A benchmark of computational methods for correcting biases of established and unknown origin in CRISPR-Cas9 screening data

CRISPR-Cas9 dropout screens stand as formidable tools for investigating biology with unprecedented precision and scale. One of their principal applications involves probing large panels of immortalised human cancer cell lines for viability reduction responses upon systematic genetic knock-out at a genome-wide level, to identify novel cancer dependencies and therapeutic targets. However, biases in CRISPR-Cas9 screens data pose challenges, leading to potential confounding effects on their interpretation and compromising their overall quality. The mode of action of the Cas9 enzyme, exerted by the induction of DNA double-strand breaks at a locus targeted by a specifically designed single-guide RNA (sgRNA), is influenced by structural features of the target site, including copy number amplifications (CN bias). More worryingly, proximal targeted loci tend to generate similar gene-independent responses to CRISPR-Cas9 targeting (proximity bias), possibly due to Cas9-induced whole chromosome-arm truncations or other unknown genomic structural features and different chromatin accessibility levels. Different computational methods have been proposed to correct these biases in silico, each based on different modelling assumptions. We have benchmarked seven of the latest methods, rigorously evaluating for the first time their ability to reduce both CN and proximity bias in the two largest publicly available cell-line-based CRISPR-Cas9 screens to date. We have also evaluated the capability of each method to preserve data quality and heterogeneity by assessing the extent to which the processed data allows accurate detection of true positive essential genes, established oncogenetic addictions, and known/novel biomarkers of cancer dependency. Our analysis sheds light on the ability of each method to correct biases arising from structural properties and other possible unknown factors associated with CRISPR-Cas9 screen data under different scenarios. In particular, it shows that AC-Chronos outperforms other methods in correcting both CN and proximity biases when jointly processing multiple screens of models with available CN information, whereas CRISPRcleanR is the top performing method for individual screens or when CN information is not available for the screened models. In addition, Chronos and AC-Chronos yield a final dataset better able to recapitulate known sets of essential and non-essential genes. Overall, our investigation provides guidance for the selection of the most appropriate bias-correction method, based on its strengths, weaknesses and experimental settings.