Statistical Challenges in Sequencing Data: Addressing Uncertainty in Analysis

The advancement of high-throughput sequencing technologies has revolutionized biological research. However, these technologies also introduce new statistical uncertainties. In this work, we present novel statistical methods that account for statistical uncertainties and best practice to mitigate bias in analyzing sequencing data. In project 1, we focus on detecting allelic imbalance on the isoform level, caused by non-coding variants in the regulatory regions. This analysis requires accounting for inferential uncertainty, caused by multi-mapping of RNA-sequencing reads. Our proposed method, SEESAW, uses Salmon and Swish to offer analysis at various levels of resolution, while accounting for inferential uncertainty. In project 2, we address the challenge of detecting cell composition changes using paired single-cell RNA-sequencing (scRNA-seq) data in cancer research. We proposed a novel approach scSTMseq, a topic modeling-based framework designed to infer cell clusters and estimate clustering uncertainty while integrating cell-level covariates, such as time. We coupled this framework with Multivariate analysis of variance for repeated measures to infer cell composition changes. In project 3, we explored the CRISPR inference/activation screens to study the regulatory roles of non-coding elements. Unlike standard scRNA-seq, scaling-based normalization introduces selection bias due to correlations between perturbation states and the sequencing depth, which complicates the case-control classification. We demonstrated this bias and proposed Poisson regression as a better alternative for normalization in the CRISPR-i/a studies.