What if the most useful thing AlphaFold3 can tell you about a protein-DNA complex is not where every atom sits, but how likely each pair of atoms is to touch?
That question drove a team of researchers from East China Normal University and Peking University to build ContactSeek, a computational framework published today in Nature that uses AlphaFold3’s contact probability outputs to dramatically improve the specificity of DNA base editors. The results upend a widely held assumption in protein engineering: that knowing the precise three-dimensional structure of a complex is the best route to understanding and improving its function. Instead, the team found that AlphaFold3’s secondary prediction — a matrix of pairwise contact probabilities between every residue and every nucleotide in a complex — captured interaction differences that the predicted 3D structure missed entirely.
The structure that was not enough
Contact probability (CP) is a side channel of information in AlphaFold3. Alongside the familiar atomic coordinates in a .cif file, every AF3 run produces a `_confidences.json` file containing a matrix of values between 0 and 1 for every token pair in the input. A value of 0.95 between Cas9 residue K1020 and a specific DNA nucleotide means AF3 is 95% confident those two moieties are in contact. It is an ensemble quantity: it reflects the distribution across AlphaFold3’s diffusion sampling runs rather than any single folded state.
First author Haowei Meng and colleagues began by mapping the genome-wide off-target profiles of adenine base editors (ABEs) using a method called dI-profiling, generating a comprehensive catalog of off-target DNA sites for Cas9-TadA8e editors at multiple guide RNAs. They then fed those off-target DNA sequences into AlphaFold3 alongside the corresponding on-target sequences, generating thousands of predicted ternary complexes of Cas9, single-guide RNA, and target DNA.
When the researchers compared how well 3D structural deviations (measured by root-mean-square deviation, RMSD) versus CP changes distinguished on-target from off-target complexes, the result was unambiguous. CP was systematically more sensitive. RMSD changes between on- and off-target complexes were often too small to be reliably parsed — sub-angstrom shifts that fell within the noise of the prediction. CP changes, by contrast, showed clear, localized differences that mapped directly onto known functional domains of Cas9. A single-nucleotide mismatch at a given spacer position produced a distinctive CP change signature in the Cas9 residues that contact that region of the DNA-RNA hybrid, even when the overall 3D structure appeared essentially unchanged.
This makes physical sense. Protein-DNA-RNA complexes are dynamic assemblies that sample multiple conformational states. A static structure captures one snapshot, but contact probability encodes the ensemble: how often a given interface forms, how stable it is, and how it responds to sequence perturbations. An off-target DNA sequence might not shift a side chain’s position by much, but it can substantially reduce the probability that the side chain forms a stabilizing contact — a difference a single structure cannot reveal but a CP matrix can.
The ContactSeek pipeline
Armed with this insight, the team built ContactSeek, a modular computational framework that turns CP matrices into actionable engineering targets. The pipeline works in three stages.
First, for each off-target DNA sequence, ContactSeek computes a delta-CP matrix: the difference between the off-target CP matrix and the on-target CP matrix. These delta-CP values capture which residue-nucleotide contacts weaken or strengthen when the DNA sequence deviates from the intended target.
Second, it identifies consensus contact regions (CCRs). These are stretches of neighboring residues in the Cas9 protein whose contact probabilities change in a correlated manner across the set of off-target sequences. A CCR is essentially a protein domain whose interaction with nucleic acids is collectively perturbed by off-target binding, making it a promising site for engineering.
Third, ContactSeek ranks residues within each CCR by a metric called contact enhancement correlation — essentially, how tightly a residue’s CP change tracks with the experimental off-target editing signal measured by sequencing. Residues whose contact probability shifts correlate strongly with off-target editing rates are the most likely specificity determinants.
From computation to better editors
The team applied ContactSeek to the widely used ABE8e adenine base editor, which fuses a catalytically enhanced TadA8e deaminase to a Cas9 nickase. ContactSeek nominated two key specificity-determining residues: K1020 in the Cas9 REC2 domain and H29 in the TadA8e deaminase itself.
K1020D was the top hit in Cas9. Cas9 residue K1020 sits in a loop of the REC2 domain near the nontarget DNA strand, and its contact probability with the DNA changed more than almost any other residue when off-target sequences were analyzed. Molecular dynamics simulations confirmed that the K1020D substitution subtly repositions the loop, reducing its contact with the nontarget strand in a way that disproportionately penalizes off-target binding while leaving on-target interactions largely intact.
On the deaminase side, ContactSeek identified TadA8e H29D as the top specificity-enhancing mutation. H29 sits near the active site pocket of TadA8e, and its contact probability with the RNA guide and DNA substrate changed systematically across off-target sequences. The H29D substitution reduced the deaminase’s promiscuous engagement with imperfectly positioned substrates.
The combined variant — ABE8e-DD (Cas9 K1020D + TadA8e H29D) — produced striking results. Genome-wide off-target profiling showed that ABE8e-DD reduced total off-target editing by 98-99% relative to the parental ABE8e at multiple guide RNAs, while maintaining on-target editing efficiency. In head-to-head comparisons against several known high-fidelity ABE variants (ABE8e-V28C, ABE8e-V106W, ABE8e-N108Q), ABE8e-DD achieved superior specificity across the board, as measured by targeted amplicon sequencing, genome-wide dI-profiling, and orthogonal R-loop assays.
Beyond adenine editors
ContactSeek is not limited to ABEs. The team demonstrated its modularity by applying the framework to an entirely different system: LbCas12a-based cytosine base editors (CBEs). Cas12a is a distinct CRISPR effector with a different guide RNA structure and DNA cleavage mechanism. Feeding off-target sequences of a Cas12a-APOBEC3A CBE into the same pipeline identified Cas12a R284E and APOBEC3A H29D as specificity-determining residues. The resulting variant showed substantially reduced off-target editing, confirming that ContactSeek generalizes across CRISPR effector families.
The framework also proved effective for reducing RNA off-target editing. Base editors, particularly those using TadA8e and APOBEC3A deaminases, can promiscuously edit RNA transcripts, causing unintended A-to-I or C-to-U changes in the transcriptome. ContactSeek adapted its approach to RNA off-targets by classifying transcriptome-wide RNA editing sites into high, medium, and low activity groups, computing delta-CP between groups, and identifying CCRs in the deaminase that correlated with RNA off-target activity. The resulting TadA8e H29D mutation simultaneously reduced RNA off-target editing without sacrificing on-target DNA activity.
A new paradigm for specificity engineering
The broader significance of ContactSeek extends beyond base editing. For years, protein engineers have relied on structural information — crystal structures, cryo-EM maps, and more recently predicted structures from AlphaFold2 — to guide rational design. The implicit assumption has been that higher structural resolution always yields better engineering insights. ContactSeek challenges that assumption by showing that contact probability, a statistical ensemble property, can outperform static structure for a task that depends on dynamic molecular recognition.
This is particularly relevant for genome editing, where the difference between specific and promiscuous activity often hinges on how a protein-DNA complex responds to mismatches, bulges, and other deviations from the perfect target. A crystal structure shows you the ideal state. Contact probability shows you how the system behaves when the substrate is imperfect — and for specificity engineering, that is precisely what matters.
The approach is also experimentally efficient. Traditional strategies for improving base editor specificity have relied on large-scale mutagenesis screens, directed evolution campaigns, or structure-guided rational design one mutation at a time. ContactSeek requires only off-target sequencing data (increasingly routine) and AF3 predictions (rapidly becoming affordable). It narrows the search space from thousands of possible mutations to a handful of high-confidence candidates in a single computational pass.
With the code and data openly available on GitHub and Zenodo, ContactSeek provides a template for how AI-driven contact analysis can accelerate the development of safer genome editing tools. As the field moves toward therapeutic applications of base editing, the ability to systematically eliminate off-target activity while preserving on-target potency will be critical. ContactSeek suggests that the path forward may not require sharper structural snapshots, but a willingness to look at the probabilities instead.