AlphaFold2 knows some protein folding principles.

Area of Science:

• Computational Biology
• Biophysics
• Structural Biology

Background:

• AlphaFold2 (AF2) excels at protein structure prediction, but its relation to the protein folding problem remains debated.
• The protein folding problem concerns the dynamic pathway to a static structure, a distinction often blurred with structure prediction.
• Current methods typically rely on Multiple Sequence Alignments (MSAs) to guide AF2's predictions.

Purpose of the Study:

• To investigate whether AlphaFold2 has learned underlying protein folding principles beyond static structure prediction.
• To explore AF2's energy landscape by removing standard inputs like MSAs and templates.
• To determine if AF2 can identify protein folding intermediates and pathways.

Main Methods:

• Operating AlphaFold2 without Multiple Sequence Alignments (MSAs) or initial templates to enable full energy landscape sampling.
• Utilizing recycling and iterative prediction strategies within AF2.
• Analyzing over 7,000 proteins to assess folding behavior based on sequence alone.

Main Results:

• A subset of proteins demonstrated folding using sequence information alone, indicating a smooth learned energy surface in AF2.
• Multiple intermediate structures were discovered by AF2, aligning with experimental data, suggesting a 'local first, global later' folding mechanism.
• For designed proteins, AF2's smooth energy landscape sometimes obscured the detection of expected folding intermediates.

Conclusions:

• AlphaFold2 appears to have learned fundamental aspects of the protein folding process, not solely structure prediction.
• AF2's ability to sample energy landscapes and identify intermediates opens new avenues for studying protein folding dynamics.
• This research provides novel insights into AF2's capabilities and facilitates the experimental discovery of folding intermediates.