Correcting nuisance variation using Wasserstein distance

• DOI: 10.7717/peerj.8594
• Publication Date: 2020-02-28T12:26:05Z
• Authors: Gil Tabak; Minjie Fan; Samuel Yang; Stephan Hoyer; Geoffrey Davis
• Keywords: Profiling (computer programming); Variation (astronomy)

Profiling cellular phenotypes from microscopic imaging can provide meaningful biological information resulting various factors affecting the cells. One motivating application is drug development: morphological cell features be captured images, which similarities between different compounds applied at doses quantified. The general approach to find a function mapping images an embedding space of manageable dimensionality whose geometry captures relevant input images. An important known issue for such methods separating signal nuisance variation. For example, vectors tend more correlated cells that were cultured and imaged during same week than those weeks, despite having identical in both cases. In this case, particular batch set experiments conducted constitutes domain data; ideal image embeddings should contain only (e.g., effects). We develop framework adjusting order “forget” domain-specific while preserving information. To achieve this, we minimize loss based on distances marginal distributions (such as Wasserstein distance) across domains each replicated treatment. dataset present results with, treatment happens negative control treatment, do not expect any treatment-induced morphology changes. our transformed (i) underlying geometric structure preserved but also carry improved signal; (ii) less present.