Embedding Techniques for Biomedical and Biological Data: Interpretable Methods and Applications

The exponential growth of biomedical and biological data, such as histopathological images, genomic sequences, and molecular networks, calls for advanced techniques to represent such data in compact, informative, and interpretable ways. Embedding methods offer an effective solution by transforming high-dimensional and complex data into representations that facilitate classification, pattern discovery, and knowledge extraction tasks. This thesis investigates a broad spectrum of biomedical and biological data embedding approaches, including classical dimensionality reduction techniques, metric learning models, granular-based approaches, and graph-based representations. The central goal is to combine high predictive performance with interpretability, crucial in clinical and scientific domains where understanding the rationale behind automated decisions is essential. A key contribution of this work is the development of GECo Graph Explanation by Communities, a novel algorithm designed to improve the explainability of Graph Neural Networks. GECo leverages community detection to identify meaningful substructures within graphs, providing clear and consistent explanations for Graph Neural Networks predictions, particularly in biomedical and biological settings. Through extensive experimentation on both public and real-world datasets, the thesis demonstrates the effectiveness of embedding techniques in addressing practical biomedical challenges while ensuring transparency, robustness, and interpretability. This work contributes to advancing trustworthy and explainable artificial intelligence systems for biomedical and biological data analysis.