ProteinNet: a standardized data set for machine learning of protein structure

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Deep learning advances protein structure prediction and design.
• Standardized datasets are crucial for method assessment and accessibility in machine learning.
• Existing protein datasets lack high-quality alignments and robust validation splits for machine learning.

Purpose of the Study:

• To create standardized datasets for training and assessing machine learning models of protein sequence-structure relationships.
• To provide a comprehensive resource for the bioinformatics community.

Main Methods:

• Developed the ProteinNet dataset series.
• Integrated sequence, structure, and evolutionary information.
• Generated high-quality multiple sequence alignments using high-performance computing.
• Created standardized training/validation splits mimicking CASP experiment difficulty using evolution-based distance metrics.

Main Results:

• ProteinNet provides programmatically accessible data formats for machine learning frameworks.
• The dataset includes comprehensive multiple sequence alignments for structurally characterized proteins.
• Validation sets are designed to accurately reflect the difficulty of protein structure prediction challenges.

Conclusions:

• ProteinNet serves as a valuable and accessible resource for developing and evaluating machine-learned models of protein structure.
• Facilitates reproducible research and advancement in the field of protein structure prediction.