Integrating morphology and molecular biology: AI-driven molecular prediction in Pathology with a focus on Cytology

Background: Recent computational pathology research proposes using artificial intelligence to infer molecular features from routine hematoxylin-eosin images, aiming to reduce the cost and time associated with biomarker analysis. Summary: Herein, we explore advancements in predicting genetic mutations, molecular subtypes, microsatellite instability, tumor mutational burden, and spatial proteomics from histological and, specifically, cytological specimens. Despite some promising predictive performances, recent evidence also suggests that current models may rely on confounding clinicopathological features or “shortcut learning” rather than true biological signals, raising important questions about their consistency when applied to broader and more diverse patient cohorts. In cytology, the field is at an earlier stage, and although AI-based molecular triage is an appealing concept, further research is needed to determine whether the predictive performance observed in histology can be replicated in cytological specimens. Key Messages: AI-based prediction of molecular features from histology shows promise but remains insufficient to replace established molecular testing. Current models may be influenced by shortcut learning and therefore require validation across broader cohorts. In cytology, AI-driven molecular prediction is still preliminary and warrants substantial further investigation.