ScaiDigest August 2023

“Thrilled to post the second edition of ScaiDigest – unraveling the cutting-edge research in our field! Today I’d like to highlight 2 recent publications on model interpretability –

Model interpretability is a major bottleneck in the adoption of artificial intelligence in the biological sciences. The ability to understand the reasoning behind model predictions is of crucial importance for translation to the bedside, as well as to build trust amongst biologists and clinicians.

I’d like to share two recent articles which demonstrate methods explicitly incorporating prior biological knowledge into the modeling and analysis of single-cell sequencing data – expiMap from Lotfollahi et al, and MAYA from Landais and Vallot.

While expiMap constructs a directly biologically-interpretable latent space based on both pre-defined and de novo-identified gene programs, MAYA enables the in-depth analysis of pathway activation allowing for the discovery of shared activation modes across cell populations and tumors. These tools are pushing the frontiers in the use of machine learning in biology, driving a shift towards the democratization of biological insights.

Having a background in both data science and biology, these two pieces particularly resonated with me, and I am excited by their ability to bring tangible biological insights from complex data such as single-cell RNA-sequencing enabling everyone to peer inside the conventional black box of artificial intelligence.

expiMap – https://www.nature.com/articles/s41556-022-01072-x

MAYA – https://www.nature.com/articles/s41467-023-37410-2“