AlphaFold as a Prior: Experimental Structure Determination Conditioned on a Pretrained Neural Network

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Machine learning, particularly AlphaFold2, has revolutionized protein structure prediction from sequence.
• Challenges remain in modeling sidechain packing, conformational dynamics, and biomolecular interactions due to limited high-quality data.
• Emerging techniques like cryo-electron tomography (cryo-ET) and high-throughput crystallography generate vast structural data, but model interpretation is a bottleneck.

Purpose of the Study:

• To improve the efficiency of structural analysis by combining experimental measurements with AlphaFold2.
• To develop an augmentation of AlphaFold2, named ROCKET, capable of refining predictions using cryo-EM, cryo-ET, and X-ray crystallography data.
• To demonstrate ROCKET's ability to capture biologically significant structural variations beyond AlphaFold2's scope.

Main Methods:

• Augmenting AlphaFold2 with ROCKET, which refines predictions using cryo-EM, cryo-ET, and X-ray crystallography data.
• Performing structure optimization in coevolutionary embedding space, rather than Cartesian coordinates, to automate complex modeling tasks.
• Utilizing differentiable crystallographic and cryo-EM target functions adaptable to other structure prediction methods.

Main Results:

• ROCKET successfully refines AlphaFold2 predictions, capturing biologically important structural variations not identified by AlphaFold2 alone.
• The method automates challenging modeling tasks, including functional loop flips and domain rearrangements, surpassing current state-of-the-art and manual modeling.
• ROCKET does not require AlphaFold2 retraining and is adaptable to multimers, ligand-cofolding, and other data types.

Conclusions:

• ROCKET offers a novel framework for integrating experimental data with machine learning for enhanced biomolecular structure prediction.
• The ability to efficiently sample barrier-crossing rearrangements opens new avenues for scalable and automated model building.
• ROCKET's extensible framework and adaptable target functions facilitate broader integration of experimental observables with machine learning in structural biology.