Anticlustering for Sample Allocation To Minimize Batch Effects

High throughput sequencing is a powerful tool for processing large amounts of DNA and RNA samples in batches. Proper experimental design and statistical methods are required to mitigate systematic technical factors due to differences in batches ("batch effects"), as data variation due to these non-biological factors can mask actual biological differences. We propose using anticlustering as an automated method to assign samples to balanced batches. Anticlustering effectively negates differences in (numeric and/or categorical) covariates among batches, and implements user-defined restrictions on the number of batches, the number of samples per batch, and whether to assign certain samples to the same batch ("must-link constraints"). A simulation study shows that anticlustering is better at achieving balance among batches than previous approaches. An application from the UCSF-Stanford Endometriosis Center for Discovery, Innovation, Training and Community Engagement ("ENACT", https://enactcenter.org/) is presented as a real-life example. In the application, multiple samples provided by an individual had to be processed on the same batch, so that comparisons among different samples of the same patient were not diluted by batch effects. The novel Two Phase Must Link (2PML) anticlustering algorithm realized the must-link restrictions while simultaneously obtaining balance among batches regarding disease stage, menstrual cycle phase, case versus control sample, and clinical site. All methods presented here are accessible via the free and open source R package anticlust (https://cran.r-project.org/package=anticlust). An interactive visualization and web-based batch assignment tool are made available in the Rshiny app "anticlust" (https://anticlust.org/).