Benchmarking of SpCas9 variants enables deeper base editor screens of BRCA1 and BCL2

For each Cas9 variant we systematically investigated on-target activity: cutting efficiency at intended sequences; and off-target activity: cleavage at unintended sites. All variants were included in a preliminary PAM-mapping assay, except HiFi-Cas9. The strongest high-fidelity and PAM-flexible variants were further assessed for off-target promiscuity, at which point HiFi-Cas9 was added to the panel (Fig. 1a). We established A375 cells stably expressing these Cas9 variants, and confirmed comparable expression levels via flow cytometry (Supplementary Fig. 1a, b).

To quantify performance among SpCas9 proteins, we designed a PAM-mapping library that reports on both PAM preferences and on-target efficacy, based on a variant’s ability to distinguish between essential²⁷ and nonessential genes²⁸. The library contains 70–100 sgRNAs per four nucleotide PAM, including all 256 possible PAMs (the canonical SpCas9 PAM is NGGN). Each of these sets includes three sgRNA 5′-types that differ in length and presence of a matched 5′ guanine: G19, a 20mer with a matched guanine; G20, a 21mer with a matched prepended guanine; and g20, a 21mer with a mismatched prepended guanine (Fig. 1b). This library (18,768 guides) was screened in duplicate at >500x coverage for three weeks (Fig. 1c). At the end of the screen we collected cells, isolated genomic DNA, retrieved the library by PCR, and sequenced to determine guide abundance

We first calculated the log2-fold-change (LFC) relative to the plasmid DNA (pDNA) (Supplementary Data 1). All variants showed good reproducibility (Pearson correlation 0.54–0.87) except Cas9-VRER (0.25), which had few active guides (Fig. 1d), explaining the poor reproducibility (Supplementary Fig. 2a). We quantified the fraction of guides targeting essential genes that were more depleted than the 5th percentile of guides targeting nonessential genes and non-targeting controls for every PAM. For WT-Cas9, this confirmed the preference for an NGGN PAM, as 95.3% of guides were active by this lenient definition (Fig. 1e), with low but detectable activity at NAGN (18.6%), NGAN (6.1%) and NCGN (4.7%) (Supplementary Fig. 2a). We examined the data via an alternative metric, calculating average recall at 95% precision for each variant, designating guides targeting essential genes as true positives and those targeting nonessential genes as false positives. These metrics produce concordant results (Pearson’s r = 0.98 for WT-Cas9) (Supplementary Fig. 2b), with an average recall of 90.3% for an NGGN PAM, 18.7% at NAGN, 5.2% at NGAN, and 4.3% at NCGN PAMs.

As expected, the high-fidelity variants were only active at NGGN PAMs (Supplementary Fig. 2a), so we included only these PAMs in subsequent analyses. We calculated precision-recall curves, including guides of all 5′-types. At 95% precision, WT-Cas9 performed best (90% recall), followed by eCas9-1.1 (40%), HypaCas9 (27%), Cas9-HF1 (25%), and evoCas9 (4%) (Fig. 1f). When discretized by 5′-type, we observed a pronounced preference for G19 guides with all high-fidelity variants; this preference was marginal for WT-Cas9 (Fig. 1g). Considering only the G19 guides, WT-Cas9 again performed best (94%), followed by eSpCas9-1.1 (90%), Cas9-HF1 (76%), HypaCas9 (74%), and evoCas9 (35%), consistent with a recent study using reporter construct screens¹⁷. That an extra 5’G, whether paired or not, greatly diminishes activity with these variants is consistent with prior reports^19,20,21. Importantly, this constraint reduces the number of potential sgRNAs four-fold. We summarized PAM activity into active (guides with fraction active >0.7) and intermediate (0.3–0.7) bins for 21mers (G20/g20) and 20mers (G19) for these variants at all PAMs (Supplementary Fig. 2c).

