A team from Peking University, led by Yi Chenci, in collaboration with Professor Li Dali's team from East China Normal University, presented a breakthrough paper in the journal Nature. In it, they described ContactSeek—an artificial intelligence framework that uses AlphaFold3 contact probabilities to systematically develop gene editing tools with enhanced precision.
This work marks the fusion of AI-based structural biology and gene editing technologies. The researchers demonstrated how protein structure prediction models can directly guide the creation of therapeutic tools.
The ContactSeek framework aims to solve a fundamental problem in base editing: maintaining high editing activity at the target site while simultaneously minimizing unwanted changes in other parts of the genome. Base editors combine the Cas protein with a deaminase, allowing one DNA nucleotide pair to be converted into another without the need to create double-strand breaks, making them important tools for both fundamental research and gene therapy development.
However, clinical application is limited due to off-target editing, where the editor alters unintended regions of the genome. The team applied the DeepMind AlphaFold3 model, which was extended to predict protein-nucleic acid complexes, to predict DNA-RNA-protein trimers at both target and off-target sites, and conduct a comparative analysis.
The main finding was that the contact probability itself provided the greatest value, rather than the structure prediction. This probability sensitively captures changes caused by single-base mutations or amino acid substitutions that might have been missed during structure prediction. By integrating this contact probability with off-target editing sequencing data, ContactSeek was able to identify key amino acid residues in both Cas proteins and deaminases that determine editing specificity.
Based on this framework, the team designed a series of novel variants of Cas9 for adenine base editors. These variants maintain efficient editing at the target site while significantly reducing off-target DNA editing. The resulting variants represent a new generation of highly precise editing tools with potential for treating genetic diseases where editing accuracy is critical, such as sickle cell anemia, beta-thalassemia, and certain inherited metabolic disorders.
The study opens a new paradigm in protein engineering for gene editing. Instead of the laborious method of trial-and-error mutagenesis or large-scale screening, ContactSeek offers a predictive AI system that indicates which amino acids should be modified to achieve the desired functional outcome. This approach is applicable not only to base editing but also to other genome editing tools, including prime editors and CRISPR-Cas systems, where the specificity of protein-nucleic acid interaction determines the safety and efficacy of therapy. The publication in Nature places Chinese research in structural biology and gene editing at the forefront of AI-driven protein engineering, demonstrating the transition of AlphaFold-class models from structure prediction to functional prediction and tool design.
The Chinese startup Moonshot has introduced the artificial intelligence model Kimi K3. According to the company, this system is capable of processing a significantly larger volume of data per query.
Kimi K3 was developed to perform tasks requiring advanced thinking, complex programming, and intensive operations.
Arthur Igreya explained how Moonshot's new product, Kimi K3, promises to raise the level of artificial intelligence development.
