Generating tertiary protein structures via interpretable graph variational autoencoders.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning

Background:

• Accurately modeling protein structural plasticity is a significant challenge in molecular biology.
• Existing methods, including AlphaFold2, primarily focus on predicting a single protein structure, which is insufficient for understanding dynamic interactions.
• Current data-driven deep learning models struggle to generate physically realistic protein tertiary structures due to limitations in capturing complex data.

Purpose of the Study:

• To develop novel deep generative models for generating realistic protein tertiary structures.
• To address the limitations of current deep learning approaches in modeling highly structured biological data.
• To improve the interpretability of generative models in protein structure prediction.

Main Methods:

• Utilized a graph variational autoencoder (VAE) framework to model protein tertiary structures.
• Represented protein tertiary structures as 'contact' graphs, enabling the application of graph-generative deep learning.
• Developed models capable of capturing rich, local, and distal constraints within protein structures.

Main Results:

• The proposed graph-generative models successfully capture complex structural constraints.
• Achieved disentangled latent representations, enhancing model interpretability by revealing the impact of individual factors.
• Demonstrated state-of-the-art performance through rigorous comparative evaluations on various metrics.

Conclusions:

• The developed graph-generative models represent a significant advancement in protein structure modeling.
• These models offer improved interpretability compared to existing deep learning approaches.
• Graph-generative frameworks hold promise for unraveling the intricate structural complexity of proteins.