A research team led by Professor Daesik Kim at Sungkyunkwan University School of Medicine published OpenABE, a structure-guided redesign of the AI-designed adenine deaminase Deam-P32, in Nucleic Acids Research online on July 20, 2026. Across 29 endogenous nuclear loci in HEK293T cells, A-to-G conversion efficiency rose from 13.02% for Deam-P32 to 41.73% for OpenABE1.1, a gap of up to 36.1-fold, placing it in the same band as ABE8e, the current state-of-the-art editor, at 42.61%. ASAP works from the paper itself as the primary source to set out what this redesign changed and what it left open.

Where AI-generated enzymes fell short of clinical-grade efficiency

Conventional adenine base editors derive mostly from TadA, a bacterial protein, which leaves them narrow in sequence diversity and prone to bystander editing. The paper's abstract splits the problem in two. TadA-derived editors suffer from bystander editing and limited diversity, while AI-designed deaminases such as Deam-P32 show lower efficiency and precision than state-of-the-art editors such as ABE8e.

Those two clauses describe exactly where generative protein design stood in genome editing. The strength of AI design is that it proposes sequences outside the natural repertoire, and such sequences occupy a different position from natural enzymes in terms of intellectual property and access. How well they work in living cells was a separate question. The 13.02% average measured here for Deam-P32 is less than a third of ABE8e's 42.61%. That is a viable research starting point and an inadequate therapeutic tool.

The route the team chose was not to run the generative model again for a better sequence. It was to take an AI-designed enzyme that already existed and rework it from a protein-structure standpoint. The title states the choice plainly: structure-guided engineering, applied to AI-derived editors.

Three changes that bought a threefold gain

The OpenABE variants are defined by three amino-acid changes to Deam-P32 plus a C-terminal extension borrowed from ABE8e. OpenABE1.1 carries the L78F substitution and the P103S/A105R substitutions with a 15-amino-acid extension from ABE8e (RQVFNAQKKAQSSIN), and OpenABE1.2 carries the same substitutions with a 17-amino-acid extension (MPRQVFNAQKKAQSSIN). As the abstract puts it, the work optimized DNA engagement and the base-contacting pocket and then added the C-terminal extensions.

Performance was measured across 29 endogenous loci in HEK293T cells.

Editor	Mean A-to-G conversion, 29 loci
Deam-P32 (AI-designed original)	13.02% ± 2.92%
OpenABE1.1	41.73% ± 1.90%
OpenABE1.2	41.03% ± 1.99%
ABE8e (prior state of the art)	42.61% ± 2.36%

Per-locus improvement spans 16.3-fold to 34.4-fold for OpenABE1.1 and 17.5-fold to 36.1-fold for OpenABE1.2. The "16 to 36-fold" figure in the abstract marks the two ends of that spread. Safety metrics come from the same locus set. Bystander cytosine editing was 1.03% ± 0.24% for OpenABE1.1 and 0.71% ± 0.15% for OpenABE1.2 against 1.81% ± 0.81% for ABE8e, while indel frequencies clustered together at 0.54% ± 0.08%, 0.63% ± 0.16% and 0.49% ± 0.12% respectively.

The number that matters is parity with 42.61%, not the 36-fold headline

The central result of this paper is parity with the leading editor rather than a large multiple over a weak baseline. A 36-fold figure inflates easily when the denominator is small. Measured against Deam-P32's 13.02%, the 41.73% average is 3.2 times higher, and at individual loci where the original barely worked at all the ratio exceeds 30. The multiple is dramatic and it does not by itself determine whether the tool is usable.

Usability is settled by the other comparison. OpenABE1.1 at 41.73% and ABE8e at 42.61% are indistinguishable once the reported error ranges are taken into account. For a lab deciding whether to adopt an editor, the operative question is not how many times better it is than its predecessor but whether it can replace the current standard. Twenty-nine loci is a wide enough base to support that parity claim, and a 29-locus average carries different weight from a handful of favorable targets.

The reading that follows concerns the division of labor between AI design and protein engineering. Generative models supply starting points from outside the natural sequence space, and structure-guided refinement raises those starting points to a usable level. This result is a case where chaining the two stages produced performance equal to the standard editor. That is a different picture from AI delivering a finished product on its own, and it maps more closely onto a workflow a lab can reproduce today.

Why 34.09% falling to 2.79% at ATC motifs carries the most weight

The largest safety gap appeared at targets containing an ATC motif. C-to-D bystander editing at the FANCF target was 34.09% ± 0.84% for ABE8e against 2.79% ± 0.12% for OpenABE1.1 and 4.89% ± 0.18% for OpenABE1.2. The team isolated nine ATC-containing targets to examine this behavior.

