Prior knowledge driven Granger causality analysis on gene regulatory network discovery

Shun Yao, Shinjae Yoo, Dantong YuBMC Bioinformatics 2015 · OpenReview

TL;DR

This paper introduces CGC-2SPR, a novel method for discovering gene regulatory networks from time-series gene expression data, especially when the number of genes far exceeds the number of time points (n >> T). By integrating prior biological knowledge with a modified Granger causality approach, CGC-2SPR significantly improves accuracy and reduces false positives compared to existing methods, offering a more robust way to understand complex biological systems.

What Problem Does This Paper Address?

Discovering gene regulatory networks (GRNs) is crucial for understanding how genes interact within a cell. Time-series gene expression data, which captures gene activity over time, is a valuable resource for this. However, a common challenge in biological datasets is the "n >> T" problem: there are many more genes (n) than available time points (T) in the data. Traditional methods like pairwise Granger causality (PGC) and even more advanced conditional Granger causality (CGC) with standard regularization techniques struggle with this imbalance, leading to a high number of false positives and inaccurate network predictions. This is because limited time points make it difficult to reliably distinguish direct causal relationships from indirect ones or random noise.

Methods

The core of this paper is the CGC-2SPR (Conditional Granger Causality using two-step prior Ridge regularization) method. It builds upon the conditional Granger causality (CGC) framework but incorporates prior biological knowledge in a novel way.

Here's a breakdown:

  • Granger Causality (GC): The foundational concept is that if past values of one time series (gene A) help predict the future values of another time series (gene B), then gene A "Granger-causes" gene B.
  • Conditional GC (CGC): This extends pairwise GC to consider multiple genes simultaneously, aiming to better distinguish direct from indirect effects.
  • Ridge/Lasso Regularization: To handle the "n >> T" problem and prevent overfitting, standard regularization techniques like Ridge (penalizing the sum of squared coefficients) and Lasso (penalizing the sum of absolute coefficients) are applied to the regression problem.
  • Prior Biological Knowledge: The key innovation is integrating prior knowledge (e.g., from databases like YeastNet or TF binding profiles) into the regularization process. This knowledge is represented as a weight matrix (W).
  • Two-Step Prior Ridge Regularization: Instead of naively applying prior knowledge, CGC-2SPR uses a two-step approach. First, a standard Ridge regression is performed to get an initial estimate of the coefficients (B*). The sign of these coefficients is then used to refine the prior knowledge weight matrix (W*). This refined matrix is then used in a second Ridge regression step, allowing the model to distinguish between positive and negative regulatory relationships more effectively.
  • Monte Carlo Significance Estimation (MCSE): To address the issue of standard statistical tests (like F-tests) failing under the n >> T condition, the authors propose MCSE. This method approximates the distribution of regression coefficients to estimate the statistical significance of discovered edges.

Datasets

  • Simulated Dataset: A custom simulation dataset was generated to mimic real biological data characteristics, specifically the n >> T problem and hierarchical modular structures found in GRNs. This dataset was generated using a linear model based on a simulated network with 1000 nodes and 1082 regulatory edges, using only 20 time points.
  • Real Biological Dataset: The "yeast metabolic cycle" (YMC) dataset was used. This dataset is characterized by n = 2935 genes and T = 36 time points. The authors focused on a subset of 2935 periodic genes.

Key Insights and Ideas

  • "1 + 1 > 2" Effect: The paper strongly emphasizes that combining prior biological knowledge with gene expression data yields a synergistic effect, leading to significantly better GRN inference than either source alone.
  • Prior Knowledge is Crucial for n >> T: The "n >> T" problem severely degrades the performance of standard methods. Prior knowledge acts as a crucial constraint, guiding the inference process and reducing false positives.
  • Two-Step Prior Integration Matters: Simply adding prior knowledge to regularization isn't enough. The two-step approach, which uses initial regression results to inform the sign of prior knowledge weights, is critical for correctly distinguishing positive and negative regulations.
  • MCSE for Reliable Significance: The proposed MCSE method provides a robust way to assess the statistical significance of inferred regulatory relationships, overcoming the limitations of traditional tests in high-dimensional, low-time-point scenarios.
  • TF Binding Score Superiority: For the yeast dataset, prior knowledge derived from TF binding scores proved more effective than general functional association networks (YeastNet), suggesting that more specific, direct regulatory information is more beneficial.

