Biomarkers are important characteristics that indicate normal biological processes, pathogenic processes, and pharmacological responses, making biomarker development crucial in the fields of medicine and life sciences. Recently, various artificial intelligence models have been developed to identify potential biomarkers, and there is an increasing need for more accurate and reliable methodologies. However, since the biological characteristics of the human body result from complex interactions among multiple features, it is important to employ methodologies that effectively reflect these interactions. Therefore, this thesis proposes biomarker discovery methods that utilize information-theoretic analysis and graph analysis to accurately reflect interactions between features. In the first study, the focus is on developing a stable feature scoring method by replacing the mutual information formula, an information-theoretic relevance measurement method. This approach facilitates faster and more reliable computation of correlations between features and diseases, aiding in biomarker discovery. In the second study, we propose methods for creating and analyzing correlation graphs, using information-theoretic measurements to generate and interpret meaningful graphs. The proposed methods have been validated through comparative experiments in various environments. The first experiment demonstrated that the selected features could be potential biomarker candidates, while the second experiment showed that the generated networks could be useful for biomarker exploration. Subsequent experiments include proposals and analyses of feature selection methods that consider graph structures among samples.
