A Biologically Interpretable Single-Cell Pathology Foundation Model for Molecular Decoding of H&E Images
En palabras de los autores
The microscopic structure of human tissues encodes molecular information that is not directly accessible by visual inspection. Current pathology foundation models improve histological interpretation but operate at multicellular resolution. It remains unclear whether routine histology contains structured molecular and cellular information recoverable at single-cell resolution, which biological properties are associated with this recoverability, and where its limits lie. Here, we introduce Loki2, a single-cell pathology foundation model trained on UniSeg, a pan-tissue resource integrating 15 million cells with segmentation, lineage annotations and measured or histology-inferred transcriptomic profiles. We examine the above question across cellular representation, molecular recoverability, and patient-level relevance. Loki2 captures cellular identity and continuous morphological variation, with interpretable latent dimensions linked to physical nuclear phenotypes through DAPI validation. Transcriptomic retrieval and in silico protein staining test molecular recoverability, with gene-level analyses revealing the strongest recovery for genes associated with morphologically distinct cellular identities and functions. Then, patient staging and biomarker discovery connect cellular morphology representations to clinical relevance. Together, these findings establish Loki2 as a biologically interpretable single-cell pathology foundation model and provide the opportunities of recovering cellular and molecular information from routine H&E histology at single-cell resolution.
Apareció: miércoles, 18 de noviembre. bioRxiv. Preprint, todavía sin revisión por pares.