Best Practices and Recommendations for Bulk RNA-Seq Analysis: A Comparative Evaluation of Common Imputation Techniques and Widely Used Complete-Case Analysis

Abstract Bulk RNA-Seq data is widely used to identify differentially expressed genes (DEGs) between groups of samples. Typically, genes with missing values are discarded in initial analyses, resulting in a loss of information. Although only a few studies apply imputation, no single method is considered the gold standard, as appropriate choice of imputation should ideally depend on the underlying cause of missingness. But identifying this cause is often challenging. Consequently, many studies rely on complete-case (CC) analysis due to its simplicity, but that introduces bias. Our simulations show that while CC controls false-positives, it lacks sensitivity, leading to under detection of truly disease-associated genes. Moreover, simple mean or median imputations may also result in bias when missing data arises from biological or technical factors, which is often a possibility in RNA-Seq data. As expression of a gene may correlate with co-regulated genes, missing gene counts could be imputed based on similarly expressed genes. The concept of imputing missing values in K-nearest neighbor (KNN) imputation algorithm aligns aptly with this biological phenomenon. If a gene count is missing, KNN imputation estimates this value using expression data from K other genes with similar expression patterns. In this article, using simulations we show that common imputation techniques for detecting DEGs perform better than CC analysis under varying percentages of missing data. On application to in-house real-data on Acute Myeloid Leukemia, we identified that compared to CC, common imputations detect more DEGs. Additionally, we provide recommendations for handling genes with low counts across all samples.