Pharma
Deployment record
In the tracker →- Organisation
- AstraZeneca
- System
- TRACE
- Stage
- Pilot
- Function
- preclinical toxicologic pathology
- 01AstraZeneca is collaborating with Tremont AI to develop AI solutions for preclinical toxicologic pathology to enhance drug safety assessments.
- 02Tremont AI emerged from the Mahmood Lab and focuses on connecting tissue findings with other evidence in preclinical safety studies.
- 03The TRACE model developed by Tremont AI uses 15 million image patches from rat tissue to train AI for identifying drug-related changes in tissue.
Show all 5Show less
- 04The collaboration aims to improve the speed and consistency of interpreting subtle treatment-related changes in tissue samples.
- 05AstraZeneca's previous acquisition of Modella AI aligns with its strategy to utilize domain-specific multimodal foundation models in drug development.
AstraZeneca
Follow Deployed on Google
A small patch of damaged tissue can carry an important warning about a drug candidate. Finding it can require a pathologist to examine large areas of otherwise unremarkable tissue, then determine whether the change is related to treatment.
Research underlying the TRACE toxicology model illustrates the difficulty: some focal abnormalities can occupy less than 1% of a tissue section. Reviewing such findings across studies requires both careful image examination and consistent interpretation.
That is the problem Boston-based Tremont AI is targeting through a research collaboration with AstraZeneca. The companies will develop and evaluate AI solutions for preclinical toxicologic pathology, part of the safety assessment conducted before medicines reach human trials.
“Foundation models can change how preclinical safety data is read, making it faster to triage and more consistent to interpret,” said Luca Weishaupt, CEO and Co-founder of Tremont AI.
From a Pathology Lab to Preclinical Drug Safety
Tremont emerged from stealth in September this year, four weeks before announcing the AstraZeneca collaboration. The Boston company spun out of the Mahmood Lab and builds on technology exclusively licensed from Mass General Brigham.
Weishaupt completed his doctorate in the Harvard-MIT Health Sciences and Technology program, working in the Mahmood Lab on foundation models for computational pathology. Tremont’s scientific co-founders are Faisal Mahmood, Long Phi Le, Guillaume Jaume and Jakob N. Kather, bringing research backgrounds in computational pathology and medical AI.
Mahmood directs the Mass General Brigham AI Institute, while Le leads computational pathology at Massachusetts General Hospital. Jaume leads a pathology AI research lab at the University of Lausanne, and Kather directs the Institute for Medical AI at Heidelberg University Hospital.
Tremont is developing models and AI agents for pharmaceutical companies and contract research organizations conducting preclinical safety studies. Its aim is to connect tissue findings with other evidence across a study, helping scientists characterize potential toxicities more systematically.
Finding the Signal in Tissue Images
Toxicologic pathologists examine tissue from preclinical studies to identify changes associated with a candidate compound. Distinguishing those changes from background variation helps characterize the drug’s safety profile and its relationship to dose.
The week's drug discovery, trials and health-system deployments. Wednesdays.
Tremont said the collaboration will evaluate models against the challenges of rare findings that can resemble background changes. The goal is to reduce repetitive review and give specialists more time for interpretation.
The company’s TRACE research, provides background on the approach. Researchers trained the model on 15 million image patches from digitized rat tissue sections and evaluated applications including lesion severity scoring and dose-response characterization.
A foundation model learns reusable visual representations that can support several downstream tasks. In this case, the research sought to avoid building a separate image model from scratch for every toxicology task.
Connecting Images With the Rest of the Study
The company describes TRACE as its image encoder, which converts tissue imagery into representations that other models can use. Its ToxScribe vision-language model generates descriptions of tissue findings, including their location and severity.
Tremont Studio brings slides together with clinical chemistry and gross pathology findings, organized by organ, dose and animal. Its Copilot is designed to answer questions about the study and link responses back to the evidence it used.
Those features address a wider interpretation problem: understanding how an image finding relates to the other observations recorded in the study. Tremont also describes slide-level discussion threads and tracked review decisions, allowing pathologists to examine findings together.
Why the AstraZeneca Partnership Fits
The collaboration gives Tremont a setting in which to develop and evaluate its models against drug-development requirements. For AstraZeneca, it brings a specialist approach to distinguishing subtle treatment-related changes from background findings, the challenge identified in Tremont’s announcement.
There is also an existing link through the founders. Mahmood and Le co-founded Modella AI, which AstraZeneca acquired in 2026. AstraZeneca identifies that acquisition as part of its strategy for domain-specific multimodal foundation models.
The company also uses computational pathology through Quantitative Continuous Scoring (QCS), which analyzes the presence and location of biomarkers in cells and tissues to help identify patients likely to benefit from treatment. The Tremont collaboration addresses a different question earlier in development: what tissue changes may reveal about a candidate drug’s safety.
Last updated
