Data-driven identification of Alzheimer’s Disease biomarkers from transcriptomic profiles using machine learning

Alzheimer’s Disease (AD) presents diagnostic and therapeutic challenges due to its complex mechanism and the limitations of conventional biomarker discovery methods. While differential expression (DE) analysis remains a common strategy in transcriptomic research, its univariate nature fails to uncover robust cross-cohort signals. This dissertation proposes a machine learning-based multivariate framework for identifying generalizable transcriptomic biomarkers of AD. By looking at three independent hippocampal datasets (VR, MAYO, and GSE) across RNA-Seq and microarray platforms, this work exhaustively evaluates gene-pair interactions using bivariate SVM classifiers built upon Relief-ranked features. The analysis reveals a set of candidate genes - including KCNIP1, CA10, CSPG5, and SLC38A2 - whose predictive power emerges through combinatorial patterns across studies. In my dissertation, I will mainly cover the methods and results of my recent journal paper titled "Cross study transcriptomic investigation of Alzheimer’s brain tissue: discoveries and limitations". This will consist of a brief introduction to the traditional differential expression analysis and its limitations, followed by how we integrated Relief and SVMs into a single pipeline. Robustness is assessed via noise injection, permutation testing, and cross-study selectivity metrics, demonstrating consistent performance under varying thresholds and perturbations. This study explores gene-gene interactions potentially involved in AD pathology but also establishes a reproducible computational pipeline for multivariate biomarker discovery with applications beyond Alzheimer's disease. The resulting methodology demonstrates the importance of generalization, interpretability, and reproducibility in transcriptomic biomarker research.