On potential limitations of differential expression analysis with non-linear machine learning models

• DOI: 10.14806/ej.28.0.1035
• Publication Date: 2023-03-08T15:01:35Z
• Authors: Gianmarco Sabbatini; Lorenzo Manganaro
• Keywords: Subtyping; Expression (computer science)

Recently, there has been a growing interest in bioinformatics toward the adoption of increasingly complex machine learning models for analysis next-generation sequencing data with goal disease subtyping (i.e., patient stratification based on molecular features) or risk-based classification specific endpoints, such as survival. With gene-expression data, common approach consists characterising emerging groups by exploiting differential expression analysis, which selects relevant gene sets coupled pathway enrichment providing an insight into underlying biological processes. However, when non-linear are involved, could be limiting since groupings identified model set genes that hidden to due its linear nature, affecting subsequent characterisation and validation. The aim this study is provide proof-of-concept example demonstrating limitation. Moreover, we suggest issue overcome innovative paradigm eXplainable Artificial Intelligence, building additional explainer get trustworthy interpretation outputs reliable each group, preserving also relations, used downstream