Abstract Accurate prediction of protein complex structures by AlphaFold3 and similar programs has been used to predict the presence of protein–protein interactions (PPIs), but this technique has never been applied to an entire genome due to onerous computational requirements and questionable utility. Here we present pooled-PPI prediction, a technique that dramatically improves the accuracy of genome-scale screens compared to a paired approach while simultaneously reducing inference time (~twofold) and the number of jobs (~100-fold). We use this technique to predict the structure of all 113,050 pairwise PPIs in Mycoplasma genitalium using only 2027 AlphaFold3 jobs. This unbiased and comprehensive dataset was highly predictive of known interactions, revealed a previously unappreciated but widespread size bias in AlphaFold interface scores, correctly identified protein–protein interfaces in macromolecular complexes, and uncovered new biology in M. genitalium . This work establishes pooled-PPI prediction as a highly scalable method for uncovering protein–protein interactions and a powerful addition to the functional genomics toolkit.
Todor et al. (Tue,) studied this question.