Proteome-scale prediction of which protein pairs directly interact

• Confidence scores are poor classifiers. Schmid and Walter ("Predictomes", Molecular Cell 2025) show that standard AlphaFold-Multimer confidence metrics (ipTM, pDockQ, avg_models) "do not reliably separate relevant PPIs from an abundance of false positive predictions". In their curated test at a 128:1 negative-to-positive ratio, at 5% FDR the best conventional metric (pDockQ2) recalled 11.3% of positives, their structure-only classifier 42.6% and the SPOC classifier (which adds omics features such as co-expression and localisation) 50.4%. After correcting for AF-Multimer's roughly 40% true-positive rate they estimate an absolute recall of about 20% across all PPIs. About 25% of random protein pairs were predicted to contact each other by their definition, which illustrates the false-positive problem.
• Ranking within a controlled screen is easier. Their SPOC classifier ranked a known partner at median rank 1 among about 1,000 decoys, and recovered the functional partners of selected genome-maintenance proteins from proteome-wide screens of more than 20,000 pairs.
• Experimental validation of novel predictions is limited. The authors state that large-scale experimental validation of novel predicted pairs is still needed; published follow-ups that test batches of high-confidence predicted pairs in cell-free or cell-based assays report that only a minority validate, but no systematic, independently reported validation rate across methods was confirmed from a primary source for this entry.
• Experimental baseline. The largest experimental human binary interactome (HuRI, Luck et al. 2020) screened about 17,500 ORFs and reports roughly 53,000 interactions among about 8,300 proteins; assay sensitivity is itself well below 100%, so ground truth for recall is uncertain.
• Not shown. No method reaches the benchmark above with calibration across species in a blind test.

Proteome-scale interaction maps would prioritise targets for disrupting host-pathogen interactions, find missing partners of disease proteins, annotate orphan proteins, and filter candidates in proximity-labelling and AP-MS experiments. Even partial progress (higher-precision ranking within a screened subset) is already used to design follow-up experiments.

A refereed, blind benchmark with held-out species or protein families and experimentally generated positives and negatives, in which one method gives at least 50% recall at no more than 5% FDR at a prevalence of at most 1 in 100, with reliability curves showing calibration, and the predicted interfaces agree with experimental structures (DockQ $\ge 0.23$) for most true positives.