Abstract
Modern cell-perturbation experiments expose cells to panels of hundreds of treatments, such as cytokines or CRISPR guides. These experiments are designed to investigate whether a particular gene is up- or downregulated by exposure to each treatment. Typical estimators of whether a gene is up- or down-regulated make many errors due to high levels of experimental noise and complex donor and batch effects. In this article we explore whether estimator error rates can be reduced by pooling measurements according to auxiliary information, such as prior knowledge of treatments that are likely to have similar effects. Because auxiliary information can be misleading and because it is generally infeasible to model its quality, we take a model-free approach to identifying a subset of estimates in which error is reliably controlled. We validate this overall approach in simulation, comparing it to an alternative based on the Benjamini–Hochberg procedure. We then use the proposed approach to reanalyze a high-throughput CRISPRa screening and a small-molecule drug screening, which leads to new discoveries about the effects of gene augmentation on immune cells and the effects of cytokines on cancer cells.

This publication has 8 references indexed in Scilit: