Scientists have turned the protein‑folding AI AlphaFold into a diagnostic tool for improving the safety of CRISPR gene editing. The researchers fed the system models of a target DNA sequence, a guide RNA and the Cas9 nuclease—the core of the CRISPR system—to generate three‑dimensional structures that match those observed in laboratory experiments.
Early attempts that also included an enzyme designed to chemically modify bases caused AlphaFold to misplace one of the proteins, prompting the team to simplify the input. Stripping the model down to just DNA, RNA and Cas9 produced reliable predictions, confirming that the nuclease alone dictates sequence specificity.
Comparing the AI‑generated structures for on‑target and off‑target DNA sites revealed a striking pattern. Roughly two‑thirds of the off‑target sites forced Cas9 into a slightly altered overall shape, but more than 95 % changed which amino acids made contact with the guide RNA. In many cases, the protein retained its global conformation while individual residues flexed to accommodate mismatched bases.
AlphaFold’s built‑in “contact probability” metric—measuring the likelihood that two atoms lie within eight angstroms—allowed the team to quantify these shifts. By subtracting the contact map of an off‑target site from that of an on‑target site, they identified precisely which amino acids altered their interactions when a mismatch occurred. The researchers named this comparative workflow “ContactSeek.”
Targeting the hotspots
ContactSeek initially produced a long list of shifting residues, but the scientists focused on clusters where many changes converged, interpreting these hotspots as regions of the protein that adapt to mismatches. They then engineered Cas9 variants in which the identified amino acids were swapped for alternatives predicted to stabilize the correct RNA‑DNA pairing.
Preliminary testing of the redesigned enzymes showed reduced binding to off‑target sequences while preserving activity on intended sites. The approach demonstrates how AI‑driven structural insight can guide precise protein engineering, offering a path toward gene‑editing tools with fewer unintended edits.
The work highlights the expanding role of machine‑learning models in biotechnology, moving beyond prediction to active design. As CRISPR therapies advance toward clinical use, methods like ContactSeek could become essential for ensuring that genome‑editing interventions are both effective and safe.
This article was written with the assistance of AI.
News Factory APP - agentic news to boost your SEO & AEO.