A comparative analysis of gene expression profiling by statistical and machine learning approaches.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Machine learning (ML) models are used to classify phenotype from gene expression data.
• Model interpretations, often gene importance rankings, aim to elucidate biological processes.
• Integrated gradients is a common explainability method for neural networks.

Purpose of the Study:

• To discuss the limitations of gene importance rankings from ML models for understanding biological processes.
• To evaluate the effectiveness of integrated gradients in identifying biologically relevant genes from gene expression data.
• To compare gene rankings from integrated gradients with statistical methods.

Main Methods:

• Experiments were conducted using RNA sequencing data from public cancer databases.
• Multilayer perceptrons and graph neural networks were trained for cancer type classification.
• Gene rankings from integrated gradients were compared against DESeq2 and other feature selection methods.

Main Results:

• A small set of top-ranked genes achieved good classification performance.
• Similar classification performance was achievable with lower-ranked genes, albeit requiring larger sets.
• Significant discrepancies in top-ranked genes between statistical and ML methods were observed, hindering comprehensive biological interpretation.

Conclusions:

• Explainability techniques can identify pathology-specific biomarkers.
• The completeness of gene sets selected by these techniques for understanding biological processes remains uncertain.
• Further research is needed to refine ML interpretability for robust biological insights.