Evaluating Autoencoder-Based Featurization and Supervised Learning for Protein Decoy Selection.

Area of Science:

• Computational biology
• Structural biology
• Machine learning

Background:

• Molecular structure data is rapidly growing, increasing the need for effective structure featurization.
• Protein structure is known to encode biological function, making structure-based featurization crucial.
• Autoencoders, a type of neural network, offer versatile architectures for generating features.

Purpose of the Study:

• To investigate and evaluate autoencoders for generating linear and nonlinear featurizations of protein tertiary structures.
• To demonstrate the utility of autoencoder-based featurizations in a practical biological context.
• To assess if these featurizations can aid in protein structure prediction tasks.

Main Methods:

• Utilized autoencoders, implemented with Keras, to derive featurizations from protein tertiary structures.
• Explored variations in autoencoder architectures to yield both linear and nonlinear features.
• Applied featurizations to the problem of decoy selection in protein structure prediction.

Main Results:

• Autoencoder-based featurizations were shown to be meaningful for detecting active tertiary structures.
• Supervised learning methods, using these featurizations, successfully identified relevant structures.
• The study demonstrates the practical applicability of autoencoders in structural biology.

Conclusions:

• Autoencoders provide a powerful and versatile tool for featurizing protein structures.
• Featurizations derived from autoencoders can effectively aid in protein structure prediction and analysis.
• This work opens new research avenues for applying neural networks in structural biology.