Visualizing the Residue Interaction Landscape of Proteins by Temporal Network Embedding

Area of Science:

• Computational Biology
• Structural Biology
• Biophysics

Background:

• Understanding protein structural dynamics is key to deciphering complex biomolecular processes.
• Dimensionality reduction techniques are vital for analyzing high-dimensional protein conformational data.
• A significant challenge lies in selecting informative input features for these reduction methods.

Purpose of the Study:

• To develop a novel method for dimensionality reduction of protein conformational landscapes.
• To integrate graph theory-based Residue Interaction Networks (RINs) with advanced autoencoder algorithms.
• To create a generalizable approach for analyzing protein dynamics in simulations.

Main Methods:

• Utilized Residue Interaction Networks (RINs) and their closeness centrality measures.
• Combined RIN centrality with EncoderMap, a hybrid neural network autoencoder and multidimensional scaling algorithm.
• Applied the method to simulations of the Trp-Cage protein and the FAT10 protein.

Main Results:

• Generated a low-dimensional embedding that meaningfully visualizes the residue interaction landscape.
• Successfully resolved intricate structural details of protein behavior while maintaining global interpretability.
• Demonstrated the method's effectiveness on both a fast-folding protein and a multidomain signaling protein.

Conclusions:

• The proposed graph-based embedding approach offers a powerful tool for analyzing protein simulations.
• This method provides generalizable insights into protein folding and multidomain interactions.
• The approach is modular and transferable to diverse protein systems.