The functional interpretation of genetic variation in the context of disease requires robust phenotypes that are amenable to experimental perturbation. Given the complexity of many genetic disorders, relevant genetic variants could ideally be evaluated within the genetic background of patient-cells. For most conditions however, such phenotypes and/or model systems remain unknown. In this talk, I will introduce how functional profiling of patient cells offers a powerful means to address these challenges. In particular, I will focus on how to leverage deep representation learning to unlock and data-mine cellular morphology as a cost-efficient (e.g. compared to scRNAseq) and particularly rich domain of cell-biology. A key challenge to the utility of both supervised and un-/self-supervised representation learning on fluorescent microscopy data however, has been the need to carefully analyze and disentangle correlate technical variation (often manifesting as “batch-effects”) from biologically meaningful signal, in order to ensure out-of-distribution (OOD) generalization. I will introduce several approaches to overcome this challenge, and will deep dive into a novel method, based on generative interventions, which facilitates batch-effect correction, improves OOD generalization, and thus opens the door to powerful new strategies to address road blocks in rare disease genomic medicine and beyond.