ModelAngelo

ModelAngelo employs a graph neural network (GNN) to directly identify Cα atom positions and trace the protein backbone in 3D density maps. Unlike traditional methods that rely on heuristics or template matching, the GNN learns to recognize the characteristic connectivity patterns of polypeptide chains. The network operates on a graph representation of the density map, where nodes represent potential atom positions and edges encode spatial relationships. This approach is robust to local resolution variations and can trace through regions of weak or ambiguous density that would stall conventional algorithms.

Once the backbone is traced, ModelAngelo automatically assigns amino acid identities by integrating sequence information with the geometric features of side-chain density. The algorithm evaluates the likelihood of each of the 20 standard amino acids at every Cα position based on:

• Side-chain density shape and volume in the cryo-EM map
• Backbone geometry constraints (Ramachandran angles)
• Sequence alignment when a reference sequence is provided This eliminates the manual, error-prone step of mutating and refining residues during model building.

ModelAngelo functions as a fully automated pipeline that takes a cryo-EM density map and an optional amino acid sequence as input and produces a near-complete atomic model. The workflow proceeds through distinct stages:

• Graph construction from the density map voxels
• GNN-based Cα tracing to identify backbone topology
• Main-chain and side-chain building using geometric refinement
• Sequence assignment and rotamer optimization
• Iterative real-space refinement against the experimental map This automation dramatically reduces the human effort required to go from a 3D reconstruction to a publishable atomic model.

ModelAngelo is designed to integrate seamlessly into standard cryo-EM processing workflows. It accepts MRC format density maps directly from RELION and cryoSPARC refinement jobs. The output atomic model is written in standard PDB format, compatible with downstream tools such as:

• Coot for manual inspection and correction
• PHENIX and REFMAC for reciprocal-space refinement
• ISOLDE for molecular dynamics flexible fitting This interoperability ensures that ModelAngelo fits into existing structural biology pipelines without requiring proprietary formats or conversion steps.

ModelAngelo is optimized to perform robustly at moderate resolutions (2.5–4.0 Å), where side-chain density is often ambiguous and manual model building becomes challenging. The GNN architecture leverages learned priors about protein geometry to make informed decisions even when local density quality is poor. Benchmarking on EMDataResource targets demonstrates that ModelAngelo consistently produces models with:

• Higher sequence registry accuracy than conventional auto-building tools
• Fewer chain breaks and connectivity errors
• Improved map-model correlation after automated refinement This makes it particularly valuable for the resolution range where most cryo-EM structures are determined.

ModelAngelo is distributed as open-source software under the Apache 2.0 license, with the source code available on GitHub at https://github.com/3dem/model-angelo. The software is implemented in Python with PyTorch for GPU-accelerated inference. Installation is available via:

• Conda package manager for environment management
• Pre-trained model weights distributed alongside the code
• Comprehensive documentation including tutorials and example workflows The open-source model enables community contributions, validation, and integration into custom processing pipelines.