Machine Learning at Genome Scale

ties of training data to be able to generalize well.In many biology-related applications, datasets can have few samples and a high number of features, which makes it challenging to apply ML/DL models.This issue is known as the "curse of dimensionality", a well known problem that can limit the efficacy of ML/DL in the biological field.Omics data can be of several types, for example: transcriptomics, proteomics, or epigenomics.Each one represents a different aspect of the biological system, and the approach of simultaneously analyzing and integrating them is referred to as multi-omics analysis.This integration is crucial because it provides a more complete and holistic view of the biological system, which would be impossible using any single omic layers alone.However, multi-omics integration is not always straightforward, as it requires approaches that handle the heterogeneity across layers while keeping the biological relevance of each individual data type and for this reason it remains an open challenge in biology.The PhD thesis fits into this broader context and is framed within the European project E-MUSE: Complex microbial Ecosystems MUltiScale modElling (https:// www.itn-emuse.com/https://www.itn-emuse.com).E-MUSE is a Marie Sk lodowska-Curie Action Innovative Training Network focused on developing new methodologies for modeling complex biological systems.The project aims to improve our understanding of microbial ecosystems, with a particular emphasis on fermented food products such as cheese.The Work Package 2 specifically focuses on data-driven approaches.Among these, it explores the "curse of dimensionality" in omics dataset and multiomics integration using statistical, network-based, and ML/DL approaches.Within this framework, the PhD thesis explores the use of ML, in particular DL methods to address these challenges.The common theme of the work is the methodological development of ML approaches at genome-scale.