Directory
A comparative analysis of gene expression profiling by statistical and machine learning approaches.
Bioinform Adv, 5(1):vbae199.
MOTIVATION: Many machine learning (ML) models developed to classify phenotypefrom gene expression data provide interpretations for their decisions, with theaim of understanding biological processes. For many models, including neuralnetworks, interpretations are lists of genes ranked by their importance for thepredictions, with top-ranked genes likely linked to the phenotype. In thisarticle, we discuss the limitations of such approaches using integrated gradient,an explainability method developed for neural networks, as an example. RESULTS:Experiments are performed on RNA sequencing data from public cancer databases. Acollection of ML models, including multilayer perceptrons and graph neuralnetworks, are trained to classify samples by cancer type. Gene rankings fromintegrated gradients are compared to genes highlighted by statistical featureselection methods such as DESeq2 and other learning methods measuring globalfeature contribution. Experiments show that a small set of top-ranked genes issufficient to achieve good classification. However, similar performance ispossible with lower-ranked genes, although larger sets are required. Moreover,significant differences in top-ranked genes, especially between statistical andlearning methods, prevent a comprehensive biological understanding. Inconclusion, while these methods identify pathology-specific biomarkers, thecompleteness of gene sets selected by explainability techniques for understandingbiological processes remains uncertain. AVAILABILITY AND IMPLEMENTATION: Pythoncode and datasets are available at https://github.com/mbonto/XAI_in_genomics.
Document Actions
