ModelAngelo: Automated Atomic Model Building for Cryo-EM

ModelAngelo is a state-of-the-art machine-learning approach designed to automate the labor-intensive process of atomic model building and protein identification in electron cryo-microscopy (cryo-EM) maps. Developed by researchers at the MRC Laboratory of Molecular Biology, ModelAngelo combines cryo-EM density information with protein sequence and structural data using a multimodal graph neural network (GNN).

By replacing manual intervention with objective, automated procedures, ModelAngelo removes significant bottlenecks in structure determination and enables the identification of unknown protein subunits in complex biological assemblies purified from endogenous sources.

Key Innovations: Multimodal Structural Intelligence

ModelAngelo utilizes a specialized three-step pipeline to translate raw cryo-EM density into complete atomic models.

• Multimodal GNN Architecture: Integrates local information from cryo-EM maps, user-provided protein sequences, and structural geometry in a single end-to-end framework.

• Three-Step Prediction Pipeline:

  1. Residue Localization: A convolutional neural network (CNN) predicts voxels containing amino acid Cα atoms or nucleic acid phosphor atoms.

  2. Graph Optimization: A GNN optimizes residue positions, orientations, and identities while predicting side-chain torsion angles.

  3. Post-Processing: Converts predicted feature vectors into a post-processed atomic model with idealized geometries.

• HMM-Based Identification: Leverages predicted amino acid probabilities to perform hidden Markov model (HMM) sequence searches, identifying proteins with unknown sequences more accurately than human experts.

• Nucleotide Backbone Mastery: Builds accurate backbones for DNA and RNA with accuracy comparable to human experts, significantly accelerating nucleic acid modeling.

• Objective Confidence Measures: Predicts a local confidence score for backbone geometry based on predicted r.m.s.d. values, stored in the B-factor column of the output coordinate file.

Performance Benchmarks

ModelAngelo sets a new benchmark for automated modeling, producing initial models of comparable quality and completeness to those generated by human experts.

Scientific Breakthroughs in Structural Determination

De Novo Identification of Axonemal Proteins

ModelAngelo was used to identify previously unassigned densities in large complexes, such as the radial spokes of Chlamydomonas reinhardtii. It identified four additional radial spoke proteins (RSP24–27) and two central apparatus proteins (FAP92 and FAP374), some of which were not even annotated in previous genome versions.

Modeling Massive Endogenous Supercomplexes

Applied to the 16.7 MDa phycobilisome supercomplex, ModelAngelo successfully identified six protein chains that had remained unidentified despite significant manual effort. The resulting models showed excellent agreement with side-chain densities and aligned closely with AlphaFold2 predictions.

Increasing Objectivity in Model Building

By informing researchers which parts of a map can be confidently interpreted and which should be left uninterpreted, ModelAngelo increases the objectivity of structure determination. This reduces manual errors and makes cryo-EM accessible to a broader scientific audience.

ModelAngelo on Tamarind Bio: Automated EM Interpretation

Tamarind Bio provides an optimized environment to deploy ModelAngelo’s high-speed modeling and sequence identification workflows.

• GPU-Accelerated Inference: Run ModelAngelo on high-performance A100 GPU clusters, reducing multi-MDa complex modeling from days of manual labor to hours of automated processing.

• Integrated Identification Tools: Use the hmm_search feature to automatically identify unknown subunits against large proteomes.

How to Use ModelAngelo on Tamarind Bio

1. Access the Platform: Log in to tamarind.bio and select the ModelAngelo tool.

2. Upload Cryo-EM Map: Provide your cryo-EM density map (ideally with resolution better than 4 Å).

3. Input Sequences (Optional): Upload a sequence file containing the protein chains expected in the complex to guide modeling.

4. Configure Modeling Mode:

   • Standard Build: For maps with known sequences.

   • No-Sequence Build: For identifying unknown proteins in unassigned densities.

5. Run Pipeline: The platform executes the CNN residue detection followed by three rounds of GNN recycling for optimal structure refinement.

6. Evaluate Confidence: Review the output coordinate file, colored by predicted confidence scores (mapped to B-factors) to identify reliable regions.

7. Refine & Export: Download the initial atomic model for final manual refinement in standard tools like Servalcat or Phenix.

Source