Abstract
Genotypic and phenotypic diversity, which generates heterogeneity during disease evolution, is common in cancer. The identification of features specific to each patient and tumor is central to the development of precision medicine and preclinical studies for cancer treatment. However, the complexity of the disease due to inter- and intratumor heterogeneity increases the difficulty of effective analysis. Here, we introduce a sequential deep learning model, preprocessing to organize the complexity due to heterogeneity, which contrasts with general approaches that apply a single model directly. We characterized morphological heterogeneity using microscopy images of patient-derived organoids (PDOs) and identified gene subsets relevant to distinguishing differences among original tumors. PDOs, which reflect the features of their origins, can be reproduced in large quantities and varieties, contributing to increasing the variation by enhancing their common characteristics, in contrast to those from different origins. This resulted in increased efficiency in the extraction of organoid morphological features sharing the same origin. Linking these tumor-specific morphological features to PDO gene expression data enables the extraction of genes strongly correlated with intertumor differences. The relevance of the selected genes was assessed, and the results suggest potential applications in preclinical studies and personalized clinical care.