{"slug": "alphafold3-guided-base-editor-cut-one-off-target-rate-from-28-to-5", "title": "AlphaFold3-Guided Base Editor Cut One Off-Target Rate From 28% to 5%", "summary": "Researchers at Peking University and East China Normal University used AlphaFold3 to redesign CRISPR base editors, reducing off-target activity from 28% to 5% in one comparison while preserving on-target activity. The study, published July 22 in Nature, introduces ContactSeek, a framework that uses AlphaFold3 contact-probability predictions to identify amino acid substitutions that reduce off-target editing. The team tested 23 substitutions across 10 Cas9 positions and found a variant that lowered off-target activity in one reported comparison from 28% to 5%.", "body_md": "# AlphaFold3-Guided Base Editor Cut One Off-Target Rate From 28% to 5%\n\n- ContactSeek combines AlphaFold3 contact-probability predictions with experimentally measured off-target sequences to identify amino acids that let base editors tolerate mismatched DNA.\n[[1]](https://www.nature.com/articles/s41586-026-10794-z) - After testing 23 substitutions across 10 Cas9 positions, the researchers found a variant that lowered off-target activity in one reported comparison from 28% to 5% while retaining similar on-target activity.\n[[2]](https://arstechnica.com/science/2026/07/team-uses-alphafold-ai-to-redesign-gene-editing-proteins-to-make-them-safer/) - The team combined Cas9 and deaminase mutations in an editor called ABE8e-DD, which showed lower detected DNA and RNA off-target editing than three existing high-fidelity editors.\n[[1]](https://www.nature.com/articles/s41586-026-10794-z)[[3]](https://news.pku.edu.cn/jxky/a3a8cf38a5364d87a2539cde5bb95303.htm) - The study did not report animal dosing or a delivery experiment. Large SpCas9-based editors also exceed the payload capacity of a standard single AAV vector.\n[[1]](https://www.nature.com/articles/s41586-026-10794-z)[[4]](https://pmc.ncbi.nlm.nih.gov/articles/PMC9258002/)\n\nResearchers led by teams at Peking University and East China Normal University used AlphaFold3 to redesign CRISPR base editors, reducing detected off-target edits while largely preserving activity at intended DNA sites. The study, published July 22 in Nature, describes a computational framework called ContactSeek that connects predicted molecular contacts with genome-wide measurements of editing errors.[[1]](https://www.nature.com/articles/s41586-026-10794-z)\n\nThe most concrete result came from a Cas9 variant whose off-target activity fell from 28% to 5% in one reported comparison while maintaining similar activity at the intended site. The finding shows that AlphaFold3 outputs can inform experimentally useful protein engineering. It does not establish that the resulting editor is safe or ready for therapeutic use.[[1]](https://www.nature.com/articles/s41586-026-10794-z)[[2]](https://arstechnica.com/science/2026/07/team-uses-alphafold-ai-to-redesign-gene-editing-proteins-to-make-them-safer/)\n\n## ContactSeek looks beyond predicted structures\n\nBase editors pair a CRISPR targeting protein, such as Cas9, with a deaminase that chemically changes a DNA base without making the double-strand break used in conventional CRISPR cutting. A guide RNA directs the editor, but sequences elsewhere in the genome can resemble the intended target closely enough to attract unwanted activity.\n\nThe researchers first mapped off-target sites for Cas9–TadA adenine base editors and then supplied AlphaFold3 with combinations of Cas9, guide RNA and on- or off-target DNA. Predicted three-dimensional structures often did not display the subtle consequences of mismatched DNA bases. AlphaFold3’s contact-probability matrices, which estimate whether particular protein and nucleic-acid residues are near one another, were more sensitive to those differences.[[1]](https://www.nature.com/articles/s41586-026-10794-z)\n\nAbout two-thirds of the analyzed off-target sites produced a changed Cas9 conformation, while more than 95% altered which amino acids contacted the RNA, according to Ars Technica’s account of the paper. ContactSeek ranked clusters of altered contacts to identify residues that might help Cas9 accommodate mismatches.[[2]](https://arstechnica.com/science/2026/07/team-uses-alphafold-ai-to-redesign-gene-editing-proteins-to-make-them-safer/)\n\n## Sequencing supported lower DNA and RNA error rates\n\nThe team tested 23 amino-acid substitutions at 10 Cas9 positions identified through the workflow. One variant produced the reported reduction from 28% to 5%, and the researchers found improvements with additional guide RNAs. They also applied the approach to a Cas12a-based cytosine editor, suggesting that the method is not limited to one Cas protein.[[1]](https://www.nature.com/articles/s41586-026-10794-z)[[2]](https://arstechnica.com/science/2026/07/team-uses-alphafold-ai-to-redesign-gene-editing-proteins-to-make-them-safer/)\n\nThe researchers separately modeled contacts between the TadA8e deaminase and RNA to address unintended RNA editing. They combined a K1020D mutation in Cas9 with an F156D mutation in TadA8e to create ABE8e-DD. Targeted amplicon sequencing, genome-wide profiling, an R-loop assay and RNA sequencing showed lower detected DNA and RNA off-target activity than several established high-fidelity adenine base editors, according to the paper.