Optimization of AsCas12a for combinatorial genetic screens in human cells

Abstract

Cas12a RNA-guided endonucleases are promising tools for multiplexed genetic perturbations because they can process multiple guide RNAs expressed as a single transcript, and subsequently cleave target DNA. However, their widespread adoption has lagged behind Cas9-based strategies due to low activity and the lack of a well-validated pooled screening toolkit. In the present study, we describe the optimization of enhanced Cas12a from Acidaminococcus (enAsCas12a) for pooled, combinatorial genetic screens in human cells. By assaying the activity of thousands of guides, we refine on-target design rules and develop a comprehensive set of off-target rules to predict and exclude promiscuous guides. We also identify 38 direct repeat variants that can substitute for the wild-type sequence. We validate our optimized AsCas12a toolkit by screening for synthetic lethalities in OVCAR8 and A375 cancer cells, discovering an interaction between MARCH5 and WSB2. Finally, we show that enAsCas12a delivers similar performance to Cas9 in genome-wide dropout screens but at greatly reduced library size, which will facilitate screens in challenging models.

Figure 1

Optimization of AsCas12a for pooled screens (a) Comparison of DNA cassettes necessary for dual knockout with Cas9 versus Cas12a. (b) Vector maps for Cas12a constructs. Point mutations for enCas12a are indicated. (c) Timeline for executing on-target library tiling screens. (d) ROC curves for guides targeting essential and cell surface genes for Cas12a and SpCas9, using viability data in A375 cells (n=1,146 and 2,468 for essential guides for Cas12a and SpCas9, respectively, and 153 and 673 for control guides). The area under the curve for each enzyme is noted in parentheses. (e) Same as (d) with 2 additional cell lines, HT29 and MELJUSO.

Figure 2

On-target design rules for enCas12a (a) ROC-AUCs for Cas12a tiling screens improve when filtered for the top half of guides by Seq-DeepCpf1 score. ROC-AUCs were calculated using guides targeting essential genes as true positives (n=1148 for all, n=573 for top half) and cell surface genes as true negatives (n=153 for all, n=76 for top half). (b) Seq-DeepCpf1 improves the fraction of active guides for 2xNLS-Cas12a. Guide scores are binned into deciles.. (c) The fraction of active guides increases along predefined PAM tier classifications for screens with the enCas12a PAM tiling library. Points represent the cumulative activity for a given tier, lines represent the fraction active for a given PAM. A dashed line is drawn at 5% activity. (d) Pipeline for training and testing machine learning models with the enCas12a tiling data. (e) Models trained on the PAM tiling data outperform models trained on the Seq-DeepCpf1 indel frequency data. Bars represent the mean spearman correlation on n=9 hold out genes from the PAM tiling data and line ranges represent the standard deviation. CNN denotes Convolutional Neural Network and GB denotes Gradient Boosted regression. (f) enPAM+GB improves the fraction of active guides for enCas12a. Guide scores are binned into deciles.

Figure 3

Prediction of off-target activity for AsCas12a (a) Schematic depicting off-target library construction and guide selection. (b) Density plots showing activity of guides in dropout screens targeting essential genes with zero, one, and two mismatches. Line is displayed at the 5th percentile of guides targeting control genes(c) Heat map displaying the fraction of guides active for each mismatch type at a given position in the guide. Guide position is numbered such that position 1 is PAM proximal. The fraction of active guides is reported from essential guides in dropout assays, vemurafenib resistance genes in vemurafenib assays, and HPRT1 guides in 6-thioguanine assays. (d) Comparison of off-target activity for Cas9, 2xNLS-Cas12a, and enCas12a. For each enzyme, the fraction of active guides at each guide position and nucleotide mismatch were ranked, and plotted in ascending order. (e) Density plots displaying measured activity of double mismatch guides targeting essential genes binned by a prediction of activity using the Cutting Frequency Determination (CFD) score. CFD activity bin and number of guides in each bin is reported.

