6/ We also show how deep learning can predict the effects of non-coding edits. Here, an erythroid but not T-cell model predicts loss of DNA accessibility upon editing. The model learns the GATA motif driving this prediction. This site is of course the intended target of Casgevy.

5/ We extensively optimized beCasKAS to be compatible with primary human T cells and 5moU mRNA delivery. Transient mRNA delivery limits off-target formation, which can be further mitigated by carefully selecting mRNA dose.

4/ We are able to see the known strand-specific editing patterns of two contrasting editors: eBE (using hAPOBEC3A) and ABE8e. Using plasmid delivery in HEK293Ts, we find these two editors have surprisingly similar absolute editing frequencies.

3/ beCasKAS makes two measurements at each off-target site. The unwound R-loop appears as a peak, and individual edits appear as mismatched nucleotides.

2/ We use an N3-kethoxal pulldown probe because it has the same selectivity for ssDNA as the deaminases used in base editing. The cell permeable kethoxal allows us to enrich for unwound Cas9 R-loops that must occur before DNA editing. This selectivity underlies our sensitivity.

1/ Happy to share our preprint from the Greenleaf Lab: beCasKAS, our method to directly detect CRISPR base editor off-targets in primary cells. We additionally show how non-coding edits can be triaged for epigenetic dysregulation using deep learning.

www.biorxiv.org/content/10.1...