This is a little outside of "ML for Omics" other than it being proteomics-related.
They scaled AlphaFold2 predictions for possible homodimers across P. furiosus, E. coli, S. cerevisiae, and H. sapiens proteomes (20-45%), analyzing those with physiological relevance and processed those into higher-order structures. The result: gorgeous structures.
They actually validated one of the predicted large structures by cryo-EM as well, showing that at least some of their predictions are reasonable.
Takeaways:
coiled-coil regions are major drivers of quaternary structure
interaction interfaces in homo-oligomers are ~70% more likely to contain disease mutations than protein surfaces
lots of these complexes appear to be symmetric.