Toward a Cross-Species Mechanistic Aging Clock with Fine-Tuned scGPT
Discipline: Computer Sciences and Information Management
Session: 2
Room: 6 - Inman
Khoi Le - Brown University
Co-Author(s): Hong Qin, Old Dominion University, Norfolk, VA
Aging is a process conserved in nearly all biological systems characterized by a progressive decline in normal function and increased risk of mortality. Despite aging’s universality, its molecular basis remains elusive, and understanding the mechanisms of aging at the cellular level may enable new strategies to combat age-related diseases. One powerful tool to study aging in recent years is single-cell RNA sequencing (scRNA-seq), which quantitatively profiles gene expression at cellular resolution. Aging clocks, or computational or statistical models that predict age, have been trained using scRNA-seq data from organism-wide aging cell atlases. However, existing aging clocks, primarily trained using regularized linear regression, can not capture non-linear molecular interactions and the heterogeneity of aging. This limitation may account for the current lack of universal aging clocks that generalize across cell types and species. We propose that the transformer architecture employed by Large Language Models can be used to construct a universal aging clock while maintaining interpretability. As a proof of concept, we fine-tuned scGPT, a foundation model pre-trained on human scRNA-seq data, to predict cellular age using data from the Aging Fly Cell Atlas, Cell Atlas of Worm Aging, and Tabula Muris Senis (Mouse Aging Cell Atlas). Gene expression values in individual cells were tokenized by mapping genes to unique IDs and assigning them to one of 51 bins based on their relative ranking within the cell, with higher bins corresponding to genes with higher expression levels. This preprocessing step addressed batch effects and processing differences across datasets. The tokenized data were then processed by a model composed of 12 stacked encoder layers, yielding cell embeddings. Finally, embeddings passed through a 3-layer classification head to produce predictions based on the datasets’ age labels. We achieved ~80% accuracy across all models with the youngest age label consistently having the highest accuracy. We then calculated SHapley Additive exPlanations (SHAP) values to identify the genes most important for age prediction. Gene Ontology enrichment analysis on high-scoring genes revealed pathways associated with aging, such as altered ribosomal protein transcript expression. Additionally, over 80% of the genes ranked in the top 10% by SHAP values were also differentially expressed (FDR < 0.05) across models trained exclusively on fly, worm, or mouse data, suggesting their relevance as aging biomarkers. Taken together, the results show that transformer-based aging clocks can reliably predict age across all cell types while maintaining interpretability, enabling a “one model fits all” approach. Future research involves examining the models’ gene embeddings and attention scores to explore learned gene interactions and regulatory networks. Moreover, tailoring the pre-training of scGPT rather than using the existing pretrained model may enhance accuracy.
Funder Acknowledgement(s): I acknowledge the support of NSF REU #2149956
Faculty Advisor: Hong Qin, hqin@odu.edu
Role: I executed the entirety of the project, including data pre-processing, model construction, model training, model evaluation, and model interpretation. Dr. Hong Qin originally conceptualized the broad idea of fine-tuning scGPT to predict cellular age across species, but I came up with all of the details for implementation. I conceptualized the idea of using model interpretation to find aging biomarkers, the GO term enrichment analysis of top SHAP genes, and the differentially expressed genes overlap analysis.