Statistical Model Selection Methods Applied to Biological Networks

Michael P. H. Stumpf, Piers J. Ingram, Ian Nouvel, Carsten WiufTrans. Comp. Sys. Biology 2005 · DOI

TL;DR

This paper argues that the common assumption of biological networks being "scale-free" is often an oversimplification. By applying rigorous statistical model selection techniques, the authors demonstrate that other fat-tailed distributions, like the log-normal and stretched exponential, provide a significantly better fit to real protein interaction and metabolic network data. This suggests a need for new theoretical models that can explain the observed network structures.

What Problem Does This Paper Address?

Biological networks, such as protein-protein interaction (PPI) networks and metabolic networks (MNs), are fundamental to understanding cellular processes. For years, a prevailing idea in the field has been that these networks exhibit a "scale-free" property, meaning their degree distribution (the probability of a node having a certain number of connections) follows a power-law. This assumption is attractive because simple network growth models can generate power-law distributions. However, fitting a power-law to empirical data has often been done in an ad-hoc manner, potentially overlooking other important characteristics of the network's structure. The authors aim to address this by using formal statistical methods to rigorously determine which functional form best describes the degree distributions of real biological networks, moving beyond visual inspection or simplified fitting procedures.

Methods

The core of this paper lies in its application of formal statistical model selection techniques. The authors propose using:

  • Maximum Likelihood Estimation (MLE): This standard statistical technique is used to estimate the parameters of various proposed probability models (e.g., Poisson, exponential, gamma, power-law, log-normal, stretched exponential) that best fit the observed degree distribution data. They employ a composite likelihood approach, treating node degrees as independent observations, which is appropriate for large networks.
  • Akaike Information Criterion (AIC): Since the models being compared are often non-nested, the AIC is used as a criterion to select the best model. AIC balances model fit with model complexity, penalizing models with more parameters. The authors calculate AIC values and then derive "Akaike weights" ($w_j$), which represent the probability that a given model is the best among the set of considered models.
  • Goodness-of-Fit Tests: To further assess how well a model describes the data, the Kolmogorov-Smirnoff (KS) and Anderson-Darling (AD) statistics are employed. These statistics quantify the difference between the empirical cumulative distribution function and the theoretical cumulative distribution function of a model. The AD statistic is particularly useful as it gives more weight to the tails of the distribution, which is relevant for fat-tailed distributions.

Datasets

The study analyzes real biological network data from several sources:

  • Protein Interaction Networks (PINs): Data was collected from five model organisms: D. melanogaster, C. elegans, S. cerevisiae, H. pylori, and E. coli. The protein interaction data was sourced from the DIP database.
  • Metabolic Networks (MNs): Data was obtained from the KEGG database, with additional information from the BRENDA database. Nodes in these networks represent enzymes, and an edge connects two enzymes if one uses the product of the other as an educt. Human and yeast metabolic network data were extracted from the global KEGG database.

Key Insights and Ideas

  • Scale-Free is Often Not the Best Fit: The most striking finding is that simple scale-free (power-law) models, even with finite-size corrections, frequently do not provide the best description of biological network degree distributions. This challenges the widespread assumption in the literature.
  • Fat-Tailed Distributions Dominate: While scale-free models are fat-tailed, other fat-tailed distributions, specifically the log-normal and stretched exponential, often offer superior fits to the empirical data. This suggests that the "fat-tailed" nature is important, but the precise functional form matters.
  • Model Selection is Crucial: The paper highlights the power of rigorous statistical model selection (like AIC) over visual inspection of log-log plots, which can be misleading and lead to ad-hoc fitting. Using all the data and a principled approach reveals nuances that simpler methods miss.
  • Need for New Theoretical Models: The fact that the best-fitting distributions (log-normal, stretched exponential) are not easily explained by simple, common network growth models suggests a significant gap in our theoretical understanding of how these complex biological networks evolve.

Key Results

  • For protein interaction networks, the log-normal and stretched exponential models consistently outperformed the standard scale-free model across most of the studied organisms. In C. elegans, the scale-free model was only marginally better than the log-normal.
  • For metabolic networks, the stretched exponential model provided a better fit than the scale-free model, and the log-normal model also performed well, outperforming the stretched exponential in terms of goodness-of-fit statistics (D and D*).
  • The paper provides quantitative measures (log-likelihoods, AIC values, Akaike weights, KS/AD statistics) that support these conclusions, demonstrating a clear preference for non-power-law fat-tailed distributions in many cases. For example, in S. cerevisiae PIN data, the Akaike weight for the log-normal model (M5) was $w_5 \approx 0.95$.

Limitations and Future Work

  • Degree Distribution is Not Everything: The authors acknowledge that the degree distribution, while important, does not capture all aspects of a network's structure. Other topological properties (e.g., clustering coefficients, path lengths) also play a role in understanding network function.
  • Finite Size and Sampling: Real biological networks are finite and the available data often represents only a sample of the complete network. The authors note that sampling properties can significantly affect the observed degree distribution, especially for scale-free networks. Future work should consider these mesoscopic and sampling effects more deeply.
  • Lack of Mechanistic Models: A significant limitation is the absence of well-established theoretical models that naturally produce log-normal or stretched exponential degree distributions in biological contexts. The authors emphasize the need for developing new models that incorporate biological constraints like modularity, cellular compartments, and evolutionary pressures.
  • Dynamic Networks: Biological networks are not static and change over time. The current analysis focuses on static snapshots, and future research should address the dynamics of network evolution.

The authors also mention that their statistical tools for network analysis are implemented in the NetZ R-package, which is available upon request, facilitating further research in this area.

Summary updated

BibTeX
@article{stumpf2005statistical,
  author = {Michael P. H. Stumpf and Piers J. Ingram and Ian Nouvel and Carsten Wiuf},
  title = {{Statistical Model Selection Methods Applied to Biological Networks}},
  journal = {Trans. Comp. Sys. Biology},
  year = {2005},
  doi = {10.1007/11599128\_5},
}
Figure