Off-target activity of high-fidelity variants

We next compared the off-target tolerance of select high-fidelity variants to WT-Cas9. To systematically assess off-targets due to mismatches with the sgRNA, we collated a set of 21 sgRNAs that were active with every high-fidelity variant in the PAM-mapping assay (all G19 guides), and included all possible single (n = 1197) and double mismatches (n = 32,319) in the sgRNA sequence, as well as 1000 non-targeting controls, for a library of 34,537 guides (Fig. 2a). We performed screens in duplicate in A375 cells stably expressing WT-Cas9; eCas9-1.1, the best-performing variant in the PAM-mapping library; and HiFi Cas9, which we had not previously assessed¹¹. Replicates were well correlated (Pearson’s r = 0.90–0.93, Supplementary Data 2) and we determined the LFC of perfect match, single mismatch, double mismatch, or non-targeting control guides (Fig. 2b). To quantitate off-target activity, we calculated the ROC-AUC measuring the separation between perfectly matched guides (true positives) and mismatched guides (true negatives) (Fig. 2c). We then calculated the probability of being active for each mismatch type and position to generate a cutting frequency determination (CFD) matrix for each variant, as done previously with SpCas9 and AsCas12a^6,29 (Fig. 2d, Supplementary Table 2). We used a logistic regression model to transform LFCs to a probability of being active, defining perfect match sgRNAs as positive controls and non-targeting sgRNAs as negative controls.

We previously generated a CFD matrix for WT-Cas9 using guides mismatched to CD33⁶, and these new results were moderately consistent (Pearson’s r = 0.61, Supplementary Fig. 3a), despite several experimental differences including 5′-type (G19 only in the present study, no requirement previously), number of genes assayed (14 vs. 1) and readout (viability vs. flow cytometry). Here, we observed a higher tolerance for mismatches at the PAM-distal end of the guide with all three enzymes, as well as for rG:dT mismatches (Fig. 2d), two trends observed previously in SpCas9 with other techniques and which have also been seen with Cas12a enzymes^5,6,29. We also found that both high-fidelity variants show greater discrimination for rG:dA and rA:dA mismatches than WT-Cas9 (Fig. 2d, Supplementary Fig. 3b), whereas other mismatches, such as rC:dA and rU:dG, are still substrates for cleavage (Fig. 2d, Supplementary Fig. 3b). Using the product of the activities of each individual mismatch in the CFD matrix, we predicted the activity of double-mismatch guides, an approach that has been used to identify problematic off-target sites for SpCas9^29,30,31. To evaluate our predictions, we generated ROC curves (Fig. 2e) and saw good discrimination between the two sets (AUC = 0.83–0.86) for all three enzymes, validating this approach for guide design.

On-target activity of PAM-flexible variants

Returning to the PAM-mapping screens, we analyzed the activity of variants that recognize alternative PAMs (Supplementary Fig. 2a). For Cas9-VQR, we observed excellent activity with all 4 NGAG PAMs (fraction active > 0.9), and poor to intermediate activity at the remainder of the NGAN PAMs (0.11–0.67) (Fig. 3a, b, Supplementary Fig. 2a). The Cas9-VRER variant, characterized to target NGCG PAMs, showed intermediate activity with GGCG (0.36) and poor activity with HGCG (0.19–0.28).

For xCas9-3.7, two PAMs showed high activity (CGGC and TGGC), with 9 additional NGGN PAMs showing intermediate activity (fraction active 0.35–0.66), and the remaining 5 NGGN showing low activity (0.16–0.29). We identified 5 additional PAMs with intermediate activity (4 NGTN, 1 NGAN) for a total of 14 intermediate PAMs (Fig. 3a, b). Legut et al. recently characterized xCas9-3.7 at all possible 64 NNN PAMs¹⁶. We z-scored sgRNAs targeting the coding regions of CD45 and CD55 used in their assay and observed concordance with essential sgRNAs from the present study at all PAM sites (Pearson’s r = 0.79), with the majority of sgRNAs centered around 0, and activity only at NGG PAMs (Fig. 3c). Further, Kim et al.²³ performed a similar PAM classification study, measuring indel frequencies at 4 and 5 nucleotide PAMs. We compared our fraction active metric against their indel frequency using WT-Cas9 and observed good concordance (Pearson’s r = 0.95, Supplementary Fig. 4a). For xCas9-3.7, we observed a similar trend (Pearson’s r = 0.90), with the vast majority of PAMs centered around 0, and the strongest activity at NGGN sites, with modest activity at some NGHN sites (Supplementary Fig. 4b).

