SLAE: Strictly Local All-atom Environment for Protein Representation.

Area of Science:

• Computational Biology
• Structural Biology
• Machine Learning

Background:

• Current protein representation methods often overlook crucial side-chain geometry and chemical details.
• Existing approaches typically rely on sequence-based models or simplified backbone graphs.

Purpose of the Study:

• To introduce SLAE, a unified all-atom framework for learning comprehensive protein representations.
• To leverage local atomic neighborhoods, including atom types and interatomic geometries, for feature extraction.

Main Methods:

• Developed SLAE, an all-atom framework utilizing residue's local atomic neighborhood.
• Implemented a novel multi-task autoencoder objective combining coordinate reconstruction, sequence recovery, and energy regression.
• Trained the model on atom types and interatomic geometries.

Main Results:

• SLAE reconstructs all-atom protein structures with high fidelity from learned latent residue environments.
• Achieved state-of-the-art performance on diverse downstream tasks through transfer learning.
• Demonstrated that SLAE's latent space is chemically informative and sensitive to environmental context.

Conclusions:

• SLAE provides a powerful, physically grounded approach to protein representation learning.
• The framework enables quantitative assessment of structural qualities and smooth interpolation of conformations at all-atom resolution.
• SLAE advances the field of computational biology by integrating detailed atomic information into protein representations.