AlphaFold2's training set powers its predictions of fold-switched conformations

Area of Science:

• Computational Biology
• Structural Biology
• Deep Learning

Background:

• Deep learning models like AlphaFold2 (AF2) predict protein structures.
• AF2 has shown success in predicting multiple conformations, including those of fold-switching proteins.
• It remains unclear if AF2 generalizes to predict novel alternative conformations beyond its training data.

Purpose of the Study:

• To assess if CFold, an AF2 implementation trained on limited data, can reliably predict alternative conformations of fold-switching proteins.
• To investigate the role of training data in AF2's ability to predict protein conformational diversity.
• To explore potential future developments for more reliable protein structure generation.

Main Methods:

• CFold, an AF2 variant, was trained on a restricted dataset of protein structures.
• CFold was used to predict alternative conformations for eight fold-switching proteins from six families.
• Various sequence sampling techniques were employed, generating 1300-4400 structures per protein.

Main Results:

• CFold accurately predicted only one alternative protein conformation not present in its training set with high confidence.
• CFold generated experimentally inconsistent structures with higher confidence than accurate predictions.
• AF2's success in predicting fold-switcher conformations appears largely attributed to memorization from its training data.

Conclusions:

• AF2's current capability to predict alternative protein conformations is primarily driven by its training dataset.
• The study highlights limitations in generalizing conformational predictions beyond learned examples.
• Future advancements, potentially using sequence pruning, could enhance the reliability of generative protein structure models.