[Paper Review] A Graph Neural Network Approach to Automated Model Building in Cryo-EM Maps

Kiarash Jamali, Dari Kimanius+1 more|arXiv (Cornell University)|Sep 30, 2022

Advanced Electron Microscopy Techniques and Applications26 citations

TL;DR

The paper presents ModelAngelo, a graph neural network that integrates cryo-EM voxel data, protein sequences, and geometry priors to automatically build atomic models in cryo-EM maps, achieving near-manual accuracy for maps better than 3.5 Å.

ABSTRACT

Electron cryo-microscopy (cryo-EM) produces three-dimensional (3D) maps of the electrostatic potential of biological macromolecules, including proteins. Along with knowledge about the imaged molecules, cryo-EM maps allow de novo atomic modelling, which is typically done through a laborious manual process. Taking inspiration from recent advances in machine learning applications to protein structure prediction, we propose a graph neural network (GNN) approach for automated model building of proteins in cryo-EM maps. The GNN acts on a graph with nodes assigned to individual amino acids and edges representing the protein chain. Combining information from the voxel-based cryo-EM data, the amino acid sequence data and prior knowledge about protein geometries, the GNN refines the geometry of the protein chain and classifies the amino acids for each of its nodes. Application to 28 test cases shows that our approach outperforms the state-of-the-art and approximates manual building for cryo-EM maps with resolutions better than 3.5 Å.

Motivation & Objective

• Automate de novo atomic model building in cryo-EM maps using a unified neural framework.
• Integrate voxel-based cryo-EM density with protein sequence information and geometric priors.
• Improve upon state-of-the-art automated model-building methods for maps at ~3 Å resolution.

Proposed method

• Construct a graph where nodes are Cα atoms and edges follow the protein chain; train a graph neural network to refine backbone positions and assign residues.
• Use a three-module, eight-layer SE(3)-equivariant GNN with Cryo-EM Attention, Sequence Attention, and Spatial Invariant Point Attention (IPA) to fuse map data, sequence embeddings, and geometry.
• Incorporate backbone frame updates and Gram-Schmidt-based alignment to maintain realistic bond geometry.
• Leverage a multi-task loss including Cα RMSD, backbone RMSD, amino acid classification, torsion angles, and per-residue confidence to train end-to-end.
• Postprocess by connecting residues into chains using peptide-bond heuristics, perform HMM-based sequence alignment against provided sequences, and recycle predictions to refine the model.
• Train with recycling rounds and denoise initial Cα positions; augment data with noise, sharpening/damping, and random rotations.

Experimental results

Research questions

• RQ1 Can a single GNN architecture jointly utilize cryo-EM voxel data, sequence information, and topology to automate protein model building in cryo-EM maps?
• RQ2 How does incorporating sequence embeddings and IPA-based geometry influence residue identification and backbone refinement compared to map-only approaches?
• RQ3 What is the performance of the proposed method on cryo-EM maps with resolutions better than 3.5 Å in terms of sequence recall and amino acid accuracy?
• RQ4 Can the model generate complete atomic models (including side chains) and handle uncertain or missing residues through training strategies like recycling and denoising?
• RQ5 How does the method compare to current state-of-the-art automated model-building tools (e.g., DeepTracer) across multiple test cases?

Key findings

• The method approximates manual model-building accuracy for maps better than 3.5 Å.
• Practical improvements arise from combining cryo-EM data, sequence embeddings, and graph topology rather than map alone.
• Ablation shows removing sequence or IPA modules degrades results; relying on map alone performs worse than DeepTracer.
• Pruning after postprocessing generally does not reduce sequence recall, indicating removal of incorrect regions.
• ModelAngelo (ModelAngelo) is slower than DeepTracer but achieves closer alignment to deposited models on tested cases.
• Across 28 test maps, the approach yields higher-quality models with integrated multi-modal information and recycling-based refinement.