[[1]](https://www.nature.com/articles/s41586-026-10794-z)[[3]](https://news.pku.edu.cn/jxky/a3a8cf38a5364d87a2539cde5bb95303.htm)\n\nHoi Yee Chu and Alan Wong, writing in an independent Nature commentary, said the work captured interaction differences correlated with on- and off-target editing by predicting thousands of editor–nucleic-acid complexes. They characterized the result as an improvement in enzyme selectivity.[[5]](https://www.nature.com/articles/d41586-026-02042-1)\n\nThose assays provide several forms of evidence for improved specificity, but they cannot establish the absence of rare or context-dependent errors. The US Food and Drug Administration’s April 2026 draft guidance says genome-editing development programs should assess both off-target editing and broader unintended changes to genome integrity with sequencing-based nonclinical studies.[[6]](https://www.fda.gov/regulatory-information/search-fda-guidance-documents/safety-assessment-genome-editing-human-gene-therapy-products-using-next-generation-sequencing)\n\n## Delivery remains a separate problem\n\nThe evidence reported so far is molecular and cell-based. The paper and accompanying university account do not describe animal dosing, tissue-specific delivery or clinical testing.[[1]](https://www.nature.com/articles/s41586-026-10794-z)[[3]](https://news.pku.edu.cn/jxky/a3a8cf38a5364d87a2539cde5bb95303.htm)\n\nABE8e-DD uses the large Streptococcus pyogenes Cas9 architecture. SpCas9-based base editors have coding sequences of roughly 5.2 kilobases before the addition of a guide, promoters and regulatory elements, exceeding the approximately 5-kilobase capacity of a standard adeno-associated virus vector. Developers can instead use split vectors, compact Cas proteins or nonviral delivery systems, but each approach introduces separate engineering and safety questions.[[4]](https://pmc.ncbi.nlm.nih.gov/articles/PMC9258002/)\n\nThe new variants will therefore need testing in therapeutically relevant cell types, animal models and realistic delivery formats. Researchers must also determine whether mutations selected around one guide and mismatch profile generalize to other targets. Ars Technica noted that ContactSeek’s changes might be more specific to particular guide–mismatch combinations than variants produced through broader directed-evolution campaigns.[[2]](https://arstechnica.com/science/2026/07/team-uses-alphafold-ai-to-redesign-gene-editing-proteins-to-make-them-safer/)\n\n## The code is public; the biological IP is unclear\n\nNo company appears among the paper’s listed author affiliations. Peking University’s announcement names government programs, university research centers and the New Cornerstone Science Foundation as supporters, rather than a commercial sponsor.[[3]](https://news.pku.edu.cn/jxky/a3a8cf38a5364d87a2539cde5bb95303.htm)\n\nThe researchers published the ContactSeek code on GitHub under the Mozilla Public License 2.0 and archived a release through Zenodo. [7] Publicly accessible materials reviewed for this article did not identify a licensing or commercialization partner. The software license, however, does not resolve whether the engineered editor proteins or their therapeutic uses are covered by separate patent applications.\n\nFor now, ContactSeek is best understood as a reproducible protein-engineering workflow with laboratory validation. Its practical value may be in helping researchers redesign an editor around a measured off-target profile before undertaking delivery development and preclinical testing.\n\n## Further sources\n\n[[1] Meng, Haowei et al., “Precise DNA base editing using AlphaFold3-based contact m… ↗](https://www.nature.com/articles/s41586-026-10794-z)\n\n[[2] John Timmer, “Team uses AlphaFold AI to redesign gene-editing proteins to make … ↗](https://arstechnica.com/science/2026/07/team-uses-alphafold-ai-to-redesign-gene-editing-proteins-to-make-them-safer/)\n\n[[3] Peking University, “AlphaFold3 contact-probability artificial intelligence fram… ↗](https://news.pku.edu.cn/jxky/a3a8cf38a5364d87a2539cde5bb95303.htm)\n\n[[4] Davis, Jessica R. et al., “Adenine Base Editing In Vivo with a Single Adeno-Ass… ↗](https://pmc.ncbi.nlm.nih.gov/articles/PMC9258002/)\n\n[[5] Hoi Yee Chu and Alan S. L. Wong, “AI identifies interactions in CRISPR complexe… ↗](https://www.nature.com/articles/d41586-026-02042-1)\n\n[[6] US Food and Drug Administration, “Safety Assessment of Genome Editing in Human … ↗](https://www.fda.gov/regulatory-information/search-fda-guidance-documents/safety-assessment-genome-editing-human-gene-therapy-products-using-next-generation-sequencing)+1 more\n\nThe stories that matter, in one email. Free — unsubscribe anytime.", "url": "https://wpnews.pro/news/alphafold3-guided-base-editor-cut-one-off-target-rate-from-28-to-5", "canonical_source": "https://mlq.ai/news/alphafold3-guided-base-editor-cut-one-off-target-rate-from-28-to-5/", "published_at": "2026-07-26 09:52:34.683005+00:00", "updated_at": "2026-07-26 09:52:36.971798+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-research", "ai-tools"], "entities": ["Peking University", "East China Normal University", "AlphaFold3", "ContactSeek", "Cas9", "Nature", "ABE8e-DD", "Ars Technica"], "alternates": {"html": "https://wpnews.pro/news/alphafold3-guided-base-editor-cut-one-off-target-rate-from-28-to-5", "markdown": "https://wpnews.pro/news/alphafold3-guided-base-editor-cut-one-off-target-rate-from-28-to-5.md", "text": "https://wpnews.pro/news/alphafold3-guided-base-editor-cut-one-off-target-rate-from-28-to-5.txt", "jsonld": "https://wpnews.pro/news/alphafold3-guided-base-editor-cut-one-off-target-rate-from-28-to-5.jsonld"}}