Erratum: Corrigendum: Network deconvolution as a general method to distinguish direct dependencies in networks

Soheil Feizi, Daniel Marbach, Muriel Médard, Manolis KellisNature Biotechnology 2015 · DOI · OpenAlex

Introduction

In scientific publishing, a corrigendum often signals a retraction or a weakening of previous claims. However, the correction for the 2013 Nature Biotechnology paper on "Network Deconvolution" (ND) presents a rare counter-example: the method is actually more powerful than initially reported. By addressing a numerical implementation error, the authors demonstrate that ND achieves up to an 87% improvement over existing tools in identifying direct biological interactions.

The Challenge of Indirect Dependencies

Complex systems—ranging from gene regulatory networks to social structures—are rife with correlations. A fundamental challenge in network analysis is distinguishing direct interactions from indirect dependencies. For instance, if Gene A regulates Gene B, and Gene B regulates Gene C, a standard correlation analysis might erroneously suggest a direct link between A and C. This "indirect noise" obscures the true underlying architecture of the system.

The original ND framework was designed to "filter out" these indirect effects using a matrix transformation. By treating the observed network as a sum of direct and indirect paths, an inverse operation can theoretically isolate the direct dependencies.

Methodology and the Scaling Correction

The core of the ND method involves a matrix transformation where the observed network is decomposed. The error identified in this corrigendum relates specifically to eigenvalue scaling. In the original implementation, the authors scaled the network by the maximum eigenvalue of the observed network. However, the underlying mathematical theory required scaling by the maximum eigenvalue of the direct network.

Correcting this implementation detail significantly enhanced the algorithm's performance across various biological and social datasets, highlighting how sensitive network algorithms can be to normalization parameters.

Key Results and Performance Gains

The re-evaluation of ND across three distinct domains revealed consistent performance boosts. In the DREAM5 benchmark for regulatory networks, the improvement of ND over existing methods like ARACNE was corrected from 75% to 87%. In social network applications, the correlation ($R^2$) with true collaboration strength remained high at 0.74.

Summary of reported vs. actual performance improvements across applications.

The visual impact of this correction is most evident in the protein folding application. The "discrimination power"—the ability to tell interacting residues from non-interacting ones—was corrected from 15% to 22%. This represents a sevenfold improvement over standard mutual information, rather than the fivefold improvement originally claimed.

Comparison of original and revised network deconvolution performance in protein folding.

Conclusion

This corrigendum reinforces ND as a robust, general-purpose method for network analysis. While the authors focused on fixing the historical record, the updated results suggest that the original 2013 paper significantly underestimated the excellence of the Network Deconvolution approach. The method remains a powerful "gold standard" for researchers seeking to map the true underlying architecture of complex systems.

Summary updated

BibTeX
@inproceedings{feizi2015erratum,
  author = {Soheil Feizi and Daniel Marbach and Muriel Médard and Manolis Kellis},
  title = {{Erratum: Corrigendum: Network deconvolution as a general method to distinguish direct dependencies in networks}},
  booktitle = {Nature Biotechnology},
  year = {2015},
  doi = {10.1038/nbt0415-424},
}
Figure