Since its release, AlphaFold has been the buzz of the computational biology community. It seems that every group in the protein science field is trying to apply the model in their respective areas of research. Already we are seeing numerous papers attempting to adapt the model to specific niche domains across a broad range of life sciences. In this blog post I summarise a recent paper’s use of the technology for predicting protein-protein interfaces.

Interactions between proteins form the basis of most function within our cells. From enzymes in metabolic pathways to cellular signalling, protein-protein interactions underly these complex biological processes. Understanding these interaction sites, know as protein-protein interfaces (PPIs), has been a sought-after goal for many scientists. These interfaces have been studied experimentally in by various methods, generating large datasets containing functional data about what proteins interact with one another in the body, their stability, and other important properties. Despite this, scientists are limited by the availability of 3D structures depicting these interactions. Structural data enables precise identification of key residues governing these interactions and empowers researchers to detect things like disease associated mutations in the body. However, since the ground-breaking work of AlphaFold2 was showcased to the world in CASP14, computational methods are rapidly filling the gap left by the resources intensive nature of resolving 3D structures experimentally. Thus, in a recent publication in nature structural & molecular biology titled “Towards a structurally resolved human protein interaction network”¹ a group of researchers set out to test AlphaFold’s ability to predict PPIs.

The group used a slightly modified version of AlphaFold known as FoldDock, and ran the model on 65,000 protein-protein interfaces gathered from two functional datasets. They then ranked these models using an estimated quality metric taken from the confidence scores associated with AlphaFold predictions. This prediction and ranking left them with a repertoire of PPI structure models grouped by confidence. They validated these confidence groupings through an experimental assay involving the linking of residues across the protein-protein interface. These experiments showed a good correlation between experimentally and computationally identified residues involved in the interaction in the high confidence protein-protein interface predictions. The authors then undertook a case study to see if the predicted models were powerful enough to discern residue mutations that would lead to the disruption of the PPI. Their conclusion from this was that only the high confidence predicted models were capable of discerning disrupting mutations. The group also examined several other aspects of PPIs including phosphorylation sites and higher-order PPIs where multiple proteins were involved.

The use of the word “Toward” in the title of a scientific article is always a good indicator of where to set expectations for the work. Although the article was an interesting and comprehensive read, there are still many questions in the protein-protein interface field that remain unanswered. The authors of the article did an excellent job of highlighting the limitations of AlphaFold in this space. It became apparent that mainly the high confidence predictions- capturing high-affinity stable protein interactions- were useful in understanding deeper questions in the implications of these PPIs.

1. Burke, D. F. et al. Towards a structurally resolved human protein interaction network. Nat. Struct. Mol. Biol. 1–10 (2023) doi:10.1038/s41594-022-00910-8.