Back

Compositional and interpretable representation of histology using AI foundation models and sparse autoencoders

Zhao, Z.; Maliga, Z.; Ogbonna, E. C.; Talemi, S. R.; Coy, S.; Gagne, A.; Lumamba, K.; Solomon, I. H.; Santagata, S.; Steyn, A. J. C.; Naidoo, T.; Sorger, P. K.

2026-06-06 bioinformatics
10.64898/2026.06.03.725182 bioRxiv
Show abstract

Light microscopy of tissue sections stained with hematoxylin and eosin (H&E) has been the foundation of histopathology for over 150 years and remains essential for diagnosis and research. The development of high-plex spatial profiling approaches able to measure protein and RNA expression at single-cell resolution augments but does not replace H&E imaging, even in research. Computational pathology (CPath) models based on deep learning promise to further increase the value of H&E imaging but interpreting these models in biological terms remains challenging. As a result, they are not widely used in spatial profiling studies. Here we describe a human-in-the-loop computational framework that leverages CPath foundation models (FMs) and sparse autoencoders (SAEs) to decompose FM embeddings and automatically identify diverse, human-interpretable histopathology features in H&E images. When FM-SAE modeling was applied to pulmonary diseases such as tuberculosis and lung cancer, human-machine interaction augmented and accelerated expert interpretation. Moreover, the resulting annotations provide a morphology-aware approach to integrating 2D and 3D mesoscale tissue architectures with molecular spatial profiling.

Matching journals

The top 4 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.