Hyperbolic Representation Learning to Capture Biological Relationships – Elisabeth Wetzer (UiT, Online)

Hyperbolic representation learning has demonstrated significant advantages in modeling hierarchical relationships within data compared to conventional Euclidean approaches. In this presentation its potential to capture biological relationships between cell types in highly multiplexed imaging data will be presented, which is critical for understanding tissue composition and functionality. Assessing the quality of learned representations that align with biological domain knowledge across geometrical spaces with differing metrics poses a challenge. Here, we propose to use an information-theoretic framework based on k-nearest neighbor estimators to compute mutual information for performance evaluation. Our findings reveal that hyperbolic embeddings preserve more biologically relevant relationships than Euclidean embeddings.

About Elisabeth Wetzer