Abstract
Cancer Research has advanced during the past few years. Using high throughput technology and advances in artificial intelligence, it is now possible to improve cancer diagnosis and targeted therapy, by integrating the investigation and analysis of clinical and omics profiles. The high dimensionality and class imbalance of the majority of available data sets represent a serious challenge to the development of computational methods and tools for cancer diagnosis and biomarker discovery. Taking into account multi-omics data further complicates the undertaking. In this paper, we describe a five-step integrative architecture for dealing with the three aforementioned problems by incorporating proteomics data, protein-protein interaction networks, and signaling pathways in order to identify protein biomarkers with a direct association to cancerous patients' overall survival (OS) and progression free interval (PFI). The core parts of this architecture are a cluster based grey wolf optimization algorithm (CB-GWO) for feature selection and a deep stacked canonical correlation autoencoder (DSCC-AE) for clinical endpoint prediction. A thorough experimental study was carried out to evaluate the performance of the proposed optimization algorithm for feature selection, as well as the performance of the deep learning model in terms of Mathew coefficient correlation (MCC) and Area under the curve (AUC) on breast, lung, colon, and rectum cancers. The results were compared to other methods in the literature. The results are very promising and show the effectiveness of the proposed framework and its ability to outperform the other algorithms and models in terms of AUC (0.91) and MCC (0.64). In addition, hub marker genes with the potential occurence of alterations in colorectal cancer, breast cancer, and lung cancer have been identified.