how AI models interpret real-world aging data. Unlike standard knowledge-based assessments, this benchmark requires models to predict chronological age, forecast ten-year mortality risk, and map molecular changes directly from raw multi-omics inputs. The evaluation revealed significant fragmentation among mainstream commercial models, with no single general-purpose LLM demonstrating consistent accuracy across clinical, epigenetic, or proteomic tasks. To address these limitations, the team fine-tuned five specialized models ranging from 0.6 billion to 9 billion parameters. Leveraging open architectures from Liquid AI and Alibaba, these Longevity-LLMs prioritized high-quality, domain-specific training data. In comprehensive benchmarking, the 9-billion-parameter L-Qwen3.5-9B achieved the highest overall ranking, surpassing frontier general models including Gemini and Claude. More granular testing demonstrated substantial performance gaps in raw data interpretation: the 0.6-billion-parameter variant reduced mean absolute error in proteomic age prediction to 5.7 years, compared to 10.1 years for the best commercial competitor. Ablation studies confirmed that specialized models rely on actual biological features rather than statistical templates, though researchers caution that predictive correlation does not equate to causal biological mechanism. Beyond model training, the framework extends into automated research workflows through Longevity Claw, an AI agent capable of orchestrating multi-step aging analyses. The agent integrates over 230 proprietary and open-source biological clocks, executes gene-set enrichment, queries external databases like PubMed and DrugAge, and dynamically adjusts its analytical pathway based on intermediate outputs. In a recent pilot targeting age-related interventions, the system autonomously screened 14 aging characteristics and proposed 328 potential intervention genes. Cross-referencing with established experimental datasets showed significant enrichment of known aging pathways, validating the agent’s capacity to surface biologically relevant signals for further preclinical investigation. The entire suite, including the benchmark, models, and agent interface, has been released to the scientific community to accelerate computational gerontology. Researchers emphasize that while AI can rapidly filter high-dimensional biological noise and prioritize candidate targets, computational predictions require rigorous validation through cellular and animal models. The publication marks a structural shift in longevity research, establishing standardized evaluation metrics and open tools that bridge machine learning capabilities with multi-omics aging biology.
