Organ-Age: A Multimodal Fusion of Transcriptomic and Radiological Signals for Organ-Resolved Biological Age Estimation

Arnav Sangle

Abstract: This study proposes Organ-Age, a multimodal framework for estimating organ-resolved biological age by integrating transcriptomic and radiological data. Using bulk RNA-seq (GTEx), chest radiographs (CheXpert), and brain MRI (IXI), modality-specific encoders generate embeddings aligned through contrastive learning and fused via a transformer-based architecture. The model outputs probabilistic organ-level age estimates, evaluated on a combined cohort exceeding 190,000 samples. The aligned multimodal approach achieved a mean absolute error of approximately 9.3 years, outperforming unimodal and unaligned fusion baselines. Residual di?erences between predicted and chronological age revealed heterogeneous aging patterns across organs. While results suggest potential for localized aging analysis, interpretability and causal inference remain limited by dataset heterogeneity and absence of matched multimodal samples. These findings support further exploration of multimodal representation learning for organ-specific aging assessment.

Keywords: biological age, organ aging, multimodal learning, contrastive learning, contrastive alignment, transcriptomics, fusion transformer, interpretability, biological aging biomarkers, probabilistic regression, multimodal representation learning