Studies presented at the Association for Molecular Pathology annual meeting evaluate how artificial intelligence can support diagnosis, reporting, and genetic test selection.
Researchers are presenting data at the Association for Molecular Pathology (AMP) 2026 Annual Meeting & Expo in Seattle on how artificial intelligence (AI) can help laboratories classify difficult cancers, produce more reliable genetic reports, and decide which specimens need rapid molecular testing. Most of the work is preliminary, and several authors note that further validation is needed.
AI Model Classifies Blood Cancer Subtypes
A team from ALMA Genomics, the University of Pennsylvania, the Massachusetts Institute of Technology, and other institutions developed ALMA 3, an AI foundation model that uses DNA methylation patterns to identify and classify blood cancers. The model was trained on more than 110,000 patient samples.
In independent testing, the model detected hematolymphoid tumors with 99.5% sensitivity and achieved a 90.9% positive predictive value when assigning specific subtypes. The study covered 81 subtypes of blood and lymphatic cancers, including several that are hard to classify with conventional methods.
The findings point to a potential role for AI-assisted molecular classification alongside conventional pathology, though additional clinical validation is under way. Pedro Orsini, MD, of ALMA Genomics led the work. Corresponding author Francisco Marchi, PhD, will give an oral presentation Friday, Nov 13, and present a poster Saturday, Nov 14 (Abstract H071).
Open-Source Platform Targets Reporting Accuracy
University of Michigan scientists developed Turnkey Precision Oncology, an open-source platform that processes DNA and RNA sequencing results, identifies genetic alterations, and generates draft reports for pathologist review. Its AI component, TPO-Agent, draws on established cancer genetics databases and scientific literature and checks that its interpretations are backed by cited evidence. The goal is to reduce the risk of unsupported or inaccurate information in clinical reports.
In testing of 50 genetic alterations across 10 patients, about 95% of the system’s claims were correctly supported by cited evidence, compared with 65% for a conventional AI approach equipped with similar research tools. Across 28 reports, the broader platform showed approximately 95% agreement with a clinical laboratory in identifying individual genetic alterations. Its AI-generated reports also showed high agreement with expert interpretations of clinically significant alterations.
The authors suggest that evidence-based AI systems with expert oversight could help laboratories produce reports more efficiently and make full cancer genetic reporting more accessible to community laboratories. How the platform performs in routine clinical practice has yet to be established.
Marcin Cieslik, PhD, of the University of Michigan led both studies. He will present Abstract I040 on the platform, and Kevin Meng-Lin, BS, a fellow in his laboratory, will present Abstract I068, which measures the accuracy, hallucination, bias, and guideline adherence of AI-assisted reporting. Both are poster presentations on Friday, Nov 13.
AI Triage May Reduce Rapid EGFR Testing
Researchers at Memorial Sloan Kettering Cancer Center and the Icahn School of Medicine at Mount Sinai tested whether AI could help pathologists decide which lung cancer specimens need rapid testing for EGFR mutations. Multiple tests can use up valuable tumor tissue and laboratory resources.
The team added AI analysis of digital pathology slides, using the TAPFM and EAGLE models, to a workflow in which six molecular pathology fellows reviewed 80 consecutive lung adenocarcinoma cases. Based on the AI findings, reviewers indicated that rapid EGFR testing could have been deferred in 26 cases, or 33%, without overlooking any of the confirmed EGFR-positive cases.
The approach was less reliable for identifying cases that could skip rapid testing for certain gene fusions. The authors said this points to the need for further refinement and human-in-the-loop oversight. They concluded that the approach offers a safe, cost-effective way to cut rapid molecular testing and preserve tissue for broader genomic analysis.
Nicholas Dcunha, MBBS, of Memorial Sloan Kettering led the work, and senior author Chad Vanderbilt, MD, oversaw it. Dcunha will present Abstract I070 at the Friday poster session.
Slide Images Show Molecular Signals
Two other studies examined whether AI can extract molecular information from routine pathology slides.
At Beth Israel Deaconess Medical Center and Harvard Medical School, researchers evaluated whether an AI model could distinguish hepatocellular carcinoma from intrahepatic cholangiocarcinoma, two liver cancers that arise from different cell types and are treated differently. Using a model that required no additional disease-specific training, the team examined 50 digitized tissue slides from The Cancer Genome Atlas public dataset. The model detected visual patterns linked to cancer-type-specific genes, especially those characteristic of hepatocytes, and RNA sequencing data from the same tumors supported the findings. Avash Das, MBBS, PhD, led the work with Vikram Deshpande, MD, and will present Abstract I056 on Friday.
At Stanford University, scientists analyzed digitized tissue slides from 367 patients with gastric cancer, using an AI foundation model to identify image features associated with mutations in four cancer-related genes. The strongest results were for PIK3CA and CDH1, with predictive performance scores of 0.80 and 0.77, on a scale where 0.50 represents chance-level discrimination and 1.0 represents perfect discrimination. The findings require validation in additional patient populations, but they suggest AI analysis of existing slides could eventually help laboratories prioritize molecular testing. Kiran Nijjer of Stanford oversaw the work and will present Abstract I069 on Saturday, Nov 14.
AMP was founded in 1995 and has more than 3,100 members across molecular diagnostics disciplines.
ID 475898336 © BiancoBlue | Dreamstime.com
