Sparse Autoencoders Reveal Interpretable Features in Single-Cell Foundation Models

Single-cell foundation models (scFMs) hold promise for applications in cell type annotation and data integration, but their internal mechanisms remain poorly understood. We investigate the structure of these models by training sparse autoencoders (SAEs) on the hidden representations of two widely used scFMs, scGPT and scFoundation. 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. Further, we find that while many features capture the information about cell types across several studies, they often fall short of unifying it into a single generalized representation. Finally, we demonstrate that SAE-derived features are causally 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.