Related Experiment Video
Updated: Apr 3, 2026
09:47
Author Spotlight: Advancing Alzheimer's Research – Exploring Early Detection and Multi-Omics Approaches
Published on: December 15, 2023
2.0K
AlphaFold as a prior: experimental structure determination conditioned on a pretrained neural network
Alisia Fadini1,2, Minhuan Li3,4, Airlie J McCoy1
1Cambridge Institute for Medical Research, University of Cambridge, Cambridge, UK.
Nature Methods
|April 1, 2026
Summary
Researchers developed ROCKET, a tool that enhances protein structure prediction by integrating experimental data. This machine learning approach improves modeling accuracy, especially for complex biological interactions and conformational changes.
Area of Science:
- Structural biology
- Biophysics
- Computational biology
Background:
- Machine learning advances protein structure prediction from sequence.
- Challenges remain in modeling protein dynamics, interactions, and low-resolution data.
- Experimental techniques like cryo-EM generate large datasets but are hard to interpret.
Purpose of the Study:
- To develop a method integrating experimental data into protein structure prediction.
- To overcome limitations of current models in capturing complex biological features.
- To enhance the accuracy and interpretability of atomic models from structural data.
Main Methods:
- Introduced ROCKET, an augmentation of AlphaFold2.
- Refined predicted protein structures using cryo-electron microscopy (cryo-EM), cryo-electron tomography (cryo-ET), and X-ray crystallography data.
- Optimized structures in coevolutionary embedding space, not Cartesian coordinates.
Main Results:
- ROCKET refines structures using experimental data, improving accuracy.
- The method captures biologically relevant structural variations missed by AlphaFold2 alone.
- ROCKET performs well even with low signal-to-noise experimental data.
Conclusions:
- Integrating experimental measurements directly into structure prediction overcomes key limitations.
- ROCKET enables scalable, automated model building without retraining.
- Provides a general framework for combining experimental data with biomolecular machine learning.
