Rapid estimation of protein folding pathways from sequence alone using AlphaFold2

Abstract

AlphaFold2 (AF2) revolutionized protein structure prediction, yet it is often conflated with the protein folding problem. Structure prediction seeks a static conformation, whereas folding concerns the dynamic process of structure formation. We challenge the current status quo, showing that AF2 has implicitly learned some folding principles. Its learned biophysical energy function, though imperfect, enables rapid discovery of folding pathways within minutes. Operating AF2 without multiple sequence alignments (MSAs) or templates forces sampling across its entire energy landscape, akin to ab initio modeling. Among over 7000 proteins, a fraction folds from sequence alone, highlighting the smoothness of AF2’s learned surface. Iterating and recycling predictions uncover intermediate structures consistent with experiments, suggesting a “local-first, global-later” mechanism. For designed proteins with optimized local interactions, AF2’s landscape becomes too smooth to reveal intermediates. These findings illuminate what AF2 has learned and open avenues for probing protein folding mechanisms and experimental intermediates.