AlphaFold as a prior: experimental structure determination conditioned on a pretrained neural network
TL;DR
Researchers developed ROCKET, a tool that combines AlphaFold2's predictive power with experimental data to automatically build accurate protein structures, even from low-resolution or noisy datasets.
Problem / question
While AI models like AlphaFold2 are excellent at predicting static protein structures, they struggle to model dynamic changes, complex interactions, and specific experimental conditions. Meanwhile, experimental methods like cryo-EM and X-ray crystallography generate massive amounts of data, but building accurate atomic models from this data is difficult, especially when the resolution is low and the structural details are blurry.
Methods
The team created a framework called ROCKET that modifies OpenFold, an open-source version of AlphaFold2. Instead of retraining the AI, ROCKET uses mathematical optimization to adjust the evolutionary sequence data fed into the model. It repeatedly tweaks this input so that the AI's structural predictions closely match the actual experimental measurements from X-ray crystallography or cryo-EM data.
Key findings
ROCKET successfully corrected AI-predicted structures to match real experimental states, including large-scale shape changes and detailed side-chain adjustments. It performed on par with human experts and existing tools at high resolutions, and uniquely succeeded at building accurate models from very low-resolution and noisy data where other automated methods typically fail.
Why it matters
This approach automates and accelerates the difficult process of building atomic models from experimental data. By using a pre-trained AI as a guide, scientists can extract meaningful biological insights from challenging datasets without needing the massive computing power required to train a new AI model from scratch.
Limitations
The method requires a lot of computer memory, currently limiting it to refining relatively small protein sections at a time. It also occasionally struggles to correct very short protein loops containing bulky components, does not automatically account for crystal packing forces, and still requires human review when working with very poor-quality maps.
Takeaway
By merging AI-based structure prediction with direct experimental feedback, ROCKET offers a powerful, automated way to solve complex protein structures that were previously too difficult or time-consuming to model.