Also, I would like to thank the NSF and NIH for funding, and DOE Argonne, NSF ACCESS, and Indiana JetStream2 for computational resources.

There are lots of other results in the manuscript so please check it out. I definitely want to thank the team of Miles Woodcock-Girard, Erin Claussen, and Samantha Fischer who did the work.

Finally, we utilize our network to build structural models of a ciliary protein complex linked to oral-facial-digital syndrome. We identified several human pathogenic mutations at the interface of OFD1 and FOPNL (CEP20), pointing to a mechanism of pathology.

Encouraged by this, we prioritized ~2,500 high-confidence DirectContacts2 protein pairs and ran them through AF3. We see excellent enrichment of high quality models consistent with the AF2 models.

First, using a compendium of computed structural models for evaluation, we see our DirectContacts2 network outperforms other networks at prioritizing protein pairs for AlphaFold2 modeling.

Previous work by @bioinfo.se‬ and @pedrobeltrao.bsky.social‬, among others, prioritized pairs using protein interaction networks, including using our hu.MAP complex map.

We expect our network to be extremely valuable in applying computational structural modeling tools (e.g. AlphaFold, Boltz) to the human interactome. It is computationally infeasible to apply these algorithms to all ~200 million possible human pairwise interactions, so prioritization is key.

We evaluated our network on a benchmark of known direct interactions from the PDB and see that it outperforms other networks including our previous attempt at solving this problem (DirectContacts doi.org/10.1371/jour...) and our hu.MAP co-complex networks!

We integrated data from 25k mass spectrometry experiments (similar to our previous work, hu.MAP3.0, doi.org/10.1038/s443...), but this time trained on direct physical interactions from the PDB. We then applied our model to ~26M human protein pairs, giving a confidence score for each pair.

Determining whether two proteins have a direct physical interaction is an important (and difficult!) problem as protein complex assembly relies on physical interactions and protein interaction interfaces are the target of many therapeutics.