Inflammation-linked aging signals in frozen single-cell foundation models: donor-aware detection and robustness...

Area of Science:

• Computational Biology
• Genomics
• Artificial Intelligence

Background:

• Single-cell foundation models (e.g., scGPT, Geneformer) are trained on human single-cell RNA-seq data without chronological age information.
• It remains unclear if these models internally represent aging biology and how to distinguish true signals from sampling biases.

Purpose of the Study:

• To develop and apply an evaluation framework to test if foundation models encode interpretable aging biology.
• To determine if apparent aging signals in models are genuine biological phenomena or artifacts of donor and cell type sampling.

Main Methods:

• Applied a nine-step evaluation pipeline to two foundation models (scGPT, Geneformer) and five PBMC datasets (~5 million cells, ~2,000 donors).
• Utilized sparse-feature decomposition, pathway-level analysis, targeted perturbations, and cell-type composition matching for validation.
• Assessed model performance against a 50-component Principal Component Analysis (PCA) baseline.

Main Results:

• Foundation models encode age but do not predict it better than a 50-component PCA.
• Sparse autoencoders identified 132 robust aging-related features, concentrated in inflammation (TNF/NF-κB, IFN-γ).
• Geneformer's NF-κB program showed a directional aging effect that persisted, albeit attenuated, after cell-type composition matching.

Conclusions:

• The developed evaluation framework reliably distinguishes biological aging signals from sampling artifacts in single-cell foundation models.
• Frozen foundation models encode a recoverable, zero-shot aging signal, primarily in NF-κB and IFN-γ inflammation submodules.
• The framework's recommendation is to report both unrestricted and composition-matched contrasts for robust signal validation.