A reformulation of the t-test in single-cell transcriptomics dramatically reduces false positives

The default differential expression tests in single-cell analysis software provide fast and simple approaches that help practitioners annotate cell populations identified through clustering algorithms. Here, we use theoretical arguments and empirical studies to show that these tests suffer from inflated false positive rates that affect the resulting annotations. We introduce LN's t-test, which tests a novel, asymptotically unbiased log-fold change estimator and avoids these false positives, without compromising speed.