Discovering Candidate Anti-Aging Perturbations Using a Foundation Model for Gene Expression

Aging is a progressive functional decline driven by complex genetic, epigenetic, environmental, and stochastic interactions that traditional linear models struggle to capture. Using human single-cell RNA-seq data from the multi-tissue AgeAnno dataset, we fine-tuned scGPT, a large transcriptomic model, to predict chronological age groups, achieving high classification accuracy. To identify genes influencing age predictions, we systematically perturbed individual genes in silico and quantified their effects, classifying them as pro- or anti-aging candidates. Our results demonstrate that scGPT does capture age-related dependencies in single-cell data and can be utilized to discover novel candidate gene perturbations-potential targets to be validated as anti-aging interventions.