Sparse Autoencoders Reveal Interpretable Features in Single-Cell Foundation Models

Flavia Pedrocchi, Florian Barkmann, Amir Joudaki, Valentina Boeva

Track: Full / long paper (5-8 pages)

Keywords: single-cell foundation models, sparse autoencoders, interpretability, steering

TL;DR: We use sparse autoencoders to interpret single-cell foundation models, revealing that they capture diverse biological signals but fragment cell type information, and demonstrate that targeted feature interventions can improve batch integration.

Abstract: Single-cell foundation models (scFMs) hold promise for applications in cell type annotation, data integration, and prediction of the effects of cell perturbations, but their internal mechanisms remain poorly understood. We investigate the structure of these models by training sparse autoencoders (SAEs) on the hidden representations of three widely used scFMs: scGPT, scFoundation, and Geneformer. The learned features reveal diverse and complex biological and technical signals, which emerge even in pre-trained models. We also observe that the encoding of this information differs between scFMs with distinct training protocols and architectures. Finally, we demonstrate that SAE-derived features are functionally related to model behavior and can be intervened upon to reduce unwanted technical effects while steering model outputs to preserve the core biological signal. These findings provide a path toward more interpretable and controllable single-cell foundation models.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Submission Number: 55