Autoencoder Assisted Cancer Subtyping by Integrating Multi-omics Data

Cancer subtypes identification can facilitate the subtype-targeted treatments for different cancers. The heterogeneous nature of the disease involves activation of several pathways, significantly affecting the patients’ survival and displays distinctive efficacy towards different drugs. Therefore, identification of cancer subtypes using genomic level data is critical. Over the years, several computational methods have been designed for multi-omics data integration and clustering. These methods mainly follow three algorithmic approaches: early, late, or intermediate integration of multi-omics data. Some of them perform clustering on concatenated data, other performs clustering on integrated similarities, whereas other, performs clustering on new representations. In this study, a deep learning framework of the Autoencoder is designed to obtain a low-dimensional representation of the early integrated multi-omics dataset. Spectral clustering is performed on the bottleneck layer of the Autoencoder to predict patients’ clusters (cancer subtypes). The Performance of the proposed Autoencoder-assisted workflow is demonstrated on three different cancer data sets taken from The Cancer Genome Atlas database. The performance is also compared with other popular early, late, and intermediate data integration methods. Furthermore, to establish the biological relevance of the identified clusters, a detailed biological analysis of the clusters obtained from the Glioblastoma multiforme dataset (GBM) is also presented.