Key Results

  • Simulation Experiments: CGC-2SPR demonstrated significant performance improvements over PGC, Ridge, Lasso, and information-theoretic methods (MRNET, ARACNE) in terms of accuracy (F₁ score). For instance, CGC-2SPR achieved an F₁ score of 0.150, substantially outperforming Lasso (0.091) and Ridge (0.046).
  • Real Data (YMC): CGC-2SPR, particularly when using TF binding score prior knowledge, identified a significantly larger number of true positive edges compared to other methods. Table 3 shows CGC-2SPR (TFBS) identifying 109 golden standard edges out of the top 10000 entries, compared to Lasso (20) and Ridge (7).
  • Computational Efficiency: CGC-2SPR, being based on Ridge regression, is computationally efficient and scalable, making it suitable for large-scale GRN inference. It processed the simulation dataset in approximately 1 hour (with parallelization), comparable to Ridge and significantly faster than Lasso or Enet which took much longer due to cross-validation.
  • Functional Prediction: The method identified a potential regulatory relationship for the previously uncharacterized gene ABM1, suggesting it might be involved in yeast's response to different glucose levels.

Limitations and Future Work

  • Prior Knowledge Selection: The effectiveness of CGC-2SPR is dependent on the quality and relevance of the prior knowledge. The paper notes that "closely relevant prior knowledge" is preferred. Future work could explore more sophisticated methods for selecting and integrating diverse prior knowledge sources.
  • Model Order Selection: Determining the optimal model order (p) for Granger causality is computationally expensive (via cross-validation) and can impact results. The authors suggest exploring more efficient methods for selecting 'p' in the future.
  • Completeness of Golden Standards: Real biological datasets often have incomplete or indirect golden standards, meaning the performance metrics might underestimate the true capabilities of the methods.
  • Lasso/Elastic Net with Prior Knowledge: The paper poses an interesting question about whether Lasso or Elastic Net, when combined with prior knowledge, could outperform CGC-2SPR, though notes their slower convergence properties. The authors mention their two-step algorithm is directly applicable to Lasso and Elastic Net.
  • Experimental Design: The authors highlight the importance of carefully designed time-series experiments, suggesting that covering multiple biological cycles and having sufficient time points are crucial for successful network inference.

Summary updated

Abstract

Background Our study focuses on discovering gene regulatory networks from time series gene expression data using the Granger causality (GC) model. However, the number of available time points (T) usually is much smaller than the number of target genes (n) in biological datasets. The widely applied pairwise GC model (PGC) and other regularization strategies can lead to a significant number of false identifications when n>>T. Results In this study, we proposed a new method, viz., CGC-2SPR (CGC using two-step prior Ridge regularization) to resolve the problem by incorporating prior biological knowledge about a target gene data set. In our simulation experiments, the propose new methodology CGC-2SPR showed significant performance improvement in terms of accuracy over other widely used GC modeling (PGC, Ridge and Lasso) and MI-based (MRNET and ARACNE) methods. In addition, we applied CGC-2SPR to a real biological dataset, i.e., the yeast metabolic cycle, and discovered more true positive edges with CGC-2SPR than with the other existing methods. Conclusions In our research, we noticed a “ 1+1>2” effect when we combined prior knowledge and gene expression data to discover regulatory networks. Based on causality networks, we made a functional prediction that the Abm1 gene (its functions previously were unknown) might be related to the yeast’s responses to different levels of glucose. Our research improves causality modeling by combining heterogeneous knowledge, which is well aligned with the future direction in system biology. Furthermore, we proposed a method of Monte Carlo significance estimation (MCSE) to calculate the edge significances which provide statistical meanings to the discovered causality networks. All of our data and source codes will be available under the link https://bitbucket.org/dtyu/granger-causality/wiki/Home .
BibTeX
@article{yao2015prior,
  author = {Shun Yao and Shinjae Yoo and Dantong Yu},
  title = {{Prior knowledge driven Granger causality analysis on gene regulatory network discovery}},
  journal = {BMC Bioinformatics},
  year = {2015},
  url = {https://openreview.net/forum?id=T-Civ1sdQ1L},
}
Figure