Abstract: The increasing availability of data has generated unprecedented prospects for network analyses in many biological fields, such as neuroscience, genomics, and ecology. A powerful statistical framework for estimating such networks is Gaussian graphical models, but standard estimators for the corresponding graphs are prone to large numbers of false discoveries. In this paper, we introduce a novel graph estimator based on knockoffs that imitate the partial correlation structures of unconnected nodes. We then show that this new estimator provides accurate control of the false discovery rate and yet large power.
Introduction
In biological research, systems ranging from gene-gene interactions to functional brain connectivity are frequently modeled as networks. Mathematically, these are often represented as Gaussian Graphical Models (GGMs), where an edge between two nodes signifies a partial correlation—a dependency that persists even after accounting for all other variables in the system.
The primary challenge in GGM estimation is inference. While popular estimators like the Graphical Lasso (GLASSO) or neighborhood selection are effective for data prediction, they are notoriously prone to "false discoveries"—identifying connections that do not exist in reality. Furthermore, these methods lack an intuitive mechanism for error control; researchers often find it impossible to tune parameters to achieve a specific target False Discovery Rate (FDR). This paper introduces KO (Knockoffs for GGMs), a framework designed to provide both mathematical guarantees and practical calibration for FDR control.
Methodology: The KO Framework
The KO strategy adapts the Knockoff framework to the context of partial correlations through a three-step process:
- Constructing Knockoffs: For every potential edge, the method generates a "knockoff" counterpart. These are sampled from a Student’s t-distribution to mimic the behavior of "null" edges (pairs of nodes with no actual connection).
- Establishing Test Statistics: The strength of the original sample partial correlation is compared against its knockoff version. A statistic $W_{ij}$ is defined such that it is positive if the original signal is stronger than the knockoff and negative otherwise.
- Data-Driven Thresholding: Since null edges are equally likely to have a stronger original signal or a stronger knockoff signal, the number of false discoveries can be estimated from the negative $W$ values. A threshold $t$ is then set to maintain the estimated FDR below a user-defined target $q$.
The authors also propose KO+, a more conservative variant that ensures exact FDR control.
Experimental Results
The KO method was validated using extensive simulations and real-world biological datasets, including fMRI brain connectivity, human microbiome data, and amphibian abundance records.
Reliable FDR Control
A critical advantage of the KO method is its calibration. In simulations across various sample sizes ($n$) and node counts ($p$), the actual FDR consistently tracked the target FDR, remaining on or below the diagonal. This level of predictable control is absent in competing methods.
Superior Statistical Power
Beyond error control, the KO method demonstrates superior "power"—the ability to correctly identify true edges. When compared to standard methods like GLASSO, neighborhood selection (MB), and others, KO consistently achieved higher power for any given FDR level.
Biological Insights
The practical utility of KO was demonstrated in several domains:
- Microbiome Analysis: KO revealed that non-smokers possess significantly more interconnected gut microbiomes than smokers ($p < 10^{-10}$), providing rigorous statistical evidence for previously qualitative observations.
- Neuroscience: In fMRI analysis, the method accurately identified that the strongest functional connections typically occur between regional counterparts in the left and right brain hemispheres.
Conclusion and Future Work
The KO framework represents a significant step forward in the reliable estimation of biological networks. By using "fake" correlations as a benchmark, it allows researchers to set meaningful error thresholds without sacrificing discovery power.
While the current framework assumes a sample size larger than the number of nodes ($n > p$), future work will focus on extending these theoretical guarantees to the high-dimensional case ($p \gg n$). Additionally, while the method is designed for Gaussian models, its robustness to non-normal biological distributions remains an area for ongoing research.
Summary updated