Researchers at AMP 2026 present data on AI models for cancer classification, reporting accuracy, and test triage.
Introduction to AI in Cancer Diagnosis Studies presented at the Association for Molecular Pathology (AMP) 2026 Annual Meeting & Expo in Seattle reveal how artificial intelligence (AI) can assist laboratories in classifying complex cancers, generating more dependable genetic reports, and streamlining decisions for rapid molecular testing. While promising, the presented work is preliminary and requires further validation for widespread clinical adoption. AI Model Classifies Blood Cancer Subtypes A collaborative team from ALMA Genomics, the University of Pennsylvania, and MIT developed ALMA 3, an AI foundation model capable of identifying and classifying blood cancers based on DNA methylation patterns. This model, trained on over 110,000 patient samples, demonstrated a 99.5% sensitivity in detecting hematolymphoid tumors and a 90.9% positive predictive value for specific subtype assignments. It successfully classified 81 types of blood and lymphatic cancers, including those traditionally difficult to diagnose, suggesting a significant role for AI in molecular classification alongside conventional pathology, pending further clinical validation. Pedro Orsini, MD, and Francisco Marchi, PhD, led this research. Open-Source Platform Targets Reporting Accuracy University of Michigan scientists introduced Turnkey Precision Oncology, an open-source platform designed to process DNA and RNA sequencing data, identify genetic alterations, and create draft pathology reports. Its AI component, TPO-Agent, ensures interpretations are backed by evidence from established cancer genetics databases and scientific literature, aiming to minimize unsupported or inaccurate information in clinical reports. In trials with 50 genetic alterations across 10 patients, the system's claims were 95% supported by evidence, outperforming a conventional AI at 65%. The platform also showed high agreement (95%) with clinical laboratories in identifying genetic alterations and expert interpretations of clinically significant changes. Researchers believe such evidence-based AI, coupled with expert oversight, could enhance reporting efficiency and make comprehensive cancer genetic reporting more accessible. Marcin Cieslik, PhD, and Kevin Meng-Lin, BS, presented these findings. AI Triage May Reduce Rapid EGFR Testing Researchers at Memorial Sloan Kettering Cancer Center and Icahn School of Medicine at Mount Sinai explored AI's potential in helping pathologists decide which lung cancer specimens require rapid EGFR mutation testing, thereby conserving valuable tissue and resources. Integrating AI analysis of digital pathology slides (using TAPFM and EAGLE models) into a workflow, molecular pathology fellows reviewed 80 lung adenocarcinoma cases. The AI suggested deferring rapid EGFR testing in 26 cases (33%) without missing confirmed EGFR-positive instances. The method was less effective for gene fusions, highlighting a need for further refinement and human oversight. The team concluded that this AI-driven triage could offer a safe and cost-effective method to reduce rapid molecular testing and preserve tissue for extensive genomic analysis. Nicholas Dcunha, MBBS, and Chad Vanderbilt, MD, were key contributors to this study. Slide Images Show Molecular Signals Two additional studies demonstrated AI's capability to extract molecular information from standard pathology slides. Researchers at Beth Israel Deaconess Medical Center and Harvard Medical School used an AI model to differentiate hepatocellular carcinoma from intrahepatic cholangiocarcinoma (liver cancers). The model, without specific prior training, identified visual patterns linked to cancer-type-specific genes, confirmed by RNA sequencing data from the same tumors. Avash Das, MBBS, PhD, and Vikram Deshpande, MD, led this work. Simultaneously, Stanford University scientists analyzed digitized gastric cancer tissue slides from 367 patients. An AI foundation model identified image features correlated with mutations in four cancer-related genes, achieving predictive performance scores of 0.80 for PIK3CA and 0.77 for CDH1. These findings suggest that AI analysis of existing slides could help prioritize molecular testing, although further validation in diverse patient populations is needed. Kiran Nijjer oversaw the Stanford research. The Association for Molecular Pathology (AMP), founded in 1995, has over 3,100 members dedicated to molecular diagnostics.
