Electronic Theses and Dissertations (PhDs)

Permanent URI for this collection

https://hdl.handle.net/10539/38017

Browse

Browsing Electronic Theses and Dissertations (PhDs) by Keyword "Deep learning"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    A Phenotype Prediction Framework for Classifying Colorectal Cancer Patients’ Response to FOLFOX Treatment: An Integrated Approach
    (University of the Witwatersrand, Johannesburg, 2024) Mashatola, Lebohang; Kaur, Mandeep
    Colorectal cancer (CRC), characterised by its prevalence and heterogeneity, poses a significant challenge in understanding drug resistance, especially in the context of FOLFOX treatment. This study presents an innovative methodology that integrates diverse data analysis approaches to address the challenge of predicting the phenotype of CRC patients resistant or sensitive to FOLFOX. The initial analysis involved dierential and co-expression analyses, identifying pivotal hub genes crucial to drug resistance in CRC, regulating intricate molecular networks. Subsequent enrichment analysis revealed their significant roles in biological functions, particularly influencing DNA repair and nuclear division. To capture inherent topological characteristics within genetic expression data, a novel technique utilising topological data analysis (TDA) was employed. By applying persistence homology to generate persistence images, the Vietoris-Rips complex was constructed using the signed-topological overlap matrix, comprehensively capturing numerous topological features, including high-dimensional Betti-1 and Betti-2. This provided valuable insights into the structural patterns of gene expression between the hub genes. Furthermore, the integration of whole-slide images enhanced understanding of tissue anatomy, which is crucial for predicting cancer stages. Using a MobileNet architecture, a deep learning model classified cancer stages, contributing to a holistic understanding of colorectal tumor microenvironments. For predictive modelling of drug resistance, a multilayer perceptron applied topological summaries generated by TDA. The developed framework, GeTopology, exhibited remarkable performance metrics, achieving an overall 83% accuracy in predicting the FOLFOX response, demonstrating a 3% improvement over a previously published phenotype prediction framework (NSCLC ) that utilised similar data modes. Robust accuracies were consistently observed in independent datasets, classifying both cancer patients and healthy individuals. The results indicated an approximate 10% increase in model prediction accuracy compared to NSCLC, emphasising the potential clinical impact of this integrative approach. In conclusion, this study advances the understanding of drug resistance in CRC by proposing a novel approach that integrates topology with histopathological images, oering transformative insights into predictive modelling and precision medicine