Mitigation of off-target toxicity in CRISPR-Cas9 screens for essential non-coding elements

Pooled CRISPR-Cas9 screens are a powerful method for functionally characterizing regulatory elements in the non-coding genome, but off-target effects in these experiments have not been systematically evaluated. Here, we investigate Cas9, dCas9, and CRISPRi/a off-target activity in screens for essential regulatory elements. The sgRNAs with the largest effects in genome-scale screens for essential CTCF loop anchors in K562 cells were not single guide RNAs (sgRNAs) that disrupted gene expression near the on-target CTCF anchor. Rather, these sgRNAs had high off-target activity that, while only weakly correlated with absolute off-target site number, could be predicted by the recently developed GuideScan specificity score. Screens conducted in parallel with CRISPRi/a, which do not induce double-stranded DNA breaks, revealed that a distinct set of off-targets also cause strong confounding fitness effects with these epigenome-editing tools. Promisingly, filtering of CRISPRi libraries using GuideScan specificity scores removed these confounded sgRNAs and enabled identification of essential regulatory elements.


Supplementary Figure 4. Validation experiments for dense-tiling screen of enhancers of GATA1.
A. Individual sgRNAs generated on-target indels in K562 after lentiviral delivery and puromycin selection, as quantified by ICE analysis 8,9 . Dark blue dots correspond to sgRNAs that reduce fitness and have low GuideScan scores, and orange dots correspond to sgRNAs that do not reduce fitness and have high GuideScan scores.
B. Competitive growth assay validated expected growth effects in these individual cell lines.
C. Individually measured growth effects correlate with the pooled screen measurements. A. Four parallel screens were conducted tiling the loci of essential growth genes GATA1, MYB, and ZMYND8 using the four platforms Cas9, CRISPRa, CRISPRi and dCas9. Shown is the full tiled region around ZMYND8 with and without filtering for high-specificity sgRNAs with the GuideScan score.
B. Full tiled region around MYB.
C. Full tiled region around GATA1.
D. Clustering of sgRNAs from the GATA1 tiling screen that target regions with expected on-target effects (exons, TSS, and enhancers).

Supplementary Figure 6. Comparison of fitness effects and specificity scores with the number of off-target binding locations.
A. Comparison of GuideScan scores with fitness effects in the tiling screen, filtered to exclude sgRNAs that are likely to have on-target growth effects by removing sgRNAs 1000 bp upstream to 1000 bp downstream of ZMYND8 or MYB coding sequences, and 1000 bp upstream of eGATA1 to 1000 bp downstream of eHDAC6. For the similar plot that includes those sgRNAs, see Figure 3C. sgRNAs with multiple perfect matches to the genome or off-target locations with only 1 mismatch were excluded.
B. For the same set of sgRNAs in A, we compared the guide enrichment from the tiling screen with the number of off-target binding locations that have 2-3 mismatches. The off-target search was done with GuideScan.
C. For comparison, the relationship between the GuideScan specificity score and the number of off-target locations for the same sgRNAs in the tiling screen library. D. Tradeoff between removing confounded sgRNAs and retaining sgRNA library density. Confounded sgRNAs were defined as the set from A with fitness effects ≤-2. sgRNAs with any perfectly-matched or 1-mismatch off-target sites are considered to have GuideScan scores <0 for this analysis.

Supplementary Figure 7. Validation of CRISPRi repression of essential enhancers with highspecificity sgRNAs.
After delivery of individual sgRNA by lentivirus, followed by puromycin selection, we performed qPCR for GATA1 mRNA levels and a Western blot for GATA1 protein levels (shown in Figure 3). The knockdown measurements are correlated.

