Transcriptomic age prediction using mixture-of-experts models reveals tissue-specific aging signatures in large-scale human RNA-sequencing data

Abstract

Transcriptomic age prediction has emerged as a powerful approach for understanding biological aging processes, yet systematic comparisons of large-scale RNA-sequencing datasets remain limited. We developed and validated a mixture-of-experts machine learning model using the ARCHS4 dataset comprising 56,877 human RNA-sequencing samples spanning ages 2-114 years across diverse tissues. Our model achieved superior performance (MAE = 7.58 years) compared to traditional approaches, with tissue-specific variations revealing lung and brain as optimal predictors while liver showed reduced accuracy. Feature importance analysis identified FOSB as the dominant age predictor (importance = 1.00), followed by complement component C4B_2 (0.95), long non-coding RNA PAX8-AS1 (0.85), mitochondrial gene MT-RNR2 (0.64), and glial marker GFAP (0.53), collectively representing stress response, immunosenescence, epigenetic regulation, mitochondrial dysfunction, and neuroinflammation pathways. Residual analysis revealed heteroscedasticity with prediction variance increasing from +-5 years in young adults to +-40 years in centenarians, indicating systematic model limitations at extreme ages. Comparison with David Sinclair's epigenetic clock approaches demonstrates that transcriptomic models achieve comparable accuracy while providing unique insights into tissue-specific aging mechanisms unavailable in blood-based methylation clocks. These findings establish transcriptomic age prediction as a complementary tool to epigenetic clocks, enabling precision aging medicine through identification of accelerated aging signatures and therapeutic targets across the human lifespan.