Impact of protein conformational diversity on AlphaFold predictions

Since proteins actually exist as an ensemble of different conformations, and not the single rigid structure that most of use for simplicity’s sake, the authors wanted to see if AlphaFold learned something about conformational diversity too. Short answer: No. Most of the predicted structures are in the holo or ligand-bound form, and predictions get worse when your protein is already conformationally diverse, probably because the evolutionary signal becomes fuzzy. Not all bad though - predicting holo forms is arguably more exciting than apo, and plDDT could be sensitive enough to detect different kinds of protein flexibility.

The case for post-predictional modifications in the AlphaFold Protein Structure Database

Great news - since AlphaFold was trained using inter-residue distances from proteins in the PDB, it learned to leave space for co-factors and post-translational modifications. The authors graft glycans on a model and explain how other kinds of modifications could be added in as well. This is definitely a call to action:

… these early results are highly encouraging to serve as a rallying point for the developers’ community to complete and enrich the predicted protein models with likely modifications, to bring them to their fullest potential and to correctly inform the next generation of structural biology studies."

Artificial intelligence reveals nuclear pore complexity

A Herculean feat - combining cryo-ET, cryo-EM, de novo structure and interface prediction, and coarse molecular dynamics to model the gigantic human nuclear pore complex (NPC) scaffold. There’s even a video

Overall a great demonstration of integrative structural biology with each step backed up by different technologies - predicted monomeric and complex structures were validated with new X-ray structures and fit to EM maps. The one that didn’t fit was remodelled and refined using RoseTTAfold and AlphaFold fragments. Information from the cryo-ET map was used to set up the MD simulation, which in turn matched previous experimental data about NPC constriction.

Applications of AlphaFold beyond Protein Structure Prediction

An interesting grab-bag of strategies, ideas, and preliminary experiments, ranging from mutational stability prediction to protein design. One of the first papers using the intermediate representations returned by AlphaFold instead of just the structure. These are essentially features per residue, and predict point mutation stability change pretty well.

Pre-training Co-evolutionary Protein Representation via A Pairwise Masked Language Model

Using pairwise information while pre-training protein sequences makes the learned sequence representations much better at contact prediction.

Multiple rereads of single proteins at single–amino acid resolution using nanopores

And finally, something a little tangential and further off, Nanopore sequencing is coming to proteins! All these advancements move us ever closer to the futuristic sci-fi handheld device that you point at a sample in the field and instantly get a screen (holographic of course) of all the (PTM-ed) proteins and their (predicted) structures.