An Explainable AI Framework for Identifying Universal Aging Signatures in Cell Embeddings

Wei Qiu, Chris Arian, Ethan Weinberger, Soo Yun, Alexander Mendenhall, Jessica E. Young, Maria Brbić, Su‐In LeebioRxiv 2025 · DOI

TL;DR

This paper introduces ACE, an explainable AI framework that disentangles aging signals from other biological variations in single-cell RNA sequencing data. ACE can identify universal aging genes across species and predict biological age, offering a powerful tool for aging research and therapeutic development.

What Problem Does This Paper Address?

Aging is a complex process characterized by a decline in physiological function and increased susceptibility to disease. Single-cell RNA sequencing (scRNA-seq) offers a powerful way to study aging at a cellular level, but interpreting these datasets is challenging. Gene expression in scRNA-seq data is influenced by many factors beyond age, such as cell type, tissue, and sex. These "background" variations can obscure the subtle, yet crucial, molecular changes associated with aging, making it difficult to identify true aging signatures and conserved aging mechanisms across different organisms. Existing methods often struggle to disentangle these confounding factors, requiring numerous specialized models or failing to capture the continuous nature of aging.

Methods

The core of this paper is the ACE (Aging Cell Embeddings) framework, a deep generative model based on a variational autoencoder. ACE is designed to explicitly disentangle aging-related variations from other biological factors. It achieves this by using two parallel encoder networks:

  • Age Encoder: Learns latent representations (embeddings) that capture aging-specific signatures.
  • Background Encoder: Learns latent representations that capture non-aging-related variations (e.g., cell type, tissue, sex, species).

To ensure effective disentanglement, ACE employs two key strategies:

  1. Prediction Networks: These networks are trained to predict the chronological age from the "age embeddings" and background covariates from the "background embeddings." This forces the embeddings to retain information about their respective factors.
  2. Hilbert-Schmidt Independence Criterion (HSIC): This statistical penalty is applied to enforce independence between the age and background latent spaces, ensuring that they capture distinct biological signatures.

ACE also includes a decoder that reconstructs the gene expression data, conditioned on both the age and background embeddings, accounting for technical factors like sequencing depth and batch effects. For interpretability, the authors use Expected Gradients (EG), an explainable AI (XAI) method, to identify key genes contributing to predicted biological age.

Datasets

The authors utilized several large-scale, publicly available scRNA-seq datasets for their analysis:

  • Tabula Muris Senis (TMS) Droplet and FACS datasets: Mouse aging atlases covering multiple tissues and cell types across different ages.
  • Aging Fly Cell Atlas (AFCA): A resource profiling aging in Drosophila melanogaster.
  • Tabula Sapiens (TS) human cell atlas: A comprehensive human cell atlas.
  • Mouse embryo dataset: High-resolution temporal profiling of developing mouse embryos.
  • Human brain aging dataset: scRNA-seq profiles from the prefrontal cortex across the human lifespan.

Key Insights and Ideas

  • Disentangling is Key: The fundamental insight is that by explicitly modeling and separating aging signals from other biological noise (cell type, tissue, etc.), we can uncover more robust and universal aging mechanisms.
  • Universal Aging Signatures Exist: ACE's ability to identify conserved aging genes across species (human, mouse, fly) suggests that there are fundamental biological processes of aging that are shared across diverse organisms.
  • Proteostasis is Central to Aging: Across species, the analysis strongly implicates protein homeostasis (proteostasis) – the processes of protein synthesis, maintenance, and degradation – as a critical conserved mechanism in aging.
  • Biological Age is More Than Chronological Age: ACE can predict biological age at both the cell and organism level, demonstrating that transcriptomic state can be a more nuanced indicator of aging than chronological age alone.
  • Local vs. Global Aging: Aging isn't monolithic; ACE can capture both global (species-wide) and local (tissue/cell-type-specific) aging trajectories, providing a more comprehensive view.

