Evaluation of colorectal cancer subtypes and cell lines using deep learning

Colorectal cancer (CRC) is a common cancer with a high mortality rate and rising incidence rate in the developed world. Molecular profiling techniques have been used to study the variability between tumours as well as cancer models such as cell lines, but their translational value is incomplete with current methods. Moreover, first generation computational methods for subtype classification do not make use of multi-omics data in full scale. Drug discovery programs use cell lines as a proxy for human cancers to characterize their molecular makeup and drug response, identify relevant indications and discover biomarkers. In order to maximize the translatability and the clinical relevance of in vitro studies, selection of optimal cancer models is imperative. We present a novel subtype classification method based on deep learning and apply it to classify CRC tumors using multi-omics data, and further to measure the similarity between tumors and disease models such as cancer cell lines. Multi-omics Autoencoder Integration (maui) efficiently leverages data sets containing copy number alterations, gene expression, and point mutations, and learns clinically important patterns (latent factors) across these data types. Using these latent factors, we propose a refinement of the gold-standard CRC subtypes, and propose best-matching cell lines for the different subtypes. These findings are relevant for patient stratification and selection of cell lines for drug discovery pipelines, biomarker discovery, and target identification.