Discovering Graphical Granger Causality Using the Truncating Lasso Penalty

Abstract: Components of biological systems interact with each other in order to carry out vital cell functions. Such information can be used to improve estimation and inference, and to obtain better insights into the underlying cellular mechanisms. Discovering regulatory interactions among genes is therefore an important problem in systems biology. Whole-genome expression data over time provides an opportunity to determine how the expression levels of genes are affected by changes in transcription levels of other genes, and can therefore be used to discover regulatory interactions among genes. In this paper, we propose a novel penalization method, called truncating lasso, for estimation of causal relationships from time-course gene expression data. The proposed penalty can correctly determine the order of the underlying time series, and improves the performance of the lasso-type estimators. Moreover, the resulting estimate provides information on the time lag between activation of transcription factors and their effects on regulated genes. We provide an efficient algorithm for estimation of model parameters, and show that the proposed method can consistently discover causal relationships in the large $p$, small $n$ setting. The performance of the proposed model is evaluated favorably in simulated, as well as real, data examples. The proposed truncating lasso method is implemented in the R-package grangerTlasso and is available at http://www.stat.lsa.umich.edu/~shojaie.

• The paper introduces the truncating lasso penalty to estimate causal Granger relationships in high-dimensional gene regulatory networks.
• It automatically determines the VAR model order by truncating redundant covariates, reducing model complexity and boosting estimation accuracy.
• Empirical results show reduced false positives and improved recall with enhanced computational efficiency compared to traditional lasso approaches.

Discovering Graphical Granger Causality Using the Truncating Lasso Penalty

The paper authored by Ali Shojaie and George Michailidis addresses a pressing challenge in the field of systems biology: the discovery of causal relationships within gene regulatory networks from high-dimensional time-course gene expression data. This is a non-trivial task, particularly when faced with the "large p, small n" problem, where the number of genes (variables) significantly exceeds the number of temporal observations (samples).

Theoretical Contributions

The main contribution of the paper is the introduction of the truncating lasso penalty, a novel method specifically designed to estimate causal Granger relationships effectively in high-dimensional settings. The truncating lasso offers two unique features:

1. Order Determination and Simplification: It autonomically identifies the order of the vector autoregressive (VAR) model, simplifying the model by truncating unnecessary covariates based on the learned order, thus effectively controlling model complexity.
2. Improved Estimation: By focusing on reducing model complexity, it enhances the estimation accuracy of conventional lasso-type approaches.

The development is grounded in the context of graphical Granger causality, a theoretical framework where causality is inferred from the predictability dynamics between time series. Here, the truncating lasso exploits temporal structures to ensure the correct temporal lag and causal links are learned.

Numerical and Empirical Results

The authors demonstrate through simulation studies that the truncating lasso outperforms traditional lasso and adaptive lasso in accurately uncovering the gene regulatory network structure:

• Performance Metrics: The analysis measures included the structural Hamming distance (SHD), F1 score, and ROC plots, providing robust evidence that the truncating lasso achieves lower false positive rates and higher recall, particularly when the sample size is restricted.
• Computational Efficiency: The iterative algorithm implemented showcases efficiency in converging to a reliable solution, showcasing linear convergence and reduced computational cost relative to the non-partitioned alternatives applied in high-dimensional VAR contexts.

Beyond simulations, the method is validated on real datasets, including known gene regulatory networks such as those derived from E-coli and HeLa cell lines. In practical scenarios, the method yielded insightful results, aligning closely with known biological truths and sometimes indicating new biological hypotheses where standard methods faltered.

Practical Implications

From a practical perspective, the truncating lasso's ability to discern temporal lag information introduces a granularity not afforded by existing methods like group lasso, which aggregates lag effects, potentially ignoring important temporal dynamics. The ease of implementation within the grangerTlasso R-package simplifies its adoption for large scale genomic studies aiming to unearth complex gene interactions foundational in cellular phenotypes.

Future Directions

The truncating lasso opens pathways to further exploration in high-dimensional causal inference. One potential avenue is its application and adaptation to other forms of biological data where causality, sparsity, and temporal dynamics are of interest, such as omics studies and time-varying networks. Additionally, integration with machine learning frameworks could refine the scalability of causal discoveries in diverse datasets.

In conclusion, this work represents a significant methodological advancement in the field of high-dimensional causal inference, providing researchers within computational biology an effective tool to decode the gene regulatory architectures essential to understanding complex phenotypic outcomes.

For detailed implementations and in-depth theoretical discussions, interested researchers are encouraged to explore the R-package grangerTlasso, which operationalizes the truncating lasso method proposed in the study.