Figure 4

Development of alternate direct repeat sequences for multiplexing with AsCas12a (a) Schematic of experimental design. (b) Average log2-fold change for direct repeats with both orientations of the BCL2L1 and MCL1 guides (n=35,883). Pearson correlation coefficient is indicated. (c) Most active variant direct repeat sequences by average log2-fold change across replicates and orientation. Nucleotide substitutions for top variants are shown below the wildtype sequence (top line). (d) Wildtype direct repeat sequence (left) and a consensus sequence for active variant direct repeats (right). Nucleotides in red denote positions that are flexible for alternate sequences. (e) Schematic of 6 multiplexed arrays with guides targeting CD47, B2M, and CD63 used in the triple knockout experiment. (f) Fraction of cells with no, one, two, or three genes knocked out, assayed by flow cytometry. (g) Comparison of knockout with arrays 2 and 6 when guides are separated by variant, or all wild-type direct repeats.

Figure 5

Validation of AsCas12a performance with synthetic lethal gene pairs (a) Schematic of library design. Numbered direct repeats reference the same sequences as Figure 4. (b) Correlation between the average log2-fold change (LFC) of target guides in position 1 versus position 2 for all three DR variants screened in OVCAR8. Pearson correlation coefficient is indicated. n=8,533 constructs for each scatter. (c) Average LFC for guide pairs versus the sum of each guide paired with controls in OVCAR8. Control points represent guide pairs with one control guide and one target guide. Regression line fit with control points only. Dashed line represents a residual two standard deviations below the mean residual for control points. (d) Density of residuals for synthetic lethal guide pairs in OVCAR8 screening with enCas12a, filtered for the top half of guides for each gene based on the enPAM+GB score. Dashed line represents two standard deviations below the mean residual of controls. Percent of pairs with a residual to the left of the dashed line is included. Labeled on the right is the number of unique constructs in the distribution. (e) Comparison of residuals for BCL2L1/MCL1 in OVCAR8 across Cas platforms. Control constructs have one target guide (BCL2L1 or MCL1) and one control guide (n=180), whereas target constructs contain a synthetic lethal guide pair (n=18). P-values were calculated using a one-sided t-test with the alternative hypothesis that the mean of the target population was less than the mean of controls. Boxes represent the 25th, 50th and 75th percentiles, whiskers show 10th and 90th percentiles.

Figure 6

Combinatorial screen identifies a novel synthetic lethality in apoptotic genes (a) Schematic of library design. Numbered direct repeats reference the same sequences as in Figure 4. (b) Example linear fit with an MCL1 guide in A375 cells. Each dot represents a guide paired with the MCL1 guide that anchors the analysis (n=322). Shaded grey around the linear fit represents the standard error. (c) Gene pairs plotted by residual Z-score in OVCAR8 versus A375 cells. Negative scores represent synthetic lethal genes, and positive scores represent buffering genes. Pearson correlation coefficient is indicated. Select gene pairs are labeled. (d) Gene interaction network. Edges are drawn at an absolute Z-score of 5. Negative edges represent synthetic lethal genes, and positive edges represent buffering genes. Nodes are laid out using the stress-minimization algorithm, Kamada-Kawai. (e) Box plot visualization of MARCH5 - WSB2 synthetic lethality. Boxes represent all guide pairs in a type (n=380 ctl:ctl, 400 MARCH5:ctl, 400 WSB2:ctl, and 200 MARCH5:WSB2), where ‘ctl’ denotes guides targeting olfactory receptor controls. Boxes represent the 25th, 50th and 75th percentiles, whiskers show 10th and 90th percentiles.

Figure 7

Genome-wide libraries for enCas12a (a) Schematic of Humagne Set A and Set B. Each library contains one multiplexed array with two guides targeting the same gene. (b) Timeline by which genome-wide screens were executed. (c) Guide-level precision recall curves for Humagne Set A, Humagne Set B, and Brunello in A375 cells. (d) Gene-level correlation of Humagne Set A + Set B versus Brunello (n=18,952 genes). (e) Comparison of the GSEA normalized enrichment scores (NES) for KEGG genes sets for Humagne Set A + Set B versus Brunello (n=164 gene sets). Pearson correlation coefficient is indicated in (d) and (e). (f) Guide-level recall for selected Cas12a and Cas9 libraries. Points in red are A375, MELJUSO, and HT29 screens described in this work. Each dot represents a cell line. Due to sample size, GeCKOv2 (n=33) and Avana (n=340) data from the DepMap are presented as box and whisker plots. Boxes represent the 25th, 50th and 75th percentiles, whiskers show 10th and 90th percentiles. (g) Recall of essential genes at 95% precision for various combinations of libraries and experimental replicates. Relative scale is calculated with reference to Humagne Set A + B screened with one replicate each.