TOX-FM: Foundation models driven over one million blades from six decades of toxicology studies

Six animal species, more than 60 organs and 14,000 studies: a family of numerical pathology models dedicated to preclinical drug safety trials.

The foundation models in digital pathology are almost all trained on human blades from the clinic. Much of histopathology, however, is done upstream in toxicology studies that assess the safety of future drugs in animals. A team of researchers presents arXiv TOX-FM, a family of models designed for this preclinical field.

An unusual corpus

The models were trained on more than 1 million digitized blades, covering more than 60 organs in six animal species and from over 14,000 toxicology studies conducted over six decades. To evaluate them, the authors established a trial bench of 18 preclinical tasks, including an existing bench of 9 clinical tasks, to compare them with the best models of human pathology in both areas.

The study also addresses open questions: should specialized models be developed for each field, and can an animal-based model be translated into humans, or vice versa? The authors learn about the promises and limitations of these models outside the clinic.