In contrast to the poor activity of xCas9-3.7, we identified 18 active and 43 intermediate PAMs with Cas9-NG¹⁴, including high activity at NGTG and NGAG PAMs, but diminished activity at NGAC and NGCC PAMs (Fig. 3a, b), consistent with prior results^14,16,23. We observed a similarly strong correlation between Cas9-NG in our assay and the results from Legut et al. (Pearson’s r = 0.78, Fig. 3c) and Kim et al. (Pearson’s r = 0.84, Supplementary Fig. 4c). Although xCas9-3.7 and Cas9-NG were both described as recognizing NG PAMs, they show little correlation (Pearson’s r = 0.24, Supplementary Fig. 4d); we note that Legut et al. observed more concordance (Pearson’s r = 0.72, Supplementary Fig. 4e).

Finally, we identified 24 active PAMs with SpG, all NGNN, consistent with initial characterization¹⁸ (Fig. 3a). 41 additional PAMs showed intermediate activity, 39 of which were NGNN and the remaining 2 NANN (Fig. 3b). We next compared Cas9-NG with SpG, and observed concordance across guides (Pearson’s r = 0.79 with all sgRNAs; r = 0.82 when filtered for NG PAMs) (Supplementary Fig. 4f, g), and PAMs (Pearson’s r = 0.9) (Fig. 3d). We found that some NGNN PAMs were more active with SpG than with Cas9-NG, while Cas9-NG had more activity at NANN PAMs (Fig. 3d, Supplementary Fig. 2a).

To understand if these variants had any 5′-type requirement, we calculated ROC-AUCs for each, using only active PAMs for each enzyme, and designating guides targeting essential genes as true positives and nonessential genes as true negatives (Fig. 3e). We found that none of these PAM-flexible variants demonstrated a marked preference, which is attractive for modalities like base editing.

Off-target profiles of Cas9-NG and SpG

To characterize the tolerance of Cas9-NG and SpG for guide-target mismatches, we selected 300 active, perfect-match sgRNAs from the PAM-mapping screens, maintaining a balance across different PAMs. We included all possible single mismatches (n = 17,775), a random subset of double mismatches (n = 60,000), and 1000 non-targeting controls, resulting in a library of 79,075 guides, including all three 5′-types (Fig. 4a). We screened this library in duplicate in A375 cells stably expressing Cas9-NG or SpG, and replicates were well-correlated (Pearson’s r = 0.81 Cas9-NG; r = 0.76 SpG; r = 0.72 Cas9-NG vs SpG, Supplementary Fig. 5a, Supplementary Data 3).

We examined LFCs of perfect match, single mismatch, double mismatch, or non-targeting control guides, considering every guide included in the library (Supplementary Fig. 5b). To ensure sensitivity to mismatched guides that maintain activity, we selected 149 of the perfect match guides with the highest effect size for subsequent analyses (Fig. 4b). We applied the same framework of assessing off-target activity by calculating the ROC-AUC, comparing mismatched guides to perfect matches (Fig. 4c). We observed separation between perfect matches and single mismatches (Cas9-NG AUC = 0.86; SpG = 0.85) and excellent differentiation between perfect matches and double mismatches (Cas9-NG = 0.97; SpG = 0.97).

We then calculated the probability of being active for each enzyme with each mismatch type and position using all 5′-types to generate a CFD matrix (Fig. 4d). We also calculated matrices separated by 5′-type and compared the probabilities of being active across guide types within each enzyme (Supplementary Fig. 5c, d). For both enzymes we found that g20 guides are the least prone to off-target cutting, followed by G20, and G19 guides. We compared the probability of being active for Cas9-NG and SpG by mismatch type using all 5′-types and observed excellent concordance (Pearson’s r = 0.97) (Fig. 4e). Cas9-NG and SpG have 7 and 6 total mutations, respectively, 4 of which are at the same residues, and one of which is the identical substitution (T1337R, Supplementary Fig. 5e). Thus, it is unsurprising that these variants behave so similarly.

Base editing with PAM-flexible variants

A major appeal of PAM-flexible variants is their potential for use in base editor screens, as the location of the perturbation is crucial for introducing the precise desired edit. While we have previously demonstrated the utility of C > T base editors (CBEs) in pooled screens, base editors capable of altering other nucleotides, such as A > G base editors (ABEs), would further expand the utility of such screens.

