Non-clinical pathology foundation models pretrained on over 1M H&E slides from six animal species across 14,000 toxicology studies.
No providers recorded yet. Browse all providers
Before a drug candidate reaches a first-in-human trial, toxicologic pathologists examine stained sections from dozens of organs in rats, dogs, monkeys and other species, looking for the microscopic changes that signal drug-induced injury. Most of those tissues show nothing, but every one has to be read. Pathology foundation models could supply the image representations that tools for this work are built on, yet nearly all of them, such as Prov-GigaPath, H-optimus-0 and Midnight, were trained on human clinical slides, mostly from oncology. Rat bone marrow or monkey adrenal gland is a different distribution, and the one large public non-clinical dataset, TG-GATEs, covers only rat liver and kidney.
TOX-FM is a family of vision transformers from Novartis Biomedical Research and Novartis that fills that gap with an in-house archive: over one million H&E whole-slide images spanning more than 60 organs and six species (rat, mouse, dog, monkey, pig and rabbit), drawn from over 14,000 toxicology studies run between 1967 and 2024. Many of these slides were digitized from archived glass specifically for the project. The models are trained with DINOv3 self-supervised distillation, and the paper, posted to arXiv in October 2026, uses them to ask how much domain-specific pretraining matters and how well pathology representations transfer between animal and human tissue in each direction.
The family has four main checkpoints: TOX-FM-100K-ViT-L (305M parameters) and TOX-FM-100K-ViT-H (636M), trained on 100,000 rat slides from 700 studies (about 1 billion patches); TOX-FM-1M-ViT-H, trained from scratch on the curated six-species set from 13,352 studies; and TOX-FM-1M-ViT-H-CL, the continual-learning model and the strongest overall. Training used 16 or 32 NVIDIA H100 GPUs for up to 2 million AdamW iterations with multi-crop and RandStainNA stain augmentation. Checkpoints were selected on a separate 12,451-slide validation set.
Evaluation used frozen encoders with linear probes, attention-based multiple-instance-learning probes and a U-Net decoder. An independent set of 19,935 slides from 289 studies supplied 18 non-clinical tasks: species, per-species organ identification, slide artifacts and pathologist-annotated lesions, as patch classification, slide classification and segmentation. Nine clinical tasks came from the eva benchmark (BACH, BreakHis, CRC, Gleason grading, PatchCamelyon, Camelyon16, PANDA, BCSS and MoNuSAC). Each domain's own models won on that domain's tasks, with TOX-FM-1M-ViT-H-CL best among non-clinical models and Midnight-12k best among clinical ones. Cross-domain performance stayed within about 95% of the strongest in-domain model. Clinical models lost the most on low-magnification species and organ tasks, where tissue architecture differs most between humans and animals. Scaling from ViT-L to ViT-H on the rat-only data improved non-clinical scores but lowered clinical ones, while adding species diversity improved both.
TOX-FM is built as a backbone for computational toxicologic pathology: screening study tissues to flag those with drug-related findings, detecting and segmenting lesions, identifying organs and species in mixed slide sets, and catching slide-preparation or scanning artifacts before review. Because the preclinical phase is the main safety gate before human dosing, these are the tasks where pathologist time is most concentrated. The authors also point to translational research, where a representation space shared between animal and human tissue could help relate preclinical findings to clinical outcomes.
TOX-FM is the first non-clinical pathology foundation model to extend beyond rat liver and kidney, and its paired benchmark is the first systematic test of how pathology models transfer between animal and human tissue in both directions. Its results bear on decisions other drug developers face: domain-specific pretraining still wins in-domain, but diverse animal data alone yields representations competitive on human cancer tasks, and dataset diversity mattered more than model size. The work is a preprint that has not been peer reviewed. The training data are proprietary Novartis study archives, and the paper includes no code, weights or data availability statement, so the models cannot be used or reproduced outside the company.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.