Reproduce “Interpretable deep learning in single-cell omics” on OpenResearch

About this paper

Manoj M Wagle, Siqu Long, Carissa Chen, Chunlei Liu, Pengyi Yang

Researchers from the University of Sydney offer a comprehensive review of interpretable deep learning models used in single-cell omics, systematically categorizing methods by their biological applications and interpretability strategies. This work synthesizes the current landscape to address the "black box" nature of deep learning, providing a structured overview of techniques for identifying key molecular features, gene programs, and networks, while also outlining challenges and future research directions.

Key insights

• Researchers employ a dual strategy for interpretability, utilizing both post-hoc feature attribution techniques and intrinsically designed models for transparency.
• Integrating prior biological knowledge (e.g., pathways, gene modules) into model architectures is a common and effective approach to enhance interpretability.
• A notable trend is the development of interpretable deep learning models specifically for multimodal single-cell data, reflecting advances in omics technologies.

Summary by alphaXiv.