We benchmarked ABE7.10³² and the newer ABE8e²⁶ and ABE8.17³³ in a small-scale assay using a reporter construct containing EGFP and two sgRNAs targeting EGFP delivered via lentivirus to MELJUSO cells (Supplementary Fig. 5a). After sequencing the target site, we quantitated the nucleotide percentage at each editable A (A5 and A8 with EGFP sg1 and A4 and A9 with EGFP sg2) using EditR³⁴. We observed the most efficient editing with ABE8e (Supplementary Fig. 6b), so we selected it for further study. We next generated a Cas9-NG version of ABE8e and tested it in the same assay. Editing levels were lower compared to WT-Cas9, but still as high as 56% (Supplementary Fig. 6b).

Base editing of BRCA1

To understand the current scope of base editing screens, we used the DNA-damage repair gene BRCA1 as an example target. For WT-Cas9, we identified 455 unique residues (24.2% of the protein) in the longest BRCA1 isoform which could be targeted to introduce a missense or nonsense mutation (Fig. 5a). With Cas9-NG, 1342 targetable residues (71.2%) were identified considering the PAMs characterized above as active or intermediate. Likewise, with SpG, 75.3% of the protein can be modified with at least one mutation.

Using BE3.9max and ABE8e (Supplementary Fig. 6c), we designed two base editor versions of each PAM-flexible variant: NG-CBE, SpG-CBE, NG-ABE, and SpG-ABE. To test these 4 Cas-BE variant pairings in a screen, we designed a library tiling across BRCA1, containing all possible guides targeting the gene, irrespective of PAM (n = 11,524), including 30 guides targeting splice sites of essential genes²⁷, 75 intergenic-targeting guides, and 75 non-targeting guides. By including all PAMs we hoped to further reinforce the PAM preferences of each variant, while also future-proofing this library for use with emerging Cas9 variants, such as SpRY¹⁸. We screened these pairings in 2 cell lines, HAP1 and MELJUSO at high coverage (>2000 cells per sgRNA) for 21 days (Fig. 5b). Since BRCA1 is essential in near-haploid HAP1 cells³⁵, we conducted a negative selection (dropout) assay in this cell line. We treated MELJUSO cells with 1 µM of cisplatin²⁵ to enhance selective pressure for BRCA1 LOF alleles. Previously, we had screened BRCA1 with a tiling library containing only NGG PAMs using a WT-CBE²⁵; we re-screened this library with WT-ABE as well.

After calculating LFCs relative to pDNA, we found that replicates were well-correlated (Pearson’s r = 0.77–0.99 CBE screens, Supplementary Data 5; 0.77–0.95 ABE, Supplementary Data 6), and thus we averaged the data across the cell lines. We examined the distribution of positive (essential splice sites) and negative (non-targeting and intergenic) controls and found that negative controls were centered around 0, while positive controls were depleted (Supplementary Fig. 6d), confirming base editing activity. To understand our ability to assay BRCA1 itself, we examined the separation of guides predicted to introduce nonsense or splice mutations (positive controls) and silent or no edits (negative controls) and calculated the AUC for each base editor and cell line (Supplementary Fig. 6e, f). We observed the best performance with the WT base editors and slightly higher performance in HAP1, consistent with our original benchmarking²⁵. In every condition, we observed clear separation between control groups, confirming that we were able to assay the BRCA1 gene effectively. Next, we compared the NG and SpG base editors to WT, filtering on guides with NGGN PAMs, and observed good concordance (Fig. 5c, d), although some guides showed less activity than with WT-CBE. We speculate that this relates to the overall decreased activity with NG and SpG compared to WT at some NGGN PAMs, which may be especially important for CBE, as continued localization of the UGI domain is necessary for proper base editing. Finally, guides in the library behaved largely similarly when paired with ABE or CBE (Pearson’s r = 0.73 SpG; r = 0.71 NG, Supplementary Fig. 6g), with some clear outliers.

We examined guides introducing coding changes along the length of BRCA1 (Fig. 5e, f) and observed strong depletion of those targeting the RING and BRCT domains, consistent with our previous findings²⁵ and the clinical importance of these regions. However, it is inappropriate to draw conclusions about specific casual mutations from the behavior of sgRNAs solely from primary screening results, as sgRNAs might deplete due to out-of-window editing, unintended edits, indels, or off-target effects²⁵.

BRCA1 Validation

We selected 18 guides (sg1-18) for validation experiments based on the magnitude of z-scores in the primary screen, regardless of PAM activity. We cloned individual sgRNAs into either ABE or CBE vectors, transduced cells, and collected samples one (early time point), two, and three weeks post-transduction. We then PCR amplified the edited locus using custom primers and deep-sequenced the edited loci to identify the causal mutations (Supplementary Fig. 7a).

