DNLC: differential network local consistency analysis

Jianwei Lu, Yao Lu, Yusheng Ding, Qingyang Xiao, Linqing Liu, Qingpo Cai, Yunchuan Kong, Yun Bai, Tianwei YuBMC Bioinformatics 2019 · DOI

TL;DR

This paper introduces DNLC, a novel computational method to identify genes and network modules whose local expression consistency changes significantly between different biological conditions. By adapting spatial statistics concepts to biological networks, DNLC offers a powerful complement to traditional differential expression analysis, enabling deeper insights into dynamic biological processes and disease mechanisms.

What Problem Does This Paper Address?

Biological systems are inherently dynamic, with gene and protein interactions constantly being activated or deactivated based on cellular conditions. While existing methods often analyze static biological networks or focus on individual gene expression changes, they struggle to capture how the consistency of gene expression within localized network regions (e.g., subnetworks) changes across different conditions (like disease states or treatment responses). This gap limits our understanding of the regulatory mechanisms underlying these dynamic biological shifts. The authors aim to address this by developing a method to pinpoint genes and modules where this local expression consistency is altered, providing a more nuanced view of biological regulation.

Methods

The core of the DNLC method lies in adapting the concept of Local Moran's I (LMI), a measure from spatial statistics used to quantify spatial autocorrelation, to biological networks.

  • Network Representation: Genes are treated as nodes in a biological network, and their expression levels are analyzed.
  • Local Moran's I (LMI) Calculation: For each gene (node) and in each sample, LMI is calculated. This quantifies how similar a gene's expression is to the expression of its immediate neighbors on the network. A high positive LMI indicates similar expression levels (high or low) among neighbors, while a large negative LMI suggests the node is an "outlier" with dissimilar expression.
  • Weighting Scheme: To focus on local consistency, a truncated Gaussian function is used to assign weights to neighboring nodes based on their shortest path distance on the network. Nodes within a distance of 2 receive a weight, while those farther away are ignored.
  • Differential Consistency (DC) Gene Selection: After calculating LMI for all genes across all samples, the method uses a generalized linear model (logistic regression for binary outcomes, Cox proportional hazards for survival) to link a gene's LMI score, the clinical outcome, and potential confounders (like age or sex). The t-statistics from this model are then analyzed using a local false discovery rate (lfdr) procedure to identify genes whose LMI scores change significantly between conditions.
  • Network Community Detection: Once DC genes are identified, a fast-greedy algorithm is used to group these genes and their direct neighbors into network communities. These communities are then analyzed using Gene Ontology (GO) enrichment to understand their biological functions.

Datasets

The paper utilizes two publicly available datasets for evaluation:

  • GSE10255: A gene expression dataset of acute lymphoblastic leukemia (ALL) subjects treated with methotrexate (MTX). This dataset contains 12,704 genes and 161 samples, with the outcome variable being the reduction of circulating leukemia cells post-MTX treatment.
  • TCGA BRCA: A breast cancer gene expression dataset from The Cancer Genome Atlas (TCGA). For this dataset, a Cox proportional hazards model was used to link gene LMI values with patient survival outcomes, adjusting for demographic variables.

The authors also generated simulated data to systematically evaluate the method's performance under controlled conditions, varying parameters like base correlation, changed correlation, node degrees, and sample size.

Key Insights and Ideas

  • Bridging Spatial Statistics and Genomics: The innovative application of Local Moran's I from spatial statistics to gene expression networks is a key conceptual leap, allowing for the quantification of local expression "consistency" in a biologically meaningful way.
  • Capturing Dynamic Network Behavior: DNLC moves beyond static network analysis by focusing on changes in local expression patterns, which is crucial for understanding dynamic biological processes and responses to stimuli or disease.
  • Complementary to Differential Expression: The method is presented as a valuable addition to traditional differential expression analysis, offering a network-centric perspective that can uncover regulatory mechanisms missed by gene-by-gene approaches.
  • Identification of Functional Modules: By integrating DC gene identification with network community detection and GO enrichment, the method can pinpoint biologically relevant subnetworks that are dynamically altered.

Key Results

  • Simulation Performance: In simulations, DNLC effectively detected artificially induced changes in local expression consistency. Its performance (measured by PR-AUC) generally improved with sample size and was robust across various correlation levels, outperforming simple differential expression (DE) analysis and showing advantages over differential network analysis (DNA) in most scenarios.
  • GSE10255 (ALL/MTX): The method identified 510 DC genes at an lfdr threshold of 0.2. Functional analysis of the largest module revealed that positive DC genes (increased consistency with MTX response) were enriched in processes like phosphorylation, peptide hormone secretion, and apoptosis, all relevant to MTX sensitivity. Negative DC genes were linked to immune cell development, apoptosis, and cell proliferation, also relevant to ALL progression.
  • TCGA BRCA (Breast Cancer): In the breast cancer dataset, DNLC identified modules where genes with altered local consistency were associated with patient survival. Positive DC genes in one module were linked to protein/peptide metabolic processes and membrane transport, while negative DC genes in another module were associated with the I-Kappa B kinase signaling pathway and genes regulating cell proliferation and survival.
  • Computational Efficiency: The computing time of the DNLC method scales roughly linearly with the number of genes and samples, making it feasible for large-scale genomic datasets (e.g., 10,000 genes and 500 samples take about 12 minutes on a single thread).

Limitations and Future Work

  • Network Dependency: The method relies on the availability of a pre-existing biological network. The quality and completeness of this network directly impact the results.
  • Focus on Ego-Networks: While efficient, the current approach focuses on ego-networks (a node and its immediate neighbors). Future work could explore more complex subnetwork structures.
  • Generalizability: The authors demonstrated the method's utility across different outcome types (binary and survival) and disease contexts (leukemia and breast cancer), suggesting broad applicability. However, further validation on diverse datasets and biological questions would be beneficial.
  • Interpretation of "Consistency": While the method identifies changes in expression consistency, the precise biological interpretation of "increased" or "decreased" consistency in specific contexts might require further domain expertise.

The paper concludes that DNLC is an effective tool for finding modules with changing gene expression consistency, complementing traditional analyses and facilitating new discoveries from gene expression data.

Summary updated

BibTeX
@article{lu2019dnlc,
  author = {Jianwei Lu and Yao Lu and Yusheng Ding and Qingyang Xiao and Linqing Liu and Qingpo Cai and Yunchuan Kong and Yun Bai and Tianwei Yu},
  title = {{DNLC: differential network local consistency analysis}},
  journal = {BMC Bioinformatics},
  year = {2019},
  doi = {10.1186/s12859-019-3046-4},
}
Figure