PAUSE: principled feature attribution for unsupervised gene expression analysis.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Deep learning models are increasingly used for analyzing gene expression data.
• Interpretability of these complex models remains a significant challenge in bioinformatics.
• Current interpretability methods are often post hoc or rely on inherently interpretable models.

Purpose of the Study:

• To propose a novel method for enhancing the interpretability of unsupervised deep learning models in gene expression analysis.
• To demonstrate that combining different interpretability approaches can be beneficial.
• To introduce PAUSE, an unsupervised pathway attribution method.

Main Methods:

• Development of PAUSE (Pathway Attribution Using Self-Supervision), an unsupervised pathway attribution method.
• Integration of PAUSE with biologically-constrained neural network architectures.
• Application to gene expression data analysis.

Main Results:

• PAUSE effectively identifies major sources of transcriptomic variation.
• Combining PAUSE with biologically-constrained models yields improved interpretability.
• The proposed approach offers a synergistic combination of interpretability strategies.

Conclusions:

• PAUSE provides a powerful tool for understanding deep learning models in genomics.
• The combination of pathway attribution and biologically-constrained models represents a promising direction for interpretable AI in biology.
• This work facilitates deeper biological insights from complex gene expression datasets.