Our collaboration with @arneelof 's lab exploring the use of AlphaFold2 for large scale prediction of structures of human complexes is now published @davidfburke @Patrick18287926 et al. nature.com/articles/s4159… We ran predictions for 65,484 pairs, generally showing good agreement with experimental models and cross-link constraints. The major caveat being that the benchmarks assume the interactions are direct.

Jointly with @arneelof's group we applied Alphafold to predict complex structures for 65,000 human protein-protein interactions. Here is what we learned in trying to generate a structurally resolved human interaction network biorxiv.org/content/10.110… In 1,465 comparisons with PDB ~50% of models are deemed accurate. A predicted confidence metric can rank higher confidence models (pDockQ>0.5 -> 80% accurate) which tends to enrich for complexes from huMap (affinity-MS) more than HuRI (Y2H)

Proteins that interact together are often relevant for the same traits and network analyses of GWAS genes can help identify trait relevant cell biology. Here @ibarrioh applied this to study 1002 traits defining a pleiotropy map of human cell biology
biorxiv.org/content/10.110… GWAS genes were selected with the locus to gene score which incorporates diverse features ( genetics-docs.opentargets.org/our-approach/p…) and a comprehensive network was assembled by @intact_project Network expansion can recover disease genes and drug targets not found by GWAS

Thousands of phosphosites have been discovered with <5% having a known function. In our new preprint @d0choa uses a machine learning approach to address this gap

If you care what phosphosites may regulate your protein/process of interest read on

biorxiv.org/content/10.110… with @AndrewJarnuczak and @juan_vizcaino we first reanalyzed most human phospho MS samples in @pride_ebi (~575 days elution time)

The varied samples gave us good coverage and the single MaxQuant run full control of FDR.

We found ~120k high confidence human phosphosites