Key Results

  • Effective Disentanglement: ACE successfully disentangles aging signatures from background factors, outperforming baseline methods (MrVI, scANVI, scVI, etc.) across multiple quantitative metrics on mouse and fly datasets.
  • Biological Age Prediction: ACE achieves high correlations for biological age prediction:
  • Cell-level: Pearson's r ≈ 0.9 on TMS Droplet dataset, with strong correlations across most tissue-cell-type pairs.
  • Subject-level: Pearson's r ≈ 1.0 for mouse embryo development and ≈ 0.9 for human brain aging, demonstrating generalization to unseen individuals and developmental stages.
  • Cross-Species Aging Genes: ACE identified conserved aging genes (e.g., H3f3b, Lars2, Cdc42, Pdcd4) that show consistent expression changes across human, mouse, and fly.
  • Experimental Validation: RNAi knockdowns of ACE-identified aging genes in C. elegans significantly impacted lifespan, validating the biological relevance of the identified genes, particularly those involved in proteostasis. Notably, Uba52 (ranked 33rd in cross-species analysis) was validated.
  • Pathway Enrichment: ACE-identified aging genes were significantly enriched in pathways related to translation, stress response, immune regulation, and extracellular matrix remodeling.

Limitations and Future Work

  • Reliance on Annotations: ACE's performance relies on accurate annotations for background factors like tissue and cell type, which can be incomplete or inconsistent across datasets. This dependence could affect model performance on less curated data.
  • Transcriptomic Data Only: Currently, ACE operates solely on transcriptomic data. The authors suggest incorporating additional omics layers (DNA methylation, chromatin accessibility, proteomics) could provide a more comprehensive understanding of aging.
  • Multi-modal Integration: Future work could extend ACE toward multi-omics integration, building on advances in multi-modal generative modeling, to enhance its biological scope.
  • Foundation Models: The disentanglement principle positions ACE as a framework for developing phenotype-specific foundation models in single-cell biology, which could be a significant future direction.

Summary updated

Abstract

Abstract Aging is a complex biological process marked by progressive physiological decline and increased disease vulnerability. Single-cell RNA sequencing offers unprecedented resolution for studying aging, yet isolating aging-related signatures remains challenging because gene expression is primarily shaped by other factors such as cell type, tissue, and sex. We present ACE (Aging Cell Embeddings) , an explainable deep generative framework that disentangles aging-related gene expression changes from background biological variation. ACE employs two latent representations: one capturing aging-related signatures and another representing non-aging-related variation in the data. Through explainable AI, ACE identifies key genes and pathways associated with aging amid dominant non-aging-related variations. Applied to large-scale mouse, fly, and human datasets, ACE uncovers aging signatures both within specific tissue-cell-type contexts and across all tissues and cell types, enabling accurate prediction of biological age. Moreover, ACE identifies aging genes conserved across species, highlighting its ability to reveal shared biological mechanisms of aging. Experimental RNAi knockdowns in C. elegans validate ACE’s findings, confirming its ability to prioritize novel aging genes affecting lifespan. ACE reveals key pathways involved in proteostasis, immune regulation, and extracellular matrix remodeling, and identifies Uba52 through the cross-species model as an important aging gene, whose knockdown in C. elegans significantly shortens lifespan. By providing interpretable and generalizable aging embeddings, ACE establishes a foundation for cross-species single-cell aging studies and translational geroscience.
BibTeX
@misc{qiu2025explainable,
  author = {Wei Qiu and Chris Arian and Ethan Weinberger and Soo Yun and Alexander Mendenhall and Jessica E. Young and Maria Brbić and Su‐In Lee},
  title = {{An Explainable AI Framework for Identifying Universal Aging Signatures in Cell Embeddings}},
  year = {2025},
  doi = {10.1101/2025.11.07.687286},
  howpublished = {bioRxiv},
}
Figure