15:00
17:00

Biomedical research relies on omics data, such as transcriptomics, to decipher the complexity of diseases such as cancer. Traditional methods for identifying transcriptomic biomarkers associated with a phenotype rely on variable selection techniques. Hybrid Ensemble Feature Selection (HEFS) methods have become increasingly popular in the last decade for their ability to guarantee the robustness of selected variables across functional and data perturbations. However, their design remains a challenge. Our first contribution aims to evaluate the HEFS-type strategy by analyzing in detail 4 scenarios applied to the identification of biomarkers in various cancers from RNA-Seq transcriptomics data. These scenarios explore combinations of two variable reduction methods (differentially expressed genes and variance) with two resampling strategies (repeated holdout by stratified balanced distribution and random stratified) for variable selection downstream of signature aggregation of thousands of wrapped machine learning models. Our results highlight the advantages of using HEFS approaches to identify biomarkers of complex diseases, given their ability to produce generalizable and stable results. In addition, we highlight critical considerations essential for the design of such strategies. Our second contribution is the development of an atlas of transcriptomic biomarker signatures from multiple cancers at various stages of progression. The integrated study of this atlas with proven biological annotation data (e.g. REACTOME) should enable the identification of genes of interest for various oncological phenotypes.

Amphi E (Bâtiment A29)