Spherical Phenotype Clustering
Contrastive learning approach using experimental metadata as learned class vectors to improve representations
Overview
Phenotypic screening experiments produce many microscope images of cells under diverse perturbations, with biologically meaningful responses that are often subtle or difficult to identify visually. The core challenge is extracting representations that reliably separate active compounds from controls and cluster phenotypically similar perturbations together.
This work proposes new adaptations of contrastive loss functions that incorporate experimental metadata as learned class vectors. The key contribution is SPC (Spherical Phenotype Clustering), a geometrically inspired variant that constrains class vectors to the unit sphere and updates them only via attractive terms — allowing phenotypically similar classes to naturally overlap rather than being forced apart.
Read the paper for more!