Statistical and deep learning methods for cancer genomic data.

Doctoral Degree. University of KwaZulu-Natal, Pietermaritzburg.Statistical and machine learning methods have been applied in broad domains including the medical field. These methods have a massive impact on healthcare by providing the support for decision making to the specialist in diagnosis and prognosis of patient disease status and disease progression. Non-communicable diseases (NCDs) remain a major challenge the world over in the 21st century, especially in developing countries where resources are limited. Recent global public health research shows an epidemiological paradigm shift from infection to non-communicable diseases, which include cancer. Cancer is considered the most devastating among all NCDs and is ranked second to malaria as the leading causes of death in the developing countries. Cancer occurs in many different types affecting all community members, where the general mechanism of cancer disease etiology is uncontrolled cells proliferation that leads to a malignant or cancerous tumor, and abnormalities at the molecular level. However, earlier detection and accurate diagnosis of cancer symptoms increase the probability of curing the condition, which has become the best strategy for fighting the disease. In the past few years, a vast amount of cancer data have been generated through new high throughput technologies. Traditional clinical and experimental approaches lack the capacity to handle such a massive scale of data. Therefore, computational methods have been introduced to biomedical investigations, including genes/biomarkers selection of cancer types and stages of the disease. Many computational tools have been developed based on different statistical and machine learning strategies and data science approaches. We used statistical, machine and deep learning methods for cancer types, subtypes, and survival prediction in this work. First, we developed a hybrid (DNA mutation and RNA expression) signature and assessed its predictive properties for colorectal cancer (CRC) patients’ mutation status and survival. In addition, we proposed a stacking ensemble deep learning approach to evaluate and compare its predictive performance for cancer types (as a multi-class classification problem) with the different standard machine and deep learning methods. Finally, we assessed the predictive performance of the Cox proportional hazard and random survival forests methods based on a signature obtained using three gene mutations (KRAS, BRAF, and TP53). However, the most significant limitation lies in the sample size being small, and there is a lack of using independent data for validation. Also, we did not consider different features such as methylation and mutation data. Moreover, it is unfortunate that the study did not include detailed simulation studies to compare the traditional statistical and machine learning methods. Overall, the most prominent finding to emerge from this investigation is that combining different data sources leads to more robust statistical significance. Also, the stacking approach is more reliable and promising compared to a single machine or deep learning. Furthermore, the RSF is a proper and striking method for survival analysis since it does not depend on any model assumptions