Those figures matter more therapeutically than the efficiency table. The purpose of an adenine editor is to convert A to G, and a cytosine changed in the same window is an unintended mutation. ABE8e's 34.09% means an unwanted change accompanied more than one edit in three in that sequence context, and 2.79% means the rate fell below a tenth of it. Given two tools with equal efficiency, this line decides the choice.

The comparison is also bound to sequence context. The 34.09% and 2.79% values come from one specific target, FANCF, and at targets without an ATC motif the gap between the two tools is closer to the 1.81% versus 1.03% seen in the 29-locus average. OpenABE's advantage is therefore not a uniform property across all targets but one that widens sharply in particular sequence contexts. Checking whether a planned target carries an ATC motif becomes correspondingly important.

Porting the enzyme from nuclear DNA to mitochondria

The team moved the same variants into a mitochondrial configuration, producing OpenABE-TALEDs. Mitochondrial DNA is hard to reach with guide RNA, so targeting relies on TALE proteins instead. The paper measured editing at mitochondrial targets including ND1-1, ND1-2, CO3, CYB and ATP6, reaching up to 32.7% at position A9 of CYB. The abstract describes these efficiencies as comparable or superior to those of ABE8e-TALEDs.

Delivery experiments used engineered virus-like particles. At the SSTR5 target, OpenABE1.1 reached 68.61% ± 0.94%, OpenABE1.2 reached 43.12% ± 0.83% and ABE8e reached 12.15% ± 0.49%, with specificity ratios reported as high as 181.26 for OpenABE1.1 and 69.59 for OpenABE1.2. In the abstract's wording, eVLP delivery further enhanced specificity and product purity.

Successful transfer to mitochondrial editing widens the scope of the work. An enzyme built for nuclear targets that also functions with a different targeting modality and in a different cellular compartment supports the reading that the improvement belongs to the enzyme rather than to one experimental configuration. Mitochondrial disease has been difficult to address with guide-RNA-based tools, and this adds one more high-efficiency candidate for that space.

The gray zones still on the table

Guide-RNA-independent off-target activity is the one category where OpenABE shows no blanket advantage over ABE8e, since the reductions reported in the orthogonal R-loop assay appear at specific sites rather than across every site tested. In the orthogonal R-loop assay measuring guide-RNA-independent off-target activity, OpenABE1.1 reduced unintended A-to-G editing relative to ABE8e at R-loop sites 5 and 6. That is not a uniform reduction across every site. For RNA off-targets, the paper states that the OpenABE variants showed a lower number of off-target sites than ABE8e.

The second gray zone is experimental scope. The primary nuclear efficiency comparison comes from the HEK293T cell line, and the eVLP results center on the SSTR5 target. Whether cell-line efficiency holds up in primary cells or living tissue is a separate question that the abstract and reported results do not settle. For therapeutic development, how these numbers move when the delivery vehicle and target tissue change is the next gate.

The third is the choice of comparators. This work is built around Deam-P32 and ABE8e as its two axes. ABE8e is a widely used reference point, and the adenine editor lineage contains other variants that a single table does not rank. OpenABE's real position gets fixed when other labs reproduce it on the same target set.

What the paper sets up for genome editing research in Korea

Three practical implications for genome editing research in Korea are visible in this result: where the starting enzyme came from, which institutions appear on the author list, and what standard of evidence the paper sets for adopting an editor. The first concerns what was improved. The starting material, Deam-P32, is an AI-designed sequence, and OpenABE is that sequence plus structure-guided changes. Reaching standard-grade performance from a starting point outside the natural enzyme lineage adds an option to how research groups source their tools.

The second is the institutional lineup. Author affiliations run through the Department of Precision Medicine and the Department of Molecular Cell Biology at Sungkyunkwan University School of Medicine, the Biomedical Institute for Convergence at SKKU, the Department of Cell and Genetic Engineering at the University of Ulsan College of Medicine, Asan Medical Center and the Department of Life Science at Ewha Womans University. Tool development and clinical proximity sit inside a single paper here, which means the path to continuing therapeutic work domestically is already wired.

The third is the standard of evidence. The paper does not argue superiority on efficiency alone; it reports bystander editing, indels, guide-RNA-independent off-targets, RNA off-targets and delivery-specific specificity as separate line items. A team that adopts an editor on the strength of a single efficiency percentage from a vendor sheet or an abstract will miss items like the ATC-motif problem this paper exposes. Checking the motif composition of the intended target first, then selecting the editor to match, is the practical standard this dataset argues for.

Source: ASAP analysis based on the Nucleic Acids Research paper "Structure-guided engineering of AI-derived adenine base editors for nuclear and mitochondrial DNA editing" (Volume 54, Issue 14, gkag737, published online July 20, 2026, DOI 10.1093/nar/gkag737)