Rapid estimation of protein folding pathways from sequence alone using AlphaFold2.

Area of Science:

• Computational Biology
• Structural Biology
• Biophysics

Background:

• AlphaFold2 (AF2) significantly advanced protein structure prediction, but its relationship to the protein folding problem remains unclear.
• Protein structure prediction focuses on static conformations, while protein folding addresses the dynamic process of achieving these structures.

Purpose of the Study:

• To investigate whether AlphaFold2 has implicitly learned principles of protein folding.
• To explore the utility of AF2's learned energy function for discovering folding pathways and intermediates.

Main Methods:

• Operated AlphaFold2 in an ab initio-like mode, excluding multiple sequence alignments (MSAs) and templates, to sample the entire energy landscape.
• Analyzed over 7000 proteins to identify those folding from sequence alone.
• Iterated and recycled predictions to uncover intermediate structures and analyzed designed proteins with optimized local interactions.

Main Results:

• AlphaFold2's learned energy function, despite imperfections, enables rapid (minutes) discovery of protein folding pathways.
• A fraction of proteins were predicted to fold from sequence alone using AF2, indicating a smooth learned energy landscape.
• Recycled AF2 predictions revealed intermediate structures consistent with experimental data, suggesting a 'local-first, global-later' folding mechanism.

Conclusions:

• AlphaFold2 has implicitly learned fundamental aspects of protein folding, extending beyond static structure prediction.
• AF2's energy landscape facilitates the exploration of folding dynamics and the identification of transient protein structures.
• The findings provide insights into AF2's learned capabilities and offer new methods for studying protein folding mechanisms.