At the early time point, we observed a wide range of editing efficiency (C > T, 0.04–60.1%; A > G, 0.2–59.1%). In all cases with <1% editing, the sgRNA utilized an inactive PAM. Samples with an intermediate or active PAM averaged 41.7% C > T editing and 37.4% A > G editing in the predicted edit window of 4−8 nucleotides, with lower but detectable levels outside of the window (Supplementary Fig. 7b), consistent with previous observations^25,32,36. Next, we examined the reproducibility of percent change with WT alleles, comparing the percentage of reads in the late versus early samples (Supplementary Fig. 7c). We found that >10% enrichment of the WT allele was reproducible across replicates, and thus considered a guide to validate if the WT allele enriched >10% from early to late samples, indicating depletion of edited alleles. By this criterion, 4/7 guides with an active or intermediate PAM that depleted in the primary screen with CBE validated, and 2/5 guides validated with ABE.

For a number of guides, we conducted validation studies with both CBE and ABE. When screened with CBE, sg2 generates a P34F mutant, which depletes from 48.0% abundance on day 8 to 11.2% on day 21 (Fig. 6a, b). Although the P34F mutation has not been documented in the ClinVar database, several publications suggest that residue 34 plays a critical role. First, it is in the RING domain, which forms a heterodimer with the RING domain of BARD1 and is necessary for E3 ubiquitin ligase activity^37,38. Additionally, of the 6 possible missense mutations introduced at this residue via Saturation Genome Editing (SGE), 5 scored as LOF, and 1 as intermediate (P34F requires mutating 2 nucleotides, so was not included)³⁵. To gain structural insight into this LOF phenotype, we visualized these residues on the crystal structure of the RING domains of BRCA1 and BARD1 (PDB IJM7). Zn²⁺ atoms stabilize the structure within the RING finger and are maintained by two binding loops, Site I and Site II³⁸. P34 falls between Sites I and II on BRCA1, and if mutated to F34, comes into close proximity to C66 on BARD1, part of Site II on BARD1 (Fig. 6c). While further experimentation is required to understand the exact mechanism, it seems likely that the P34F substitution destabilizes the interaction between BRCA1 and BARD1. We also examined sg2 with ABE and observed an average of 54.3% editing at positions 4 and 5, resulting in an E33G mutant. This allele remained constant across timepoints analyzed (57.9% day 8; 54.6% day 21), indicating that, in contrast to the P34F mutation, the E33G mutation is not LOF (Fig. 6d, e). Indeed, when profiled by SGE, it was classified as intermediate³⁵. Further, this residue does not come in close contact with the Zn²⁺ atoms or their binding loops (Fig. 6f).

We also validated sg10 with both base editors. With CBE, we observed the predicted H1767Y mutation, as well as a second Q1768X mutation caused by out-of-window editing (26.7% C > T editing at C9). Alleles with the single H1767Y mutation depleted from an average of 35.9% on day 8 to 20.8% on day 21 (Fig. 6g) while the WT allele enriched from 35.2 to 64.4%. Alleles containing only the H1767Y mutation depleted, indicating this mutation is likely sufficient for LOF (Supplementary Fig. 7d). These results are concordant with the SGE data, as both H1767Y and Q1768X individually score as LOF³⁵. With ABE, sg10 introduces either an N1766S mutation (47.2% A > G conversion, A4) or N1766G mutation (18.7% A > G conversion, A3; 47.2% A > G conversion, A4) and an H1767R mutation (59.1% A > G conversion, A7), resulting in a LOF phenotype, while the WT allele enriches from 42.1% to 58.7% (Fig. 6g, Supplementary Fig. 7e). Given that the H1767R mutation occurred alone and did not deplete substantially (10.5% day 8; 9.8% day 21), it is likely that the mutants at position 1766 cause the LOF phenotype. Notably, Findlay et al. classified H1767R as functional, which is concordant with our observation; however, they also found that every missense mutation introduced at N1766 is functional, including N1766S³⁵. We did not capture any alleles with a mutation only at this position, so cannot make definitive conclusions about the role of N1766S or N1766G. It remains possible that the observed LOF arises from a combinatorial effect of both mutations.

