AlphaFold Moves from Structure Prediction to Editing Guidance: Gene Editing Finally Finds Its 'Navigator'

Just like moving from handwriting input methods to smart suggestions, or from manual exposure to autofocus, the essence of every technological leap is turning "trial-and-error based on experience" into "probabilistic prediction." The gene editing field is experiencing a similar inflection point.

Over the past decade, CRISPR and other base editors have allowed scientists to modify DNA sequences as easily as correcting typos. But one core issue has remained unresolved: How to ensure the editing tool only modifies the target location without accidentally harming other genes? Off-target effects are like defusing bombs in a minefield—a slight tremor could cause disaster. Traditional solutions rely on experimental screening and structural biology analysis, which are inefficient, costly, and often only offer remedial measures after the fact.

But this time, the story is different. In July 2025, the Yi Chengqi research group at the School of Life Sciences, Peking University, published a paper in Nature proposing an AI framework called ContactSeek. Its core isn't learning protein structures from scratch, but standing on the shoulders of AlphaFold3—using the protein contact probabilities predicted by AlphaFold3 to guide the design of base editors.

This news is noteworthy not because another AI tool appeared, but because it completed a "cognitive closed loop." AlphaFold3 was originally used to predict 3D protein structures. Its output "contact probability"—the likelihood of amino acid residues being spatially close—is essentially a probability map. The cleverness of ContactSeek lies in using this map as navigation data, telling the base editor: where are the safe landing zones, and where are the minefields prone to missing the target.

Let me try to break down this logic. Base editors work by fusing a DNA-binding protein (like Cas9 nickase) with a deaminase, converting C bases to U at specific locations, which then become T. However, the deaminase itself has preferences for DNA sequences. If the sequence near the target matches its preference, off-target editing may occur. Previously, researchers needed extensive experiments to test the performance of different deaminases in various contexts, equivalent to groping in the dark.

ContactSeek's approach is: First, let AlphaFold3 predict the contact probability when the target DNA binds with the deaminase. High contact probability indicates stable binding and low off-target risk; low contact probability indicates unstable binding and high risk of accidental harm. Then, the AI framework provides an "editability score," directly telling experimenters which editor is suitable for this target, or how to optimize the editor sequence.

From an industry trend perspective, this marks AI moving from "understanding biology" to "designing biology." The AlphaFold series previously solved the problem of "reading"—translating amino acid sequences into spatial structures. Now, ContactSeek turns the result of "reading" into a tool for "writing." This transformation has huge potential in gene therapy, synthetic biology, precision agriculture, and other fields. For example, for monogenic hereditary diseases like sickle cell disease and thalassemia, if base editing can achieve near-zero off-target precision, clinical translation risks will be significantly reduced.

However, I must also point out that this technology is currently in the early validation stage. The improved editing efficiency and reduced off-target rates shown in the Nature paper are results on specific targets. Large-scale promotion requires more data for training, as well as validation across different species and cell types. Additionally, AlphaFold3's prediction accuracy for protein complex structures still has limitations, especially regarding dynamic conformational changes. Whether ContactSeek can handle these complexities remains to be seen in future iterations.

Nevertheless, I believe the significance of this breakthrough lies not in the technology itself, but in the methodology. It proves that the "prediction-validation-optimization" closed loop works in bioengineering. Previously, we relied on high-throughput screening to hit the jackpot; now, we rely on AI to calculate probabilities and then design precisely. This is like moving from hand-forging to CNC machining; improvements in tool efficiency bring industrial-level changes.

For readers, I have a specific action suggestion: If you follow gene therapy or synthetic biology, keep an eye on ContactSeek's subsequent papers and open-source code (if released). More importantly, watch out for...

Original link: https://www.ithome.com/0/981/478.htm