A-rich PAM Cas9 editors prove more predictable: a single-nucleotide microhomology at -4 sets insertions versus deletions
Synopsis
The A-rich PAM-recognizing Cas9 editors SpyMac and iSpyMac produce a higher share of templated insertions and fewer random insertions than conventional G-rich PAM SpyCas9 across about 37,000 targeting sites. The work attributes the difference to a one-base microhomology between the nucleotide at -4 and the PAM, which converts insertions into deletions through microhomology-mediated end joining. The authors also built CRIfuser, a discrete diffusion model that predicts editing profiles after staggered cleavage.
Interpretation
A-rich PAM-recognizing Cas9 editors behave differently in precision editing: SpyMac and iSpyMac yield a higher proportion of templated insertions and a lower proportion of random insertions than conventional SpyCas9. SpyCas9 had been widely described as making blunt double-strand breaks 3 bp upstream of the PAM, with staggered cleavage tied only to short 5′ overhangs of 1–3 nucleotides. Across 37,266 targeting sites, 11.11%, 16.43% and 15.2% of sites exceeded 50% templated insertions for SpyCas9, SpyMac and iSpyMac respectively; templated events made up 31.04%, 53.01% and 46.64% of 4-bp insertion reads.
The sequence preference of insertions and deletions is explained by a one-base microhomology between the -4 nucleotide and the PAM: A at -4 pairs with NAA and G at -4 pairs with NGG, and microhomology-mediated end joining then produces deletions that suppress the corresponding insertions. The long-standing observation that G/C at -4 suppresses templated insertions in the SpyCas9 insertion motif had lacked a mechanistic explanation. In paired NAA and NGG libraries, deletions rose significantly for A-rich PAM editors when -4 was A and for SpyCas9 when -4 was G, while -4 C or T controls showed no significant change.
Loss of PolQ reshapes the indel spectrum: 2–10 bp random insertions fall significantly, 1-bp templated insertions rise significantly, and 1-bp random insertions stay unchanged. Polθ had mainly been linked to alternative end joining such as microhomology-mediated end joining. About 180 million sets of editing outcomes were analyzed across six single-cell clones lacking the entire coding region, the polymerase domain or the helicase domain, with high inter-clonal consistency in editing efficiency.
CRIfuser represents editing outcomes as a two-dimensional distribution of the two ligation-end positions and generates predicted profiles by iterative discrete denoising, achieving lower negative log-likelihood than inDelphi, FOREcasT and Lindel. Earlier prediction models did not consider the predictable templated insertions produced by staggered cleavage. It reached the lowest negative log-likelihood loss on the SpyCas9, SpyMac and iSpyMac datasets.
Perspective
The conclusions come from paired-library comparisons of SpyCas9, SpyMac and iSpyMac in HEK293T cells and from PolQ domain-deletion single-cell clones built in that background; CRIfuser's prediction range covers deletions up to 25 bp and templated insertions up to 6 bp around the cleavage site.
The authors attribute the drop in random insertions under PolQ deficiency to synthesis-dependent microhomology-mediated end joining and explain the rise in 1-bp templated insertions by pathway competition between cNHEJ and MMEJ; these mechanistic explanations remain inferences. The paper is a preprint that has not been peer reviewed.
