Streamline Protein Structure Representation with AtomSurf on Vecura

What is AtomSurf?

AtomSurf is a powerful, learnable protein structure encoder that enables the transformation of complex 3D protein structures into high-quality, geometric embeddings. It achieves this by simultaneously encoding two complementary views: an atom/residue graph for connectivity and sequence context, and a molecular surface mesh for shape and electrostatics. By integrating these views, AtomSurf provides a versatile and robust featurization framework for structural biology.

It helps researchers convert raw PDB files into machine-learnable features suitable for a wide array of downstream geometric deep-learning tasks. It is especially useful for scientists working on binding-site prediction, protein-protein interaction analysis, ligand identification, and mutation stability studies.

What can users do with AtomSurf on Vecura?

With AtomSurf on Vecura, users can:

• Generate detailed surface meshes from PDB files automatically using MSMS integration.
• Extract high-dimensional graph representations of protein residues and bonded structures.
• Compute rich per-vertex and per-node embeddings that combine 3D geometry with evolutionary information (via optional ESM2 integration).
• Utilize the resulting structural features as input for specialized downstream models, such as those predicting binding affinity or interaction hot-spots.

What the output means

The output provides four distinct data artifacts: raw surface mesh data, a residue-level graph structure, per-vertex surface embeddings, and per-node graph embeddings. These files are saved in the .pt format, ready to be loaded into your custom downstream PyTorch models for classification, regression, or generative modeling tasks.

This output should be used to support scientific decision making. It does not replace experimental validation.

Why this matters

In structural biology and drug discovery, the ability to accurately represent a protein's 3D surface and internal connectivity is paramount. Traditional methods often struggle to capture both the solvent-exposed geometric features—which drive molecular recognition—and the deep-seated covalent/sequence context that dictates protein function.

AtomSurf bridges this gap by creating a unified encoding framework. By providing a standardized, pre-trained featurization tool, it democratizes access to complex geometric deep-learning workflows, allowing researchers to focus on their specific scientific questions rather than the intricacies of structural data preprocessing.

• Developed by: Research group associated with the original AtomSurf paper (arXiv:2309.16519)
• Source: Official GitHub Repository
• Reference: AtomSurf: Surface Representation for Learning on Protein Structures (arXiv 2309.16519)