Spherical Phenotype Clustering

Phenotypic screening experiments comprise many images of the same cells perturbed in different ways, with biologically significant variation often subtle or difficult to see by eye. The specialized nature of the morphological changes and the fact that large quantities of data can be produced quickly makes training new machine learning models attractive. A byproduct of the experimental setup is knowledge of which well an image originated from and the treatment applied. This contrasts with consumer images which do not guarantee an associated categorisation. We propose a non-parametric variant of contrastive learning incorporating this metadata. The method is tested on the BBBC021 benchmark dataset and in HaCat cells treated with the JUMP reference compound set. On BBBC021 we attain higher NSC and NSCB scores than existing unsupervised (or weakly supervised) methods. In the HaCat cells we attain significantly better quantitative results (>10%) than CellProfiler or SimCLR and qualitative clustering reflecting underlying biology.