Author Correction: Improved prediction of protein-protein interactions using AlphaFold2

Patrick Bryant, Gabriele Pozzati, Arne ElofssonNature Communications 2022 · DOI · OpenAlex

Abstract

Predicting protein-protein interactions (PPIs) accurately is a cornerstone of structural biology. While AlphaFold2 has revolutionized monomeric structure prediction, assessing the quality of predicted multimeric complexes requires specialized metrics. This paper discusses the development and subsequent refinement of the pDockQ score, a quality assessment metric derived from a sigmoidal fit of AlphaFold2 confidence scores and interface contact density. We highlight a recent technical correction to the parameters of this scoring function, ensuring more precise validation of predicted protein interfaces.

Introduction

The release of AlphaFold2 marked a paradigm shift in computational biology, providing high-accuracy models for individual protein chains. However, the biological function of proteins often emerges through their interactions with other proteins. Extending AlphaFold2 to predict these complexes (protein-protein docking) introduced a new challenge: distinguishing between biologically relevant interfaces and structural "hallucinations"—predictions where the model forces two proteins together without a physical or evolutionary basis.

The pDockQ Metric

To address this, researchers developed pDockQ (predicted DockQ). Unlike the standard pLDDT (predicted Local Distance Difference Test), which measures local confidence in a single chain, pDockQ evaluates the interface between proteins.

The metric is calculated using a combination of two primary features:

  1. Average Interface pLDDT: The mean confidence score assigned by AlphaFold2 to residues participating in the interaction.
  2. Number of Interface Contacts: A count of the physical interactions between the two chains, representing the "packing" density of the interface.

These features are combined into a single input variable ($x$) and passed through a sigmoidal function to map the raw AlphaFold2 output to a value between 0 and 1, where higher values indicate a higher likelihood of a correct interaction.

Technical Correction and Parameter Refinement

In a recent author correction published in Nature Communications, the parameters for the pDockQ sigmoidal curve were updated. The scoring function follows the standard sigmoidal form:

$$pDockQ = \frac{L}{1 + e^{-k(x-x_0)}} + b$$

The correction provides optimized coefficients that more accurately reflect the relationship between AlphaFold2 outputs and the gold-standard DockQ metric. The updated parameters are:

  • Maximum value ($L$): 0.724
  • Midpoint ($x_0$): 152.611
  • Steepness ($k$): 0.052
  • Offset ($b$): 0.018

These values replace the initial estimates ($L=0.958, x_0=160.11, k=1, b=0.001$). The most significant change is in the steepness ($k$), which was adjusted from a very sharp transition ($k=1$) to a much smoother curve ($k=0.052$). This refinement allows for a more nuanced assessment of interface quality across a broader range of predicted complexes.

Implications for Interactome Research

The pDockQ score is instrumental for large-scale interactome screening. By applying a threshold (typically around 0.23), researchers can filter out low-confidence predictions, significantly reducing the false-positive rate in proteome-wide interaction studies.

While pDockQ is currently optimized for AlphaFold2, its success underscores the importance of context-aware confidence metrics. As structural biology moves toward modeling entire cellular environments, empirical scoring functions like pDockQ will remain essential for validating the complex web of protein interactions that drive life.

Conclusion

The refinement of the pDockQ parameters ensures that the scientific community has access to the most accurate tool for validating AlphaFold2 multimer predictions. This correction highlights the iterative nature of computational biology, where empirical models are continuously tuned against experimental data to provide more reliable insights into the molecular machinery of the cell.

Summary updated

Abstract

This score is created by fitting a sigmoidal curve (Fig.
BibTeX
@inproceedings{bryant2022author,
  author = {Patrick Bryant and Gabriele Pozzati and Arne Elofsson},
  title = {{Author Correction: Improved prediction of protein-protein interactions using AlphaFold2}},
  booktitle = {Nature Communications},
  year = {2022},
  doi = {10.1038/s41467-022-29480-5},
}
Figure