An unsupervised multi-omics clustering based on ensemble deep learning reveals subgroups of cancer patients

Abstract

ancer is a complex disease caused by the abnormal behavior and interaction of different bio entities (e.g., genes, proteins and epigenetic factors), which are profiled using omics technologies (e.g., transcriptomics, proteomics and epigenomics). Traditionally, each omic have been analyzed individually through different methods, however, the major source of the cancer complexity lies on that interaction among the elements involved. Amongst the most common computational techniques in this field, machine learning, a branch of artificial intelligence that builds data-driven models, is the key due to its capacity to transform biological data into knowledge. Ensemble machine learning and deep learning, two areas which are making a substantial impact on the field, have been usually treated as independent methodologies. Besides, most of research focused on supervised learning, that requires a prior knowledge for the labelling of the dataset. Therefore, we developed MOEDC (Multi-Omics Ensemble Deep Clustering), an unsupervised multi-omics (transcriptomics-proteomics-epigenomics) clustering based on ensemble deep learning for stratifying cancer patients. MOEDC was developed using the kidney renal clear cell carcinoma (KIRC) dataset from The Cancer Genome Atlas (TCGA), where it found two clusters with different prognosis and an accuracy comparable with state-of-the-art (SOTA) models. To further validate MOEDC, we used it for clustering the bladder urothelial carcinoma (BLCA) dataset, where it found three clusters that were more associated to clinical features than those generated by the SOTA models. We elucidated the clinical and biological characteristics of the clusters of both datasets through differential analysis by showing key biomarkers that might be useful in future applications. Thus, MOEDC worked successfully on stratifying cancer patients and could be used on other similar tasks.