Debiased Projected Two-Sample Comparisonscfor Single-Cell Expression Data

We study several variants of the high-dimensional mean inference problem motivated by modern single-cell genomics data. By taking advantage of low-dimensional and localized signal structures commonly seen in such data, our proposed methods not only have the usual frequentist validity but also provide useful information on the potential locations of the signal if the null hypothesis is rejected. Our method adaptively projects the high-dimensional vector onto a low-dimensional space, followed by a debiasing step using the semiparametric double-machine learning framework. Our analysis shows that debiasing is unnecessary under the global null, but necessary under a “projected null” that is of scientific interest. We also propose an “anchored projection” to maximize the power while avoiding the degeneracy issue under the null. Experiments on synthetic data and a real single-cell sequencing dataset demonstrate the effectiveness and interpretability of our methods.