Researchers from Peking University and East China Normal University developed the ContactSeek AI framework, published July 25, 2026, in Nature. By utilizing AlphaFold3 contact probability, the team engineered Cas9 variants for adenine base editors that maintain high on-target activity while significantly reducing off-target DNA editing.
The primary hurdle in base editing is the tendency for editors to modify unintended genomic sites, a flaw that has limited their clinical utility.
Contact Probability vs. Structural Prediction
The researchers initially attempted to use AlphaFold to model the full CRISPR complex, including the target DNA, guide RNA, Cas9 sequence, and the enzyme that modifies bases. To fix this, the team simplified the input to only the DNA, RNA, and Cas9 protein, as the latter is the primary driver of sequence specificity.
This simplification revealed a critical distinction: while structural predictions were useful, contact probability proved most valuable. While about two-thirds of off-target sites caused the Cas9 protein to adopt a slightly different structure, over 95 percent altered which amino acids contacted the RNA.
Contact probability measures the chance that two items, such as nucleotides or amino acids, are within a distance of eight Angstroms. By comparing these probabilities between on-target and off-target sites, the team could identify exactly which amino acids in Cas9 shift when a mismatch occurs between the guide RNA and the DNA.
Engineering Precision Base Editors
ContactSeek does not just predict structure; it identifies the specific amino acid residues in both deaminases and Cas proteins that determine editing specificity. The researchers focused on clusters of amino acids that shifted when bound to off-target sites, treating these areas as regions adapting to mismatched bases.
Using this AI-driven predictive framework, the team engineered a new generation of Cas9 variants for adenine base editors. These tools are designed to convert one DNA base to another without requiring double-strand breaks, which minimizes the risk of genomic instability.
This approach replaces the traditional, labor-intensive process of trial-and-error mutagenesis or large-scale screening with a systematic method for identifying which amino acids to modify to achieve a specific functional outcome.
Clinical Applications for Genetic Disorders
The ability to minimize off-target effects is paramount for treating genetic diseases where precision is a safety requirement. The developed variants have potential applications in treating sickle cell disease, beta-thalassemia, and certain inherited metabolic disorders.
Because the framework is generalizable, its utility extends beyond base editing. The researchers indicate it can be applied to other genome editing tools, including prime editors and various CRISPR-Cas systems, where the specificity of protein-nucleic acid interactions determines the safety and efficacy of the therapy.
Comparative AI Approaches in Nuclease Design
The ContactSeek methodology represents one of several recent AI-driven efforts to refine molecular scissors
. While the Peking University team used contact probability to refine existing Cas9 variants, other researchers have taken a different route via inverse protein-folding models.
For example, a team led by Jennifer Doudna used an AI model to reverse engineer protein sequences based on a desired three-dimensional structure to create a synthetic nuclease called SynTnpB. Unlike the ContactSeek focus on contact probability, the Doudna team coupled their AI model with evolutionary constraints to ensure mutations did not disrupt crucial binding areas.
| Approach | Primary AI Mechanism | Target Outcome |
|---|---|---|
| ContactSeek | AlphaFold3 Contact Probability | Reduced off-target editing in Cas9 adenine base editors |
| SynTnpB | Inverse Protein-Folding + Evolutionary Constraints | Synthetic nuclease with high efficiency and low off-targets |
The Shift Toward Functional Prediction
The publication of the ContactSeek framework in Nature marks a transition in how AlphaFold-class models are used. These models are moving from simple structure prediction—determining what a protein looks like—to functional prediction, which determines how a protein behaves and how it can be redesigned for therapeutic use.
By integrating contact probability with high-throughput off-target editing sequencing data, the framework provides a blueprint for designing tools that are safer for human use.
Keep reading
