Abstract
The characterization of Tumor MicroEnvironment (TME) is challenging due toits complexity and heterogeneity. Relatively consistent TME characteristicsembedded within highly specific tissue features, render them difficult topredict. The capability to accurately classify TME subtypes is of criticalsignificance for clinical tumor diagnosis and precision medicine. Based on theobservation that tumors with different origins share similar microenvironmentpatterns, we propose PathoTME, a genomics-guided Siamese representationlearning framework employing Whole Slide Image (WSI) for pan-cancer TMEsubtypes prediction. Specifically, we utilize Siamese network to leveragegenomic information as a regularization factor to assist WSI embeddingslearning during the training phase. Additionally, we employ Domain AdversarialNeural Network (DANN) to mitigate the impact of tissue type variations. Toeliminate domain bias, a dynamic WSI prompt is designed to further unleash themodel's capabilities. Our model achieves better performance than otherstate-of-the-art methods across 23 cancer types on TCGA dataset. Our code isavailable at https://github.com/Mengflz/PathoTME.