Abstract 4021: Precision diagnostics for early melanoma detection using spatial biology and AI guided image analysis

Abstract Accurate diagnosis of melanoma is particularly challenging, as malignant lesions often resemble benign naevi. The gold standard for diagnosing melanocytic lesions relies on the examination of morphological cell features using hematoxylin and eosin (H&E) stained tissue samples. This process however is often subject to variability, lacking the sensitivity to precisely define cell types with a high degree of accuracy. This project aims to overcome histopathological limitations by accurately defining transcriptional cell states, indicative of invasive potential, in benign and malignant lesions. We propose that this will enable more precise definitions of true melanomas in a sea of benign or indolent lesions, which constitute the bulk of excised lesions sent for diagnosis. To identify these “hidden” cell states, we performed spatially resolved whole transcriptome profiling (Visium CytAssist) of whole lesions using a large progression series of samples (n=75): ranging from healthy skin, naevi (benign and low to high grade dysplastic naevi), melanoma in situ, T1a melanomas, and melanoma arising from pre-existing naevi. To enable deconvolution of the Visium ‘spots’, we performed Single Cell RNA sequencing on healthy skin, naevi and thin melanomas, to determine marker genes specific to each cell type present in human skin. This analysis revealed a change of melanocyte cell states along melanoma progression, from conventional naevus to T1a melanomas. We further identified naevus and melanoma-specific signatures which can be spatially resolved, enabling classification of lesions as “benign” or “malignant”, as well as identifying previously hidden regions of malignant potential. We further utilised single cell resolved Xenium (transcriptomic) and Orion (proteomic) platforms to validate the spatial expression of benign and malignant genes and to finely characterise immune cell populations within the tumour microenvironment. Additionally, all specimens have been assessed for a panel of 361 cancer genes, facilitating a comprehensive characterization of the mutational profile driving melanoma progression. By integrating the molecular data with spatial gene expression analysis, we have developed an innovative deep learning model that enhances diagnostic precision for melanoma based on H&E-stained slides. Together, we have defined melanocyte cell states across melanoma progression and uncovered key tumour-microenvironment interactions. We envision that these data will permit AI-guided diagnosis to accurately delineate benign and malignant melanocytic lesions. Citation Format: Yung-Ching Kao, Samuel X. Tan, Xiao Tan, Harald Oey, Katie J. Lee, Tuulia Vallius, Roxanne Pelletier, Andrew Causer, Chenhao Zhou, Darren Smit, Jade O'Neill, Albert Xiong, Blake O'Brien, Angus Collins, Peter Karl Sorger, H. Peter Soyer, Kiarash Khosrotehrani, Quan Nguyen, Mitchell S. Stark. Precision diagnostics for early melanoma detection using spatial biology and AI guided image analysis [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 4021.