10th U6+ Consortium of African Universities International Conference

Non-clinical diagnostics and predictive modeling in statistical genomics increasingly rely on artificial intelligence to extract clinically meaningful insights from high-dimensional molecular data. However, translating such models from research to practical use remains challenging due to the computational complexity of genomic datasets and the need for scalable solutions that can operate in resource-constrained environments. This study presents a scalable multiobjective machine learning framework for efficient biomarker gene selection and diagnostic prediction using microarray gene expression data from colorectal carcinoma samples. Using Asymetrix oligonucleotide arrays, gene expression profiles from 62 tumour and normal tissue samples were analyzed to identify a minimal yet highly informative biomarker set capable of robust classification. The proposed hybrid approach combines a two-stage feature selection strategy designed to reduce computational burden while preserving predictive performance. First, a Bayesian t-test was applied to eliminate non-informative genes and substantially reduce dimensionality. Second, a multi-objective optimization strategy integrated with a Support Vector Machine (SVM) classifier sequentially selected genes by balancing classification accuracy, model simplicity, and generalizability. This design prioritizes model parsimony, enabling efficient inference and reducing the computational demands typically associated with high-dimensional genomics. From an initial set of approximately 2,000 genes, the framework identified only three (3) biomarker genes that achieved 90.1% prediction accuracy on independent test data, demonstrating strong robustness and transferability. Unsupervised validation using principal component analysis and complete-linkage hierarchical clustering showed near-perfect separation between tumour and normal samples, supporting the biological relevance of the selected biomarkers. Overall, this work demonstrates that high-performance genomic AI models can be made computationally efficient, interpretable, and scalable. By transforming highdimensional data into lightweight predictive models, the proposed framework supports practical deployment of AI-driven diagnostics in settings with limited computational infrastructure.