HTRgene: a computational method to perform the integrated analysis of multiple heterogeneous time-series data: case analysis of cold and heat stress response signaling genes in Arabidopsis

Hongryul Ahn, Inuk Jung, Heejoon Chae, Dongwon Kang, Woosuk Jung, Sun KimBMC Bioinformatics 2019 · DOI

TL;DR

This paper introduces HTRgene, a novel computational method for integrating multiple, heterogeneous time-series gene expression datasets. HTRgene identifies "response order preserving DEGs" by analyzing how gene expression changes over time across different experimental conditions, offering a more robust way to discover stress response genes.

What Problem Does This Paper Address?

Understanding how organisms respond to environmental stresses like cold and heat is crucial for fields like agriculture and medicine. Gene expression data, particularly time-series data, offers a powerful window into these responses. However, analyzing multiple such datasets simultaneously presents a significant challenge. Datasets often come from different experiments with varying conditions (e.g., stress intensity, time points, tissue types, ages), making direct comparison difficult. Existing methods for identifying differentially expressed genes (DEGs) typically focus on individual datasets or simple pairwise comparisons, failing to leverage the full potential of integrated, heterogeneous data. This leads to inconsistent results and a missed opportunity to uncover more reliable stress response mechanisms. The core problem is how to integrate these diverse datasets to find genes that consistently respond to stress in a predictable temporal order.

Methods

HTRgene tackles the integration challenge through a four-step process:

  1. Normalization and Consensus DEG Detection: Raw gene expression data from various sources (microarray, RNA-Seq) are first normalized. Then, genes that are consistently differentially expressed across multiple samples (consensus DEGs) are identified. This step filters out genes that only show a response in a single, potentially noisy, experiment.
  2. Co-expression-based Clustering: To manage the complexity of analyzing individual genes, HTRgene clusters consensus DEGs based on their co-expression patterns. This reduces the problem from gene-level analysis to cluster-level analysis.
  3. Response Time Vector Determination: For each gene cluster, HTRgene determines a "response time vector." This vector represents the time point at which a gene's expression statistically changes before and after a specific point. This is achieved using a greedy, hill-climbing approach to find an optimal response time that maximizes the separation score within the cluster.
  4. Ordering Gene Clusters: Finally, the gene clusters are ordered based on their response time vectors. This creates a "response schedule," identifying genes that not only respond to stress but do so in a consistent temporal order across different datasets. The output is a set of "response order preserving DEGs" and their associated "response phases."

Datasets

The HTRgene method was validated using two sets of heterogeneous time-series gene expression data for Arabidopsis thaliana:

  • Cold Stress: 28 time-series sample datasets.
  • Heat Stress: 24 time-series sample datasets.

These datasets were collected from public repositories like GEO and ArrayExpress. The paper does not introduce new benchmark datasets but rather uses existing, publicly available, heterogeneous data to demonstrate the utility of HTRgene.

Key Insights and Ideas

  • Response Order Preservation is Key: The central idea is that while the exact timing of a stress response might vary across experiments, the order in which genes respond to each other is likely conserved. This insight is crucial for integrating heterogeneous data.
  • Clustering for Scalability: Analyzing millions of genes individually across multiple datasets is computationally prohibitive. Clustering genes with similar expression patterns first dramatically reduces the complexity, making the analysis feasible.
  • From Genes to Clusters: Shifting the focus from individual gene response times to cluster response times allows for a more robust estimation, mitigating noise from individual gene measurements.
  • Biological Mechanisms Revealed: The method successfully reproduced known biological pathways for cold and heat stress, validating its ability to uncover biologically relevant information from complex, integrated data.

Key Results

  • Cold Stress Analysis: Identified 425 response order preserving DEGs grouped into 12 response phase clusters.
  • Heat Stress Analysis: Identified 272 response order preserving DEGs grouped into 7 response phase clusters.
  • Reproduced Known Pathways: The HTRgene analysis results aligned well with established biological mechanisms for both cold and heat stress response in Arabidopsis, including key transcription factors and downstream genes.
  • Improved Accuracy: Compared to existing DEG detection tools (limma, ImpulseDE), HTRgene demonstrated a significantly higher F1 score, indicating better accuracy in identifying known stress response genes. The ordering step in HTRgene was shown to improve precision by over three-fold, albeit with a two-fold decrease in recall.

Limitations and Future Work

  • Computational Complexity: While clustering reduces complexity, determining the optimal response time vector for each cluster still involves a computationally intensive search. The paper mentions an O(n!) complexity for individual gene analysis, which is mitigated but not entirely eliminated.
  • Exclusion of Non-Stress Data: The current implementation excludes non-stress-treated time-series samples (control data). The authors suggest that utilizing this data could be a valuable direction for future research.
  • Circadian Rhythms: The paper notes that circadian rhythms can introduce noise and lead to the exclusion of relevant genes. Future work could explore methods to account for or integrate circadian effects.
  • Generalizability: While demonstrated for cold and heat stress in Arabidopsis, the method's applicability to other stresses, organisms, or data types could be further explored.

Summary updated

BibTeX
@article{ahn2019htrgene,
  author = {Hongryul Ahn and Inuk Jung and Heejoon Chae and Dongwon Kang and Woosuk Jung and Sun Kim},
  title = {{HTRgene: a computational method to perform the integrated analysis of multiple heterogeneous time-series data: case analysis of cold and heat stress response signaling genes in Arabidopsis}},
  journal = {BMC Bioinformatