For as long as structural biology has guided protein engineering, the logic has been elegantly simple: determine where every atom sits, identify the parts that touch the substrate, and redesign those parts to improve function. If you want a CRISPR enzyme to stop editing the wrong DNA sequences, you find the residues that contact those sequences and mutate them away. The model is static. The premise is that seeing is engineering.
A study published July 22 in Nature now challenges that premise at its foundations. The work, led by co-first authors Haowei Meng, Zhixin Lei, Yongchang Yan, Liren Wang, Sihan Zhang, and Xichen Rao of Peking University, with corresponding authors Dali Li of East China Normal University and Chengqi Yi of Peking University, introduces ContactSeek, an AI framework that does not ask where a CRISPR complex is. It asks where its parts could be. And that shift from static structure to probabilistic contact is producing results that conventional engineering has not matched.
The problem ContactSeek addresses is one of the most stubborn in genome editing. Base editors, the molecular tools that rewrite individual DNA letters without cutting the double helix, have enormous therapeutic potential. But they are not precise enough. Off-target editing — unintended changes at sites that resemble but do not match the intended target — remains a barrier to clinical use. Directed evolution and rational design have produced incrementally better variants, but these approaches are slow, labor-intensive, and often sacrifice on-target activity for specificity.
The Peking University team took a different approach. Instead of engineering first and testing later, they asked whether DeepMind’s AlphaFold3 (AF3) could predict, before any mutation was made, which parts of a base editor were responsible for its off-target behavior. The answer turned out to depend on what they measured.
AlphaFold3 is best known for predicting three-dimensional protein structures with atomic accuracy, a capability that earned its creators a Nobel Prize. But Meng and colleagues realized that the predicted structures themselves were less informative than something else the model produces as a byproduct: contact probability, a per-residue metric that captures the likelihood of side-chain atoms being within approximately 5 angstroms of a nucleic acid strand. Where a resolved structure tells you where a residue is, contact probability tells you how likely it is to be there — a probabilistic map of molecular intimacy rather than a single snapshot.
This distinction proved critical. When the team compared AlphaFold3’s predicted structures of Cas9 bound to on-target versus off-target DNA sequences, the structural differences were too small to be useful. Root-mean-square deviation of atomic positions, the standard metric for comparing three-dimensional structures, barely registered. Contact probability, by contrast, changed detectably. The differences were subtle — sub-angstrom in scale — but they were consistent, and they correlated with experimentally measured off-target editing across multiple guide RNAs.
The insight that probability is more sensitive than position became the engine of ContactSeek. The framework works in four stages. First, the team maps genome-wide off-target sites for adenine base editors (ABEs) using a method called dI-profiling, which captures where the editor acts across the entire genome. Second, AlphaFold3 predicts the ternary complex of Cas9, guide RNA, and DNA for both on-target and each off-target sequence, extracting contact probability matrices for every residue. Third, ContactSeek computes delta-CP: the difference between off-target and on-target contact probability at each residue, identifying consensus contact regions (CCRs) — clusters of neighboring Cas9 residues whose contact patterns shift consistently between on-target and off-target complexes. Fourth, within these CCRs, residues with the strongest contact changes are nominated as specificity-determining residues and prioritized for mutation.
The results are striking by the standards of a field accustomed to marginal gains. A single mutation in Cas9, K1020D, reduced ABE8e off-target editing by 83 percent at the HEK4 guide RNA site and 69 percent at ABEsite16, while preserving on-target activity. Combining that Cas9 mutation with a ContactSeek-identified mutation in the TadA8e deaminase domain, H29D, produced a variant the team calls ABE8e-DD that achieved a 99.2 percent reduction in genome-wide off-target editing at ABEsite16. This single variant outperformed multiple high-fidelity ABE variants developed through earlier directed evolution campaigns, including ABE8e-V28C, V106W, and N108Q.
The modularity of the framework is perhaps its most important feature. ContactSeek is not specific to Cas9 or to adenine base editors. The team applied it to a different editor entirely: LbCas12a-based cytosine base editors (CBEs) using the APOBEC3A deaminase. A Cas12a mutation, R284E, reduced genome-wide off-target editing by 72.4 percent at the RUNX1 guide RNA site. Combining this with an APOBEC3A H29D mutation produced Cas12a(R284E)-A3A(H29D), which achieved a 95.1 percent reduction in off-target editing across multiple guide RNAs. The same framework, the same logic, a completely different editor.
The broader significance of ContactSeek extends beyond any single result. Traditional structure-guided engineering treats macromolecular complexes as static assemblies: the structure is the ground truth, and engineering decisions follow from it. ContactSeek treats the complex as an ensemble of possibilities. A residue that contacts off-target DNA with high probability but on-target DNA with low probability is a specificity-determining residue, even if the distance between the two states is too small to measure by conventional structural alignment. The method extracts signal from noise that static structures cannot resolve.
This probabilistic approach has a natural home in genome editing, where the difference between a therapeutic and a hazard can be a single misplaced base pair. But it likely extends further. The same reasoning — that contact probability distributions contain information that static structures do not — could apply to any biomolecular system where specificity matters: antibody-antigen recognition, enzyme-substrate selectivity, protein-protein interaction engineering. Any process where the difference between right and wrong targets is encoded not in where a protein sits but in where it could sit is a candidate for ContactSeek-style analysis.
For now, the immediate impact is on base editing. The ABE8e-DD variant, with its two mutations and near-complete elimination of off-target editing, sets a new benchmark for adenine base editor fidelity. The Cas12a(R284E)-A3A(H29D) variant does the same for cytosine base editors. Both were identified without directed evolution, without high-throughput screening, without trial and error. ContactSeek predicted them, and the experiments confirmed the predictions.
The researchers have made ContactSeek v1.0.0 publicly available on GitHub and Zenodo, along with all AlphaFold3 prediction outputs and sequencing data. The code is designed to be run on any Cas effector and any deaminase, given experimentally mapped off-target sites. The limiting factor is no longer engineering capacity but the quality of off-target data fed into the model.
What ContactSeek ultimately demonstrates is that the resolution of a structure is not the resolution of the question. A three-dimensional model accurate to the atom can still miss the interaction patterns that distinguish one DNA sequence from another. Probability, not position, turned out to be the relevant metric. That is a lesson that structural biology, still catching up to the AI tools that transformed it, will be absorbing for some time.