AlphaFold2 and its limits
Mapping protein-protein interaction networks is essential for understanding the dynamics of cellular functions and their cross-regulation. Precise knowledge of interaction sites makes it possible to specifically perturb the proteins in these networks and understand the synergies and competitions that ensure cell function.
Unfortunately, a great amount of structural information is still lacking to provide a detailed understanding of the organisation of interaction networks. The AlphaFold2 artificial intelligence programme has demonstrated a remarkable ability to predict the structures of protein assemblies that have co-evolved over long time scales. Its performance remained poorly characterised for assemblies involving intrinsically disordered regions, which often mediate transient and dynamic interactions during evolution.
Protein fragmentation strategy
In a study published in the journal Nature Communications, researchers from the AMIG team at the I2BC department have shown that AlphaFold2 performs poorly if large disordered regions are used directly for prediction (40% success rate). A protein fragmentation strategy was found to be particularly well adapted to predicting the interfaces between folded domains and small protein motifs that fold on contact with the partner. Applied on a large scale on more than 900 complexes, this strategy achieved a success rate of almost 90%, a very encouraging result for the systematic screening of protein interaction networks. Nevertheless, the study calls for vigilance with regard to the risks of detecting false positives, which will be at the heart of future developments in artificial intelligence strategies such as AlphaFold2.
ALL OF THis work HAs BENEFITED FROM ACCESS TO GENCI'S HPC RESOURCES (
Jean Zay supercomputer).
Contacts Institut CEA-Joliot: