A1008
Title: Estimating causal effects with proximal inference methods in single-cell CRISPR screens
Authors: Catherine Wang - Carnegie Mellon University (United States) [presenting]
Kathryn Roeder - Carnegie Mellon University (United States)
Larry Wasserman - Carnegie Mellon University (United States)
Abstract: Single-cell CRISPR screen experiments allow for the tests and estimation of causal effects by perturbing a genomic region of interest and measuring changes in gene expression; however, insufficient adjustment for confounding may result in bias, producing large false positive rates. Typically, when the confounding variables are unmeasured, casual estimands cannot be identified and estimated, but the proximal setup allows for causal inference by taking advantage of two types of proxy variables: treatment-inducing confounding proxies and outcome-inducing confounding proxies. In the single-cell CRISPR setting, scientists measure many CRISPR perturbations and genes' expressions that may act as such proxies. A methodology for analyzing causal effects between many perturbation-gene pairs using proximal inference is proposed. Estimation procedures are explored, including two-stage least squares approaches, generalized method of moments approaches, and a plug-in and one-step corrected approach. These are designed to take advantage of the wide selection of proxies, each of which may be individually weak. Finally, the assumptions required for proximal inference are discussed and evaluated, and results are shown in simulated and real-world single-cell CRISPR screens.
