Beyond Pathway Boundaries: A Degree-Aware Network Clustering Test for Gene Sets

Technical Summary: Beyond Pathway Boundaries: A Degree-Aware Network Clustering Test for Gene Sets

Problem Statement
Current interpretation of gene lists relies heavily on Over-Representation Analysis (ORA), a method plagued by three fundamental limitations: the assumption of fixed pathway boundaries, the incorrect assumption of gene independence, and a heavy dependence on the choice of the background set. While network-based methods attempt to address these issues by leveraging interaction-network modularity, they introduce a new critical flaw: hub bias. In curated biological networks, highly connected genes (hubs) are often overrepresented due to being well-studied. Under naive null models, these hubs appear artificially clustered, leading to spurious significance. Existing correction strategies are insufficient; edge permutation destroys the very topology the test aims to condition upon, while propagation methods often obscure the confounding factor within parameter tuning.

Methodology: MANGO
The authors introduce MANGO (Moran's Autocorrelation for Network Gene Over-representation), a statistical framework designed to answer a specific conditional question: Does a gene set's spatial autocorrelation on a fixed biological network exceed what its degree composition alone would predict?

The core of MANGO involves computing Global Moran's I under a rigorous null model that conditions on two factors:

1. The fixed structure of the biological network.
2. The binned degree distribution of the specific gene set being tested.

By stratifying genes into degree bins (e.g., a ten-bin approach), MANGO generates a null distribution that accounts for the inherent connectivity of the genes in the set, effectively neutralizing hub bias. The method further decomposes significant signals to the component and gene levels, allowing for the localization of clustering within specific network modules.

Key Results
Benchmarking and simulation studies demonstrate the efficacy of the degree-stratified approach:

• False Positive Control: Uniform null models produce a false positive rate of 1.0 when testing hub-enriched gene sets that lack genuine biological clustering. In contrast, the ten-bin degree-stratified null reduces this rate to 0.0 without sacrificing statistical power.
• Power Preservation: For degree-typical signals, the Area Under the Curve (AUC) remains high ($\ge$ 0.98). The difference in AUC ($|\Delta \text{AUC}|$) between the proposed method and ideal detection on degree-typical signals is negligible ($\le$ 0.004).
• Simulation Validation: Pathway-spiking simulations confirm that MANGO can detect real biological clustering across diverse pathway sizes and degree profiles.
• Real-World Application: Applied to the FIGI colorectal cancer GWAS (204 SNPs), the gene set was found to be degree-typical (KS $p = 0.83$), yet exhibited highly significant spatial autocorrelation ($p < 0.001$). A component-level jackknife analysis localized the entire signal to a single 24-gene module. This module spans TGF-$\beta$, Wnt/cadherin, and related pathways, containing four identified bottlenecks (SMAD3, MYC, CTNNB1, PTPN1) that align with established colorectal cancer driver biology.

Significance and Claims
The paper positions MANGO as a necessary evolution in gene set analysis that resolves the tension between network topology and statistical validity. Its primary contribution is the demonstration that degree composition is a critical confounder in network-based gene set testing. By conditioning the null model on the degree distribution, MANGO eliminates the artifact of hub-driven clustering while preserving the ability to detect genuine biological signal. The method provides a robust, topology-preserving alternative to ORA and flawed network-based corrections, offering a more accurate lens for interpreting gene lists in the context of complex biological networks.