AlphaFold2 knows some protein folding principles.

AlphaFold2 (AF2) has revolutionized protein structure prediction. However, a common confusion lies in equating the protein structure prediction problem with the protein folding problem . The former provides a static structure, while the latter explains the dynamic folding pathway to that structure. We challenge the current status quo and advocate that AF2 has indeed learned some protein folding prin- ciples, despite being designed for structure prediction. AF2's high-dimensional parameters encode an imperfect biophysical scoring function. Typically, AF2 uses multiple sequence alignments (MSAs) to guide the search within a narrow re- gion of its learned surface. In our study, we operate AF2 without MSAs or initial templates, forcing it to sample its entire energy landscape - more akin to an ab initio approach. Among over 7,000 proteins, a fraction fold using sequence alone, highlighting the smoothness of AF2's learned energy surface. Additionally, by combining recycling and iterative predictions, we discover multiple AF2 interme- diate structures in good agreement with known experimental data. AF2 appears to follow a "local first, global later" folding mechanism. For designed proteins with more optimized local interactions, AF2's energy landscape is too smooth to detect intermediates even when it should. Our current work sheds new light on what AF2 has learned and opens exciting possibilities to advance our understanding of protein folding and for experimental discovery of folding intermediates.