Base editors: attachment site beats linker length

In Short. A decade of base-editor engineering treated linker length as the dial on where the deaminase hits DNA; a new geometric model says changing that length barely moves the window—attachment site does.

---

Base editors promised a cleaner kind of CRISPR: change a letter without cutting both DNA strands. The practical question was always which letters. The set of protospacer positions an editor converts—the activity window—decides whether a wanted change arrives alone or with bystander neighbours. For a decade the engineering story sounded tidy. Lengthen the linker between Cas and deaminase and the window should broaden; shorten or stiffen it and the window should narrow. The theory was geometric folklore with laboratory clothes on.

Picture the working assembly. A Cas protein binds a guide and peels one DNA strand into an R-loop, exposing bases on the displaced strand. A cytidine or adenosine deaminase, fused somewhere on that Cas scaffold, has to reach those bases. Two design variables seemed obvious: where you bolt the enzyme on (the attachment site) and how long the flexible string between bolt and enzyme is (the linker’s contour length). Labs lengthened linkers, tried rigid ones, inserted deaminases into scaffold loops, replaced domains, and circularly permuted the protein. The reasoning, stated plainly in the early literature, was that physical distance and rigidity set window width.

So when Anees Ahmed Mahaboob Ali and Everette Jacob Remington Nelson of Vellore Institute of Technology put that proposition into a quantitative model, hopes for the old dial were still high (bioRxiv). Their geometric construction uses only linker statistics, steric exclusion, and R-loop geometry. Three parameters—capture radius, displaced-strand persistence, and link steepness—were fitted to two previously reported profiles and then held fixed. Predictions were scored against 50 architectures from seven studies, with each substrate coordinate withheld from the score. If tether geometry were the whole story, the model should relocate windows when the tether moves.

Practice had other ideas. Varying contour length by a factor of sixteen—XTEN linkers from 3 to 48 residues in their length series—did not shift the predicted window at all. The peak stayed at protospacer position 5 or 6; the centre of mass wandered by about 1.4 nucleotides with no ordering by length. Changing the attachment site did move things. Across 185 sterically buildable single-linker designs at thirteen attachment sites, three composition classes, and seven contour lengths, the predicted peak never left positions 3 to 12. Positions 1–2 and 13–20 were the mode of no design. On an intact scaffold, a single linker therefore cannot place the peak outside that band; across sampling seeds it sits within 5 to 11.

That bound is a ceiling on where a fused enzyme can act, not a forecast of where it does. The paper is careful on the distinction. Window position agrees with measurement to roughly one nucleotide on their corpus, yet does not beat a constant predictor that always guesses the crowded centre of the accessible band. Window width is not reproduced. Transfer to Cas12a fails by five to six nucleotides in a way that, after the authors exclude reach, capture radius, sterics, and substrate treatment, localises to the fusion junction rather than to how far the tether can stretch. Geometry sets a reachable set. Chemistry and junction detail still decide occupancy inside it.

The failure is a useful reminder that a decade of linker folklore was turning the wrong dial. Architectures that relocated windows substantially did so by inserting the deaminase into scaffold loops, replacing a domain, or permuting the protein—not by lengthening a string. The narrow practical takeaway is correspondingly narrow. If you are building a tethered base editor and care about window position, treat attachment site as the design variable and linker length as a secondary constraint. This preprint is a design-geometry claim about what tether models can and cannot promise. It is not a new clinical editor, and inventing editing rates for patients would miss the point.■