Additionally, screens with sg9 and CBE introduced a mutation in the BRCT phosphopeptide binding motif (G1727K), a conserved motif in several DNA damage repair proteins, that allows association with proteins phosphorylated by ATM (Supplementary Fig. 7f, g)³⁷. Although this mutation has not been documented in ClinVar, G1727R and G1727E mutations are pathogenic, and G1727V is categorized as LOF by SGE³⁵, indicating that substitutions are not easily tolerated at this position. We also screened sg15 with both base editors, introducing a C64Y mutation with CBE, and C64R and L63P mutations with ABE (Supplementary Fig. 7h–k). This guide did not validate with NG-CBE, but did previously with WT-CBE²⁵ and NG-ABE. While we were unable to parse the effects of these individual mutants based on the spectrum of alleles in our data, C64R scored as LOF with SGE and L63P is pathogenic in ClinVar, so it is likely that both of these mutations contribute to the LOF phenotype.

All validation results are summarized in Fig. 6h. We identified 5 guides, which utilized intermediate or active PAMs, that introduced deleterious mutations with one or both base editors; 2 mutate the RING domain, and 3 the BRCT domain. 3 of 8 guides with intermediate or active PAMs introduced benign edits with both base editors, indicative of false positives in the primary screen. This aligns with our observations from previous WT-CBE BRCA1 and BRCA2 screen validations where 5/13 guides represented false positives from the primary screen²⁵. None of the 10 guides that utilized an inactive PAM validated, and sequencing showed little editing at these sites, reinforcing the PAM-specificity of these Cas9 variants and highlighting the necessity of validating primary screening results to avoid drawing conclusions from off-target effects.

Base editing of BCL2

Since its FDA approval in 2016, Venetoclax, which targets the anti-apoptotic protein BCL2, has been administered to patients with chronic lymphocytic leukemia (CLL), small lymphocytic lymphoma (SLL), or acute myeloid leukemia (AML)³⁹. Unfortunately, many develop resistance to treatment. Many of these tumors have single amino acid mutations in BCL2, and several other mutations that lead to drug resistance have been characterized in human cells or mice^40,41,42. We set out to identify additional resistance-causing mutations in BCL2, as a better understanding of these mechanisms can improve patient monitoring, allow for tailored treatment plans, and inform the design of new, mutation-agnostic drugs.

We designed a tiling library targeting BCL2 with Cas9-NG, generated both CBE and ABE versions, and screened in triplicate at high coverage (>10,000 cells per sgRNA) in MOLM13, an AML line sensitive to BCL2 inhibition. We treated cells with 62.5 nM Venetoclax for 14 days (Fig. 7a). LFCs for untreated cells were calculated relative to pDNA, and LFCs for Venetoclax-treated cells were calculated relative to untreated arms (Supplementary Data 7). We calculated z-scores for each guide relative to intergenic controls. 16 guides enriched with a z-score >3 with either or both base editors (Supplementary Fig. 8a), and 68.8% (11/16) of these are predicted to edit between positions 100–175, a region containing the P2 and P4 pockets responsible for Venetoclax binding (Fig. 7b). Notably, several sites of resistance mutations observed clinically (G101, D103, F104) fall within this region^40,41,42.

BCL2 Validation

We chose five sgRNAs to validate with both base editors, including three guides predicted to make missense edits at residues 103-105 (sg19–21), and two guides predicted to make missense edits at residues 148/149 (sg22) and 169 (sg23), which are denoted on the 3D structure of BCL2 in complex with Venetoclax (Fig. 7c). We performed validation screens as described above, with each of the guides individually transduced into MOLM13 cells in duplicate. The conditions from the primary BCL2 tiling screens were replicated, and following isolation and amplification of genomic DNA, we deep-sequenced the targeted loci.

Editing levels at the early time point ranged from 11.7–55.4% with ABE and 10.9%–47.9% with CBE in the predicted window (Supplementary Fig. 8b). While all predicted edits were based on the canonical 4–8 window, we observed editing in the 3–10 window with both ABE and CBE, as well as low levels of C > T editing farther afield (sg22) which led to several unpredicted amino acid substitutions (Supplementary Fig. 8b). As before, we used the relative abundances of the WT allele at the early and late time points to evaluate whether a guide validated. If the WT allele depleted by more than 10% under the selective pressure of Venetoclax, we considered it to be validated (Supplementary Fig. 8c). 8 of 8 guide-BE combinations validated as true positives, and 2 of 2 validated as true negatives. Further, sg20, 21, and 22 validated with both ABE and CBE, whereas sg19 and sg23 validated with the base editor with which they scored in the primary screens. With the non-scoring base editor, these sgRNAs were predicted to make either a silent edit (sg23, CBE), or no edit (sg19, ABE). That top hits from the BCL2 screens had a higher validation rate than those from the BRCA1 screens is likely because the former is a positive selection screen, which presents fewer opportunities for off-target activity to score.

