When can AlphaFold predict the oligomeric states of proteins?

Homooligomerisation is a prevalent and important process that many proteins undergo to form the quaternary structures required for biological function. However, determining oligomeric states and structures experimentally remains technically challenging and time-consuming for many proteins. Here, we show that the protein structure prediction tools AlphaFold2-Multimer and AlphaFold3 can be used to quickly and accurately predict oligomeric states and structures for a range of soluble and membrane proteins. Across over 4700 proteins, AlphaFold2-Multimer provides reliable oligomeric state predictions in the majority of cases, however accuracy is more limited for proteins lacking close structural representatives in the AlphaFold training set, highlighting the dependence of these methods on robust training data. Together, our results suggest both the utility and current limitations of AlphaFold-based oligomeric state prediction, highlight cases where multiple physiologically relevant assemblies may be plausible, and provide practical guidance for minimizing computational cost, identifying challenging cases, and applying these methods to proteins lacking experimental structural data.