From Clinical to Preclinical: Transfer and Interpretability of a Human Pathology Foundation Model in Rabbit Histology

Research involving non-human species is fundamental to preclinical drug discovery and toxicological assessment. In this study, we evaluate the transferability of a pathology foundation model—previously demonstrated mainly in human diagnostic tasks—to rabbit histopathology at $5\times$ magnification, establishing, to our knowledge, the first performance baseline in this domain. We evaluate frozen UNI2-h embeddings on a new dataset of $\sim 1,000$ rabbit whole-slide images acquired at $5\times$ magnification from 17 animals inoculated with VX2 tumor cells and treated with high-intensity focused ultrasound. Five of these slides, each from a different animal, are densely annotated across muscle, necrosis, immune infiltrate, and mixed categories. Linear probes trained on those five slides under leave-one-animal-out cross-validation perform multi-label classification of three tissue categories at macro-F1 0.84 and macro-AUROC 0.97, and predict QuPath-derived cell counts at $R^2 = 0.89$ (per-animal range 0.85 to 0.95), despite simultaneous shifts in species and resolution. These results characterize how far human-trained representations extend to preclinical rabbit histology, and whether low-magnification archives can be analyzed without rescanning. The black-box nature of these embeddings limits their interpretability. To address this, we train a sparse autoencoder (SAE) on UNI2-h embeddings from five million rabbit H\&E tiles, decomposing them into sparse, interpretable features corresponding to distinct biological characteristics. Extending Le et al. (2025) to a non-human species at low magnification, we identify latents whose activation correlates with QuPath-derived cell counts even at $5\times$ magnification. In a separate exploratory analysis of tumor and immune-cell-rich regions, we find that some features activate selectively in immune-cell-rich regions, whereas others activate selectively in tumor regions, enabling these tissue categories to be distinguished through individual SAE features. These findings provide evidence that a pathology foundation model pretrained on large-scale human clinical histology contains transferable biological representations that capture meaningful structures in rabbit tissue, and offer an interpretable framework for preclinical digital pathology.