A gene-editing safety problem that has mostly been studied by cataloguing where CRISPR-Cas9 cuts DNA it shouldn’t has a new structural answer for why it happens. A team based primarily at Peking University used DeepMind’s AlphaFold3 to model, residue by residue, how Cas9 and a companion base-editing enzyme physically interact with on-target versus off-target DNA and RNA sequences — then used that structural map to redesign the proteins, cutting measured off-target activity from 28 percent down to 5 percent in their tests.
What the study found
The work, published in Nature in July 2026 as “Precise DNA base editing using AlphaFold3-based contact modelling” (DOI: 10.1038/s41586-026-10794-z), comes from researchers at Peking University’s State Key Laboratory of Gene Function and Modulation Research, the Peking-Tsinghua Center for Life Sciences, and East China Normal University, led by Haowei Meng, Zhixin Lei, and Yongchang Yan among others. The team built a framework they call ContactSeek, which uses AlphaFold3’s contact-probability outputs — its predictions of which amino acid residues physically touch which DNA or RNA bases — to compare Cas9’s predicted structure when bound to its intended target against its predicted structure when bound to known off-target sites.
That comparison surfaced a clear structural signature. Roughly two-thirds of the off-target binding sites the team examined caused Cas9 to adopt a measurably different structural conformation than it takes at the correct target. More strikingly, over 95 percent of off-target sites altered which amino acid residues were in contact with the guide RNA at all — meaning most off-target binding isn’t just a weaker version of on-target binding, it is a structurally distinct interaction. ContactSeek flags these “consensus contact regions” — positions where the residue-RNA contact map consistently shifts between on-target and off-target complexes — as the most promising targets for mutation.
Working from that map, the team tested 23 amino acid substitutions across 10 key positions in Cas9 and in TadA8e, the deaminase enzyme used in adenine base editors built on a Cas9 scaffold. Their best combined variant — carrying mutations in both proteins — outperformed existing high-fidelity adenine base editors on specificity in their assays, and the same contact-modelling approach generalized to a Cas12a-based cytosine base editor, suggesting it isn’t a one-off fix specific to a single enzyme.
Cataloguing off-target sites versus explaining them
Most prior work on Cas9 specificity has approached off-target editing empirically: sequence a genome after editing, find where unintended cuts landed, and try to predict future off-target sites from sequence similarity to the guide RNA. That approach works well enough to be useful — an 18-base guide sequence should, by chance alone, appear roughly once every 70 billion bases, far less often than the human genome’s roughly 3 billion bases would predict a random collision — but it treats the enzyme itself as a black box and says little about the physical mechanism that lets Cas9 tolerate a partial sequence mismatch in the first place.
ContactSeek’s contribution is mechanistic rather than statistical: it asks what changes inside the protein-RNA-DNA complex, not just where the mismatch sits in the sequence. That distinction matters for engineering. A prediction method built on sequence similarity can flag a risky site but offers no direct route to fixing the enzyme; a structural map of exactly which contacts destabilize on off-target binding points directly at which residues to mutate. It’s the difference between a symptom list and a diagnosis.
An open-science-infrastructure story as much as a biology one
The result is also a demonstration of what happens when a general-purpose structural-prediction tool gets repurposed as diagnostic infrastructure. AlphaFold3, developed by DeepMind and the Isomorphic Labs team and made available for non-commercial research use, was built to predict static protein structures and complexes — not to compare on-target and off-target binding conformations for a genome-editing enzyme. ContactSeek is an example of a research group building purpose-specific analysis on top of that shared prediction infrastructure rather than every lab needing to develop its own structural-modelling capability from scratch. That reuse pattern — a widely available AI structural-prediction tool becoming the substrate for a specific engineering advance months or years after its original release — is becoming a recurring shape of AI-enabled biology research, and it raises the same open-science-infrastructure questions CASRAI tracks elsewhere: reproducibility of AI-assisted findings, transparent methods reporting, and equitable access to the compute and models that increasingly gate this kind of work.
Why the mechanism matters for gene-editing biosafety
Off-target editing is the central safety concern standing between CRISPR-based therapies and broader clinical deployment, and it is also a recurring theme in dual-use and research-security discussions about gene-editing tools generally — see CASRAI’s dictionary entry on Dual-Use Research of Concern (DURC) for how that oversight framework applies to research with legitimate benefit and misuse potential. A structural, mechanistic account of why off-target binding happens is more useful to that oversight conversation than a purely empirical catalogue of observed off-target sites, because it supports engineering fixes that can be reasoned about and verified independently, rather than tuned against one dataset and left opaque. It also sits squarely inside the broader push toward mechanistically understood, reproducible research tools that CASRAI’s Reproducible AI experiment entry addresses: an AI-assisted structural claim about a widely used research reagent is far more useful to the field if the contact-probability pipeline behind it is documented well enough for another lab to rerun.
None of this means off-target risk in CRISPR-based gene editing is solved. ContactSeek’s redesigned variants were validated in the reported assays, not across the full range of therapeutic delivery contexts, and generalizing structural fixes from one Cas9 variant to the many engineered variants now in clinical and research use will take further work. But shifting the field’s default question from “where did this guide RNA cut unintentionally” to “which structural contact made that cut possible” is a meaningful change in how gene-editing tools get engineered for safety — one that leans directly on open, reusable AI structural-prediction infrastructure rather than proprietary, single-lab modelling pipelines.
For background on the broader oversight landscape around CRISPR applications, see CASRAI’s guides on CRISPR Ethics: Germline Editing, He Jiankui, and Oversight and CRISPR Germline Editing Legal Status by Country, and the dictionary entry on the AI Research Tool category that AlphaFold3 and similar structural-prediction systems belong to.
