Gene regulatory network discovery using pairwise Granger causality.

Area of Science:

• Computational Biology
• Systems Biology
• Bioinformatics

Background:

• Gene expression data analysis is crucial for understanding biological processes and drug development.
• Granger causality (GC) is a powerful statistical method for inferring gene regulatory networks from time-series data.
• Pairwise GC application to high-dimensional gene expression data can lead to spurious causalities due to limitations in model scope and order selection.

Purpose of the Study:

• To investigate the emergence of false discoveries in pairwise Granger causality detection for gene regulatory networks.
• To evaluate the effectiveness of model validation techniques in mitigating spurious causalities.
• To apply validated pairwise GC to a human HeLa cell-cycle dataset and analyze network properties.

Main Methods:

• Simulated time-series gene expression data were used to assess pairwise versus full-model GC.
• Model validation techniques, including Akaike Information Criterion (AIC), were employed to select appropriate vector autoregressive model orders.
• Validated pairwise GC was applied to the human HeLa cell-cycle dataset to construct a gene regulatory network.

Main Results:

• Pairwise GC analysis, especially with insufficient model order, significantly increases false discoveries compared to full-model approaches.
• Model validation techniques effectively reduce the number of spurious causalities.
• Analysis of the HeLa cell-cycle dataset revealed network hubs predominantly act as sources of regulatory interactions, a novel observation.

Conclusions:

• Model validation is essential for accurate gene regulatory network inference using pairwise Granger causality.
• Akaike Information Criterion is a suitable model order selection tool, though caution is needed for very short time series.
• The identified network hubs' role as interaction sources provides new insights into cell-cycle regulation.