A “concerning” observation is that protein interactions inferred from experimental data are not more likely to show direct interactions in AlphaFold models than what we normally consider weak evidence, e.g., coexpression or text mining.
If we trust AlphaFold, it means either the experimental data are very noisy, or the interactions inferred from them are primarily indirect. Either way, we need to think more deeply about what it means for protein-protein interactions in the cellular context.
This project actually took a long time. One of the first questions we studied using this idea was this paper, which used structural modeling and ancestral sequence reconstruction to examine ligand-receptor (sex peptide, SPR) evolution. academic.oup.com/mbe/article/... @official-smbe.bsky.social
In our recent paper in PLoS Computational Biology, we used AlphaFold2 Multimer to predict the structural details of ~28,000 protein-protein interactions from high-confidence STRING pairs in Drosophila, and identified interesting patterns in disordered regions. journals.plos.org/ploscompbiol...
Most predicted dimers involve binding through disordered regions, not just ordered regions. We generated a DmelPPI website melppi.github.io, so you may download publishable data for your pairs of interest.
We wrote a review on using machine learning to study evolutionary genetics and molecular evolution in Trends in Genetics. It is open access—please take a look if you are interested in this topic www.sciencedirect.com/science/arti... @cp-trendsgenetics.bsky.social
Please circulate: I will have a postdoc opening soon, ideally with a start date sometime this summer or fall. If you're interested in empirical population genetics, computational methods development, and/or evolutionary modeling via simulation, don't hesitate to reach out.
For more info on the lab:
