The End of Aging Clocks: Training Foundation Models to Reason in Aging and Longevity
Zhavoronkov, A.; Aladinskyi, V.; Aliper, A.; Miftakhutdinov, Z.; Reymond, M.; Naumov, V.; Zagirova, D.; Pushkov, S.; Sidorenko, D.; Shayakhmetov, R.; Galkin, F.
Show abstract
The aging clock paradigm has yielded dozens of specialist models that can estimate chronological age or mortality from virtually any biodata type. Yet each such model operates within a fixed modality, relies on a predetermined feature set, and produces limited biological interpretation. Here, we report Longevity-LLM v0.1, a Qwen3-14B model fine-tuned through supervised and reinforcement learning regimes on DNA methylation, proteomics, clinical biomarker, and RNA expression data. Longevity-LLM achieves high ranks in the recently announced Longevity Bench, including such tasks as cancer survival and RNA- or proteome-based age prediction. After reinforcement fine-tuning, the model achieved a 4.34-year MAE in epigenetic age prediction, surpassing the Horvath multi-tissue clock. In addition to age prediction, Longevity-LLM can carry out numerous other tasks, including proteomic profile generation, for which it significantly outperforms all frontier LLMs. These results demonstrate that a single modestly sized LLM can match or replace purpose-built aging clocks across data modalities. This work constitutes an interim report from the initial sprint of our Multi-Modal AI Gym for Science (MMAI), an initiative dedicated to building foundation models for drug discovery and aging research.
Matching journals
The top 7 journals account for 50% of the predicted probability mass.