Hamza Zidoum is an Assistant Professor in the Department of Computer Science at Sultan Qaboos University (SQU), Oman, with a long academic and international research career in computer science, artificial intelligence, and data science. He earned his PhD and MSc in Computer Science from Franche-Comté University, France, and holds an engineering degree in Electronics and Control/Automation from Ferhat Abbas University, Algeria. His academic experience spans multiple countries and institutions, including Monash University (Australia), UAE University, Bilkent University (Turkey), and several French universities, before joining Sultan Qaboos University in 2002, where he continues to teach and conduct research. His teaching portfolio covers machine learning, programming in Python, data structures and algorithms, artificial intelligence, intelligent systems, and advanced topics in deep learning and data science. His research and supervision work focuses on machine learning, deep learning, medical image analysis, natural language processing, optimization, and applied AI, with recent interests including large language models, neural differential equations, and explainable AI. He has supervised numerous MSc and PhD theses in areas such as healthcare analytics, solar energy forecasting, cybersecurity, and computer vision. He is also active in research projects funded internally at SQU, particularly in AI for medical imaging, disease prediction, and optimization systems, and contributes to curriculum development, accreditation, and academic committees. Alongside teaching and research, he has played an active role in mentoring students, organizing programming competitions, and contributing to applied AI research relevant to Oman’s national priorities.

Beyond Black Boxes: How Autoencoder Architecture Shapes Multi-Omics Representations in Breast Cancer Can the wiring of a neural network shape what it learns about cancer biology — without ever seeing a clinical label?This talk presents a controlled study of how biological structure influences unsupervised representation learning from multi-omics breast cancer data. Standard autoencoders treat thousands of molecular features as flat, unstructured inputs. We ask a simple but underexplored question: what happens when we encode known protein–protein interaction (PPI) networks directly into the model architecture or training objective? We compare four autoencoder variants — from a plain baseline to a graph convolutional model — trained on 941 breast cancer patients across four omics modalities (TCGA-BRCA), keeping everything constant except the form of structural bias. No clinical labels are used during training.The results are striking. Introducing PPI structure via the loss function dramatically stabilises learned representations and preserves latent geometry. Encoding it into the architecture goes further — producing the only model to achieve significant survival stratification (p = 0.024), on par with the gold-standard PAM50 molecular subtypes. Clinical signals invisible to unstructured models — including metastatic and nodal status — emerge spontaneously across 61–86% of latent dimensions in graph-based models. But structure is not a free lunch. Each form of bias introduces specific trade-offs between stability, dimensionality collapse, and reconstruction quality. The key message: how you incorporate biological knowledge matters as much as whether you incorporate it.