• Protein interfaces involving disordered protein regions are underrepresented in training data for structure prediction models.
• Benchmarking of structure prediction models should consider different modes of protein binding.
• AlphaFold-Multimer varies in performance for prediction of protein interfaces involving purely ordered vs involving disordered protein regions.
• Experimental validation efforts of predicted interfaces lack insights into how disorderedness affects validation probability.

A structural understanding of protein–protein interactions is a key component of many facets of applied molecular biology research. AlphaFold-Multimer (AF-MM) provided a breakthrough in the ability to predict protein–protein interface structure. However, the available training data for this model and the resulting benchmarking and validation efforts show a bias toward interactions between more ordered regions of proteins. Here we highlight some of the successes and limitations of AF-MM and discuss available methods and future directions to enable balanced prediction of all interface types.