Chronos: a cell population dynamics model of CRISPR experiments that improves inference of gene fitness effects

Abstract

CRISPR loss of function screens are powerful tools to interrogate biology but exhibit a number of biases and artifacts that can confound the results. Here, we introduce Chronos, an algorithm for inferring gene knockout fitness effects based on an explicit model of cell proliferation dynamics after CRISPR gene knockout. We test Chronos on two pan-cancer CRISPR datasets and one longitudinal CRISPR screen. Chronos generally outperforms competitors in separation of controls and strength of biomarker associations, particularly when longitudinal data is available. Additionally, Chronos exhibits the lowest copy number and screen quality bias of evaluated methods. Chronos is available at https://github.com/broadinstitute/chronos .

Fig. 1

Overview of Chronos. a Illustration of the model for the simplified case where an sgRNA j is introduced targeting essential gene g in cell line c. Stochastic double-strand repair divides the population of infected cells into two groups, one with (top) and one without (bottom) successful gene knockout, which proliferate at different rates. Successful knockout probability is the product of a per-sgRNA probability p_j and a per-cell line probability p_c. The population of cells measured at each subsequent time point is a mix of these two populations. Chronos infers the relative change in growth rate r_cg. b Typical workflow of a Chronos run. Readcounts driven by outgrowing clones are removed, then gene fitness effects inferred using the readcount matrix, sequence map, and guide map. The inferred gene fitness effects are then corrected for copy number effects. c Number of cell lines in each lineage for the Achilles and Project Score datasets used for comparison

Fig. 2

Comparison of positive-negative control gene separation across methods for Achilles and Score datasets. a The distribution of all gene fitness effects for unexpressed (negative control) genes and common essential (positive control) genes. Unexpressed genes are identified individually for each cell line (Methods). b Global separation (pooled across cell lines and genes) between gene scores for common essential genes and unexpressed genes in the cell line where they are unexpressed. Separation computed using null-normalized median difference (NNMD). More negative values indicate stronger separation. Black arrows indicate the direction of improved performance. c NNMD for individual cell lines. Boxes indicate the interquartile range (IQR) of NNMDs. Whiskers extend to the last point falling within 1.5 x IQR of the box, and lines indicate medians. d Estimated false positive rate, based on the total percentage of unexpressed genes scoring in most depleted 15% of gene scores within cell lines. e Fraction of unexpressed genes scoring as false positives in individual cell lines. f Area under the precision/recall curve (PR AUC), where recall is the number of common essential gene scores that can be recovered at a given precision. g Fraction of possible common essential gene hits identified at 90% precision in individual cell lines

Fig. 3

Performance improvement with additional time points. The shaded area shows the 95% confidence interval for the (7 choose n) possible permutations of n measured time points that can be supplied to an algorithm. “Average” results are the performance achieved by taking a subset of the seven individual runs with a single late time point with a given algorithm and taking the median of their gene fitness effects. Joint results are obtained by running Chronos with a subset of multiple late time points simultaneously

Fig. 4

Performance on selective dependencies. a For each identified oncogene, the NNMD between cell lines with and without the canonical biomarker. b The results of a when aggregating results across oncogenes with signal. c PR AUC for separating cell lines with an indicated alteration from those without for individual oncogenes. d The results of c when aggregating across oncogenes. e The number of known expression addictions (y-axis) found to have a Pearson correlation lower than X (x-axis). f The area under the curves of e for the individual datasets

Fig. 5

The copy number effects. a Distribution of correlations for uncorrected gene log fold changes with their own copy number across cell lines in Achilles for common essential and nonessential genes. b Lowess smoothed trends for the mean-centered log fold change of known essential and known nonessential genes as a function of copy number. c Per-gene correlations of gene fitness effects with its own copy number, binned by mean gene fitness effect. BAGEL2’s copy correction is supplied by CRISPRCleanR. The boxes show the IQR for the correlations of genes in the given bin, whiskers extending to the last data point within 1.5 the IQR from the median

Fig. 6

Prevalence and cause of differences between Chronos and CERES estimates. a, b The Chronos/CERES gene effect profile correlation for genes with Chronos mean effect greater than -0.5 and standard deviation greater than 0.1 (y-axis) plotted against the mean correlation of the gene’s sgRNAs (x-axis). Genes for which CERES estimates a single sgRNA has at least 0.2 greater efficacy than the others are highlighted. c, d For the gene TCEAL7, the relationship between individual sgRNA log fold-changes (LFCs) and gene effect estimates by either Chronos or CERES. Points are cell lines. Lines show best-fit regression of each sgRNA to the algorithm’s gene effect estimate with shaded 90% confidence intervals. Sequence labels of sgRNAs are truncated to the first four nucleotides for clarity. e CERES vs Chronos gene effect scores for TCEAL7. Dot color indicates whether the gene was expressed or not in each cell line