A new use for AlphaFold: making gene editing less error-prone

Off-target edits have long been the safety bottleneck for CRISPR-style gene editing: cuts that land on DNA sites only resembling the intended target. A Nature paper puts AlphaFold3, the protein-structure predictor trained to watch how molecules fold, to a new use: flagging the residues in Cas-family editor proteins that drive those wrong edits so the proteins can be redesigned.

Gene-editing systems combine a guide RNA that picks the target, a Cas protein (most famously Cas9) that cuts, and a third component that varies by editing mode. Cas9 enforces specificity by tolerating only a limited number of mismatches between the guide and the DNA, but it still tolerates some, which is where off-target edits slip in. Prior fixes have included picking guide RNAs that minimize similarity to other sites and engineering Cas variants with lower off-target tendency, each with limited room to spare.

The new work, re-reported by Ars Technica, modifies AlphaFold3's contact-prediction layer to surface the residues in Cas-family proteins that mediate off-target activity. The authors released the pipeline as ContactSeek on GitHub, and a Hyper.ai explainer calls the approach contact-modelling-driven base editing.

The paper reduces the rate of off-target edits, not the existence of the problem; delivery, immunogenicity, and editing-efficiency trade-offs are untouched. The next test is whether the contact-residue approach generalizes to other Cas-family members and editing modes.