AI for toxicologic pathology.
What toxicologic pathology is
Toxicologic pathology helps decide whether a new medicine is safe enough to test in people, and at what dose. Before a drug candidate reaches patients, preclinical safety studies test it in a control group and several dose groups. Toxicologic pathologists, specialists in drug safety assessment, then examine the tissue from those studies under the microscope.
A toxicology study is hundreds of slides, plus the clinical chemistry and gross findings that explain them. The pathologist looks for treatment-related changes, which must be told apart from normal biological variation, and connects each finding to the controls, every dose group, the other organs and the clinical data. These judgments are critical to the compound’s safety profile and to the dose at which it can be given safely.
Drug discovery is generating candidate compounds faster than ever, in part due to AI. Each one still has to answer two questions: is it safe for humans, and at what dose does it become toxic? Today the work depends on expert interpretation that is semi-quantitative and subject to variability, and as drug development accelerates it is getting harder to scale.
Where AI helps in a toxicology study
AI for toxicologic pathology means models built for preclinical tissue and for the way a toxicology study is designed: by organ, dose group, sex and animal. General-purpose models tend to break down where findings are rare and difficult to discern from background findings. Toxicology-specific foundation models are built for those conditions. Tremont AI builds them, and the software where pathologists use them.
Seeing the tissue: TRACE
A whole-slide image is a gigapixel of tissue. TRACE, Tremont AI’s vision encoder for toxicologic pathology, turns it into a representation the rest of the stack can work with, one that keeps what a pathologist would look at and drops what they wouldn’t. It is a representation of tissue learned at very large scale on preclinical data, and every other Tremont AI model runs on it. Downstream tasks built on TRACE include quality control, tissue segmentation, nuclear segmentation, captioning and abnormality ranking.
Describing the finding: ToxScribe
ToxScribe is a vision-language model trained across tox species and magnifications. Shown a region or a whole slide, it describes the finding, where it is and how severe, in the words a pathologist would use. On four liver fields from a carbon tetrachloride study, one per dose group, ToxScribe calls the control normal, picks up the first vacuoles at 100 mg/kg and grades the change as it becomes marked at 300.
Connecting the whole study: agents and Copilot
A finding on a slide only means something next to the rest of the study. Agents correlate what the models find with clinical chemistry, organ weights and gross observations. In Tremont Studio, Tremont Copilot answers questions about the study in front of you, such as what the incidence was at the top dose or which animals drove it. It reports incidence and severity by dose group, plotted against the clinical chemistry, in pathology’s own language, and every answer links back to the slides it read.
Reviewing together: peer review on the slide
Peer review used to mean exported images and email chains. In Tremont Studio, discussion lives on the slide. A pathologist pins a finding and starts a thread anchored to the exact lesion, and the peer review pathologist answers on the slide, on the record. Findings are graded, resolved and signed off, and every step is tracked.
The pathologist decides
Foundation models can turn the reading of a study into a measurable and reproducible process, and free pathologists from repetitive review to focus on interpretation. They do not take the decision away. Tremont Studio turns the models’ output into a reviewable record, and the pathologist decides and signs off.
Tremont AI’s products are intended for research and preclinical use, not for clinical diagnosis.
Where the technology comes from
Tremont AI is a Boston company and a spinout of the Mahmood Lab at Mass General Brigham. Its foundation models build on technology exclusively licensed from Mass General Brigham. The team includes co-founder and CEO Luca Weishaupt, PhD, who worked on foundation models for computational pathology in the Mahmood Lab, and toxicologic pathologist Milind Dalvi, DVM. The scientific co-founders are Faisal Mahmood, PhD and Long Phi Le, MD, PhD (Mass General Brigham), Guillaume Jaume, PhD (University of Lausanne) and Jakob N. Kather, MD (Heidelberg University).
Tremont AI emerged from stealth on September 29, 2026. In October 2026 it announced a research collaboration with AstraZeneca to develop and evaluate AI solutions for accelerated safety screening in preclinical toxicologic pathology studies.
Who it is for
Tremont AI works with pharmaceutical and CRO (contract research organization) pathology teams. It works with a small number of partners at a time, with early access evaluated on the partner’s own studies. To talk about your program, write to contact@tremont.ai.
Frequently asked questions
What is toxicologic pathology?
Toxicologic pathology is the microscopic examination of tissue from preclinical drug safety studies. Toxicologic pathologists look for treatment-related changes, tell them apart from normal biological variation, and relate each finding to the controls, the dose groups and the rest of the study. Their judgments help decide whether a new medicine is safe enough to test in people, and at what dose.
What does Tremont AI do?
Tremont AI builds AI for toxicologic pathology: foundation models (TRACE, ToxScribe) and the Tremont Studio platform for preclinical drug safety studies. Its models read the tissue from a study and weigh what they see against everything else the study recorded.
What is TRACE?
TRACE is a vision encoder for toxicologic pathology, a representation of tissue learned at very large scale on preclinical data. It turns a gigapixel whole-slide image into a representation other models can work with, and every other Tremont AI model runs on it.
What is ToxScribe?
ToxScribe is a vision-language model trained across tox species and magnifications. Shown a region or a whole slide, it describes the finding, where it is and how severe, in the words a pathologist would use.
What is Tremont Studio?
Tremont Studio is one workspace for the whole toxicology study: whole-slide images, clinical chemistry and gross findings in one place with AI predictions beside them, live review threads between pathologists, and Tremont Copilot, which answers questions about the study with incidence by dose group and the slides behind each answer.
Who is Tremont AI for?
Pharmaceutical and CRO (contract research organization) pathology teams that review preclinical safety studies.
Does AI replace the toxicologic pathologist?
No. The models describe, rank and connect findings, and every answer links back to the slides it came from. The pathologist reviews the record, decides and signs off.
Where does Tremont AI’s technology come from?
Tremont AI is a spinout of the Mahmood Lab at Mass General Brigham, and its foundation models build on technology exclusively licensed from Mass General Brigham.
Is Tremont AI for clinical diagnosis?
No. Tremont AI’s products are intended for research and preclinical use, not for clinical diagnosis.