Neural representations of cryo-EM maps and a graph-based interpretation

Area of Science:

• Structural Biology
• Biophysics
• Computational Biology

Background:

• Advances in cryo-electron microscopy (cryo-EM) generate high-resolution structural data.
• Protein structure determination requires interpolating discrete voxel data into continuous spatial domains.
• Existing interpolation methods may lack accuracy for complex macromolecular structures.

Purpose of the Study:

• Introduce a novel neural cryo-EM map data format.
• Enable accurate parameterization of cryo-EM maps with continuous density and gradient data.
• Demonstrate the utility of this format through graph-based interpretations of experimental cryo-EM maps.

Main Methods:

• Developed a novel neural cryo-EM map format using neural networks.
• Created graph-based interpretations from 115 experimental cryo-EM maps (1.15–4.0 Å resolution).
• Compared accuracy against conventional tri-linear interpolation.

Main Results:

• Neural cryo-EM map interpolation achieved <0.01 mean absolute error, significantly outperforming tri-linear interpolation (up to 0.12 MAE).
• Graph interpretations from atomic resolution maps (>1.6 Å) provided >99% residue and 85% atomic coverage with 0.19 Å RMSD.
• Lower resolution maps yielded 84% residue coverage, with accuracy correlated to experimental resolution and data quality.

Conclusions:

• The continuous and differentiable neural cryo-EM map format facilitates conversion to alternative formats like graphs.
• Graphs derived from high-resolution maps accurately identify atom locations, serving as input for predictive modeling.
• This approach can generalize to transform any 3D grid data into a continuous, differentiable format for deep learning applications.