Protein model quality assessment using rotation-equivariant transformations on point clouds.

Area of Science:

• Computational biology
• Structural biology
• Machine learning

Background:

• Machine learning for protein structure prediction has advanced significantly, impacting basic science and drug discovery.
• Effective numerical representations of macromolecular structures (e.g., graphs, 3D grids, distance maps) are crucial for machine learning applications.
• Existing methods often involve complex, customized machine learning models.

Purpose of the Study:

• To explore a novel, conceptually simple numerical representation for protein structures in a machine learning context.
• To evaluate the performance of this new representation in a blind experiment during CASP14.
• To assess its competitiveness against highly complex, established methods.

Main Methods:

• Representing atoms as points in 3D space with associated features, starting with element type.
• Employing neural network layers with rotation-equivariant convolutions to update atomic features.
• Aggregating information from all atoms to the alpha carbon level, then to the protein level for prediction.

Main Results:

• The proposed method achieved competitive results in protein model quality assessment.
• The approach demonstrated strong performance and generality despite its simplicity.
• It requires minimal prior information and was trained on relatively limited data.

Conclusions:

• A simple representation of atoms as 3D points with rotation-equivariant convolutions offers a competitive alternative for protein structure prediction.
• This method shows promise for protein model quality assessment, even when compared to sophisticated models like AlphaFold 2.
• The approach highlights the potential of simpler, more generalizable machine learning strategies in structural biology.