We next examined the enrichment of specific alleles to determine causal resistance mutations. With sg20 we saw strong enrichment for the D103E mutation caused by a C > A transversion at position 5 (Supplementary Fig. 8d). In this case, all missense mutations were the result of out-of-window edits or transversions, highlighting the necessity for direct sequencing of the edited locus. When paired with ABE, sg20 edited at A4, resulting in enrichment of the D103G allele, which likely disrupts the α2 helix (Fig. 7d–f). When we examined sg21 with CBE, we also saw enrichment for missense mutations at D103; substitution for an asparagine (D103N) was most favored upon Venetoclax treatment, but we also observed enrichment of the D103Y allele, as well as the dual replacement of D102 and D103 with asparagine (Supplementary Fig. 8e). The D103 residue falls within the P4 pocket and is important for hydrogen binding between the azaindole moiety of Venetoclax and BCL2⁴¹. Both D103E and D103Y have been previously recorded in patient samples bearing the G101V mutation^41,43, and Blombery and colleagues have shown that D103E mutagenesis causes the P4-binding pocket to more closely resemble that of BCL-xL, which is not inhibited by Venetoclax. With ABE, sg21 predominantly enriched for the F104L mutation that increases the P2-binding pocket volume⁴⁴ and likely disrupts a hydrogen bond between Venetoclax and the side chain of F104 (Fig. 7g, h, Supplementary Fig. 8f). For sg19 we saw C > T editing at positions 5, 7, 8, and 9, introducing an S105F missense mutation in 94.5% of edited alleles at the early time point (Supplementary Fig. 8g). In all cases this mutant enriched during treatment with Venetoclax, and dual editing of S105F and R106C enriched further still (Fig. 7i, j). Interestingly, the strongest LFC was seen with a rare in-frame deletion that removes R106.

The preceding three sgRNAs introduced edits at positions 102–106, which are located in the α2 helix. Edits at the α5 helix and the non-core α6 helix also enriched in the primary screen. With sg22 we observed several resistance mutations at positions 148–152. With CBE, an A149T mutation comprised 94.2% of all edited alleles on day 7 (Supplementary Fig. 8h), which alone was able to confer resistance, but when V148I occurred in combination, we saw further enrichment (Fig. 7k, l). When sg22 was screened with ABE, we saw A > G editing at positions 3, 4, 9 and 10, leading to a V148A substitution in all edited alleles. This edit alone was sufficient to cause Venetoclax resistance, and we observed secondary edits which also enriched during drug treatment (Supplementary Fig. 8i).

The final validated sgRNA (sg23) was predicted to make a silent edit with CBE, though we observed low levels of editing in a large window (C0-C18), resulting in non-resistant missense edits. With ABE all edited alleles carried the L169P missense mutation (Supplementary Fig. 8j). Interestingly, this edit only enriched when V170A mutagenesis was observed in tandem (Fig. 7m). This resistance mechanism is particularly interesting, because residues 169 and 170 do not come in direct contact with Venetoclax. Mapping of these mutations onto the crystal structure of BCL2 shows the potential of a larger structural impact, whereby substitution with a significantly smaller side chain on the inner face of the helix, or disrupting the helix altogether with a proline substitution, may create vacated space that may cause additional conformational changes (Fig. 7n). A summary of the performance of sgs19−23 in both screens is provided (Supplementary Fig. 8k).

By leveraging PAM-flexible Cas9-NG paired with ABE and CBEs, we were able to densely tile BCL2 and identify nucleotide substitutions that confer resistance to Venetoclax, including three previously-documented mutations (F104L, D103E, D103Y), and several resistant mutations that, to our knowledge, have not been reported. This screen demonstrates the power of tiling base editing screens in a positive selection setting, and identifies a condensed region of BCL2 (100−175) harboring many resistance mutations, which may be particularly interesting for follow up with more exhaustive forms of mutagenesis.