Theses (MSc. Molecular Medicine)

Nasopharyngeal Carcinoma (NPC) is a cancer of the cells lining the nasopharynx where distant metastasis has been the cause of treatment failure. Most cases were diagnosed at loco-regionally advanced stages due to several reasons including symptoms being often inconclusive at early stages, the aggressive nature of NPC tumours with early local and lymphatic infiltration and some challenges in the initial diagnosis stage during imaging screening including challenging positioning of the tumour as well as the subjective judgment of the clinician. Thus, being able to identify tumours that could potentially metastasise would offer better chances of improving patient prognostication. Here, this study aimed to test whether machine learning could be utilised to predict NPC metastasis using a transcriptomic dataset obtained from RNA sequencing. The algorithms tested were supervised algorithms including Naïve Bayes, k-Nearest Neighbours (kNN), Neural Network, Logistic Regression, Random Forest and Support Vector Machine (SVM). Unsupervised algorithms including hierarchical clustering and multidimensional scaling (MDS) were also performed to test class separation. Using leave one out cross validation (LOOCV) method, the algorithms were trained and evaluated using an internal dataset that has been filtered with the top 3 differentially expressed genes (DEGs) ranked by ReliefF based on their importance. The trained algorithms were then validated against an independent dataset with known classes. For M1 and M0 classification, the highest prediction performance was achieved by SVM, Neural Network and Logistic Regression where all three models classified 6/6 M0 and 6/6 M1 accurately with a perfect score of 1.00 for Area under the ROC Curve (AUC), calling accuracy (CA), F1, precision, and recall metrics. Additionally, the classification of NPC and non-NPC samples was also evaluated. The validation using an external dataset with Logistic Regression, Naïve Bayes, and Random Forest models managed to predict 7/7 NPC samples and 3/4 non-NPC samples and achieved the highest score of 0.909 for accuracy, 0.906 for F1, 0.920 for Precision and 0.909 for Recall Based on the performance metrics score, Logistic Regression consistently performed the best across both predictions and it can be concluded that machine learning appears to be a valid tool used in the prediction of cancer metastasis. Thus, potentially these models could be an accurate, assistive method to complement the current diagnostics tool for the prognostication of NPC patients.