AlphaFold 2 and the future of structural biology

Exceeding expectations, AlphaFold 2 is strikingly successful at predicting the three-dimensional folds of a peptide chain based on its amino acid sequence. This was a fact agreed upon by Kresten Lindorff-Larsen, Guillermo Montoya, Ed Tate, Wouter Boomsma and Robert Tampé during a panel debate at the 2021 ISBUC Annual Meeting. However, we should be wary of taking the predicted structural models from AlphaFold 2 at face value. Predicting a single conformation of an isolated protein chain, only tells a small part of the story and leaves out ensembles and dynamics. Furthermore, the panelist agreed that although the world was surprised by AlphaFold 2’s striking results, it was only a matter of time before the problem cracked, as the stream of “new” unique protein folds in the Protein Data Bank (PDB) had dwindled, meaning that the vast majority of new structures are made from folds already represented in the PDB.

The important question is what we can learn from AlphaFold 2?

Rather than viewing AlphaFold 2 as a competition, structural biologists need to work with it; to see what we can learn and how it can be a useful tool in our research. For example, the built-in tool that assesses the quality of the predictions has proven to be highly accurate. What is really interesting is that when AlphaFold 2 does a poor job of predicting structure, this in itself has proven to be an excellent predictor of intrinsically disordered regions. In addition, AlphaFold 2 tends to predict ligand bound states of proteins, though no ligands are present. For now, it remains unclear whether this is caused by a bias in the PDB towards these states or if they are somehow naturally preferred based on sequence. This bias gives structural biologists an exciting new research area to explore.

What’s next for AlphaFold?

A natural next step for AlphaFold would be to tackle predictions of protein dynamics. However, the panel was not in agreement on how easily this would be achieved. Some thought it is likely that this development could be driven by conformational landscapes determined, for example, by cryo-EM. Others argued that the necessary information is already contained in the sequence and could be coaxed out without additional experimental structures.

Regardless, all agreed that there is a vibrant and exciting future in store for structural biology, where the interface with other biological fields is likely to be significantly strengthened by AlphaFold 2.       

By Matilde K. Nordentoft, Milena Timcenko & Casper de Lichtenberg.

Reporting on a panel discussion between Kresten Lindorff-Larsen, Guillermo Montoya, Ed Tate, Wouter Boomsma and Robert Tampé. Moderated by Henriette Autzen at the 2021 ISBUC Annual Meeting.