Building trans-omics evidence: using imaging and 'omics' to characterize cancer profiles

Utilization of single modality data to build predictive models in cancer results in a rather narrow view of most patient profiles. Some clinical facet s relate strongly to histology image features, e.g. tumor stages, whereas others are associated with genomic and proteomic variations (e.g. cancer subtypes and disease aggression biomarkers). We hypothesize that there are coherent "trans-omics" features that characterize varied clinical cohorts across multiple sources of data leading to more descriptive and robust disease characterization. In this work, for l 05 breast cancer patients from the TCGA (The Cancer Genome Atlas), we consider four clinical attributes (AJCC Stage, Tumor Stage, ER-Status and PAM50 mRNA Subtypes), and build predictive models using three different modalities of data (histopathological images, transcriptomics and proteomics). Following which, we identify critical multi-level features that drive successful classification of patients for the various different cohorts. To build predictors for each data type, we employ widely used "best practice" techniques including CNN-based (convolutional neural network) classifiers for histopathological images and regression models for proteogenomic data. While, as expected, histology images outperformed molecular features while predicting cancer stages, and transcriptomics held superior discriminatory power for ER-Status and PAM50 subtypes, there exist a few cases where all data modalities exhibited comparable performance. Further, we also identified sets of key genes and proteins whose expression and abundance correlate across each clinical cohort including (i) tumor severity and progression (incl. GABARAP), (ii) ER-status (incl.ESRl) and (iii) disease subtypes (incl. FOXCl). Thus, we quantitatively assess the efficacy of different data types to predict critical breast cancer patient attributes and improve disease characterization.