XA4C: eXplainable representation learning via Autoencoders revealing Critical genes

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• Representation Learning (RL) models like autoencoders are powerful for transcriptome analysis but lack interpretability.
• Identifying essential genes for functional studies is challenging with current RL methods.
• Traditional methods like Differentially Expressed (DiffEx) genes may not capture complex gene interactions.

Purpose of the Study:

• To develop an interpretable machine learning approach for identifying critical genes in transcriptome data.
• To propose a novel definition of "Critical genes" as those contributing significantly to learned representations.
• To validate the utility of Critical genes in understanding cancer biology.

Main Methods:

• Implemented an eXplainable Autoencoder for Critical genes (XA4C) using eXplainable Artificial Intelligence (XAI).
• XA4C quantifies gene contribution to latent variables for prioritizing Critical genes.
• Applied XA4C to gene expression data from six cancer types.

Main Results:

• Critical genes identified by XA4C capture essential cancer pathways.
• Critical genes show minimal overlap with traditional Hub or DiffEx genes.
• Critical genes exhibit higher enrichment in disease (DisGeNET) and cancer (COSMIC) gene databases, suggesting novel biological discoveries.

Conclusions:

• XA4C enables interpretable transcriptome analysis using RL.
• Critical genes identified through explainable RL offer a powerful tool for studying complex gene interactions in cancer.
• This approach has the potential to uncover significant unknown biological mechanisms.