Researchers used AlphaFold-derived contact predictions to choose Cas9 changes, then tested the engineered proteins and measured off-target activity falling from twenty-eight percent to five percent for one variant. The ContactSeek workflow compared predicted molecular contacts at intended and unintended edit sites. Researchers used the output to select amino-acid positions for modification.
Twenty-three substitutions across ten positions were tested experimentally. A selected variant retained similar intended activity while reducing measured off-target activity. The sequence is stated as reported because later official or investigative records may refine the first public account.
The method also worked with a Cas12-based system. The work was published in Nature. No human therapeutic outcome was tested. Attributed claims are preserved as claims rather than treated as independent proof.
Structural prediction can narrow an experimental search without proving biological safety. Gene-editing risk includes delivery, immune response and unintended genomic change beyond one assay. Reproducible laboratory validation creates an evidence layer independent from the model’s confidence.
The available record also connects a verified development with its operating context. The ContactSeek workflow compared predicted molecular contacts at intended and unintended edit sites. Structural prediction can narrow an experimental search without proving biological safety.
A separate factual check preserves the distinction between the event and the explanation offered for it. Researchers used the output to select amino-acid positions for modification. Gene-editing risk includes delivery, immune response and unintended genomic change beyond one assay.
The reporting boundary remains material: Generalization across guides, cells, organisms and clinical delivery remains unknown.
The next observable records are independent replication and broader mismatch testing and preclinical comparison with established high-fidelity editing systems.