Introduction to AI in Cancer Diagnosis
Studies presented at the Association for Molecular Pathology (AMP) 2026 Annual Meeting & Expo in Seattle reveal how artificial intelligence (AI) can assist laboratories in classifying complex cancers, generating more dependable genetic reports, and streamlining decisions for rapid molecular testing. While promising, the presented work is preliminary and requires further validation for widespread clinical adoption.
AI Model Classifies Blood Cancer Subtypes
A collaborative team from ALMA Genomics, the University of Pennsylvania, and MIT developed ALMA 3, an AI foundation model capable of identifying and classifying blood cancers based on DNA methylation patterns. This model, trained on over 110,000 patient samples, demonstrated a 99.5% sensitivity in detecting hematolymphoid tumors and a 90.9% positive predictive value for specific subtype assignments. It successfully classified 81 types of blood and lymphatic cancers, including those traditionally difficult to diagnose, suggesting a significant role for AI in molecular classification alongside conventional pathology, pending further clinical validation. Pedro Orsini, MD, and Francisco Marchi, PhD, led this research.
Open-Source Platform Targets Reporting Accuracy
University of Michigan scientists introduced Turnkey Precision Oncology, an open-source platform designed to process DNA and RNA sequencing data, identify genetic alterations, and create draft pathology reports. Its AI component, TPO-Agent, ensures interpretations are backed by evidence from established cancer genetics databases and scientific literature, aiming to minimize unsupported or inaccurate information in clinical reports. In trials with 50 genetic alterations across 10 patients, the system's claims were 95% supported by evidence, outperforming a conventional AI at 65%. The platform also showed high agreement (95%) with clinical laboratories in identifying genetic alterations and expert interpretations of clinically significant changes. Researchers believe such evidence-based AI, coupled with expert oversight, could enhance reporting efficiency and make comprehensive cancer genetic reporting more accessible. Marcin Cieslik, PhD, and Kevin Meng-Lin, BS, presented these findings.
AI Triage May Reduce Rapid EGFR Testing
Researchers at Memorial Sloan Kettering Cancer Center and Icahn School of Medicine at Mount Sinai explored AI's potential in helping pathologists decide which lung cancer specimens require rapid EGFR mutation testing, thereby conserving valuable tissue and resources. Integrating AI analysis of digital pathology slides (using TAPFM and EAGLE models) into a workflow, molecular pathology fellows reviewed 80 lung adenocarcinoma cases. The AI suggested deferring rapid EGFR testing in 26 cases (33%) without missing confirmed EGFR-positive instances. The method was less effective for gene fusions, highlighting a need for further refinement and human oversight. The team concluded that this AI-driven triage could offer a safe and cost-effective method to reduce rapid molecular testing and preserve tissue for extensive genomic analysis. Nicholas Dcunha, MBBS, and Chad Vanderbilt, MD, were key contributors to this study.
Slide Images Show Molecular Signals
Two additional studies demonstrated AI's capability to extract molecular information from standard pathology slides. Researchers at Beth Israel Deaconess Medical Center and Harvard Medical School used an AI model to differentiate hepatocellular carcinoma from intrahepatic cholangiocarcinoma (liver cancers). The model, without specific prior training, identified visual patterns linked to cancer-type-specific genes, confirmed by RNA sequencing data from the same tumors. Avash Das, MBBS, PhD, and Vikram Deshpande, MD, led this work. Simultaneously, Stanford University scientists analyzed digitized gastric cancer tissue slides from 367 patients. An AI foundation model identified image features correlated with mutations in four cancer-related genes, achieving predictive performance scores of 0.80 for PIK3CA and 0.77 for CDH1. These findings suggest that AI analysis of existing slides could help prioritize molecular testing, although further validation in diverse patient populations is needed. Kiran Nijjer oversaw the Stanford research. The Association for Molecular Pathology (AMP), founded in 1995, has over 3,100 members dedicated to molecular diagnostics.