Supplementary Figure 8. Parallel screens of CTCF loop anchors with Cas9, CRISPRi/a, and dCas9.
A. The CTCF motif-targeting sgRNA library was used in parallel screens to compare the CRISPR-Cas9 platforms. All screens shown here were maintained at 3000x coverage (cells per sgRNA), whereas the Cas9 screens shown in Figure 1 were maintained at 11,000x coverage.
B. Growth effects measured in this screen were validated with individual competitive growth assays. Validation of Cas9 effects shown in Figure 1. Error bars are standard deviation of three technical replicates.
C. Reproducibility between biological replicates. For CRISPRi/a, sgRNAs ≤1000 bp from the TSS of an essential gene identified in a previous CRISPRi/a gene screen were excluded to avoid on-target artifacts.
D. Low-specificity guides are significantly enriched among CTCF motif-targeting guides with fitness effects when using CRISPRi/a. P-value from Fisher's exact test, using a 2x2 table of the numbers of guides in each quadrant based on the thresholds drawn in black lines. Numbers in corners correspond to the number of CTCF site-targeting guides in the quadrant. sgRNAs with > 1 perfect matches to the genome or > 0 off-target locations with only 1 mismatch were excluded from this analysis, as before. Notably, the Cas9 screen shown here was maintained at lower coverage and thus resulted in noisier data than the replicates shown in Figure 1. It showed a significant, but less pronounced, enrichment for low-specificity guides among the guides with fitness effects (Fisher's exact test) than in the higher quality screen data shown in Figure 1, showing that experimental noise can disguise the confounding effect of off-target activity.
E. Clustering of low-specificity sgRNAs reveals that each perturbation has off-target activity that reduces cell fitness with a unique subset of the low-specificity sgRNAs. Shown are the subset of low-specificity sgRNAs that have a guide enrichment ≤-2 in at least one replicate. A. We retrieved data from a published growth screen where sgRNAs were targeted to the TSS of known essential and nonessential genes 10 , in different cell types. The marked depletion of sgRNAs targeting non-essential genes was unexpected and the authors discussed the need for further investigations to clarify the source of these effects. Here, we found that these sgRNAs have low specificity scores, implicating off-target activity. However, the enrichment was not significant, possibly due to the small number of sgRNAs in the dataset. B. We retrieved data from a published growth screen where sgRNAs targeted the TSS of genes with Cas9 11 . We excluded sgRNAs with any off-target sites with only 0 or 1 mismatch as determined by the GuideScan search tool. There is a significant enrichment for fitness effects with low-specificity sgRNAs (Fisher's exact test).

Supplementary Figure 10. Filtered library designs for regulatory elements and splice sites.
A. ccREs were retrieved from the ENCODE SCREEN database and their distribution of lengths is shown.
B. Various GuideScan score filtering cutoffs were applied to the sets of sgRNAs overlapping the ccREs. 89% of ccREs can be targeted with ≥5 sgRNAs with GuideScan scores > 0.2, enabling CRISPRi/a screens of ccREs with high-specificity libraries.
C. Fraction of splice sites that can be targeted with sgRNAs within a window (-20 to +10 bp), after filtering out low-specificity sgRNAs.

Supplementary Data 1: sgRNA libraries used in this study
The sgRNA sequences and the sgRNA scores from GuideScan are provided in a separate Excel file.

Deep learning models for TF binding prediction used for CTCF library design:
To select CTCF sites in the category of loop anchors without annotated CTCF binding sites, we one-hot encoded as 4*1000 binary matrices following previously established practices 18,19 . "N" bases were encoded as zeros. We separately encoded each sequence and its reverse complement.
We used an architecture featuring three convolutional layers, with each followed by a rectified linear unit (ReLU), followed by a max-pooling layer. The convolutional filters of the first layer can be interpreted as picking up sequence patterns revealing whether a peak is a CTCF peak or a DNase peak without CTCF, the filters in the following layers identify combinations of those patterns, and the max-pooling layer encodes the assumption that a single sequence pattern combination should not occur multiple times within a short region. The first convolutional layer had sixty (4*15) filters with stride 1*1, the second convolutional layer had 60 (1*15) filters with To identify regions within ChIP-seq peaks that are important for making positive predictions, we scored the importance of every nucleotide in each positive example using DeepLIFT with gradient times input, which computes the product of each input and gradient with respect to that input (Shrikumar, Greenside, & Kundaje, 2017). Since most CTCF ChIP-seq peaks that were correctly predicted had at least one region with high DeepLIFT scores and we wanted to select less than one hundred guides, we filtered the regions with high DeepLIFT scores by cross-correlating the scores starting at each index within the sequence with the log-odds of the CTCF PWM from JASPAR 14 , where we used a pseudo-count of 0.0001 and a background of 52% GC content when computing the log-odds. This was procedure was carried out for each sequence and its reverse complement, and the top two motif hits across both were retained. (Note that some of these motif hits would not be identified by scanning the sequence for the CTCF motif because the regions with important DeepLIFT scores are not always those with the best matches to the CTCF motif.) Motif hits within 20 bp of a higher-scoring motif hit as well as those with log-odds scores ≤0.5 were removed. Motif hits in peaks without a previously identified CTCF motif hit were retained; the intuition is that these sequences are imperfect matches to the CTCF motif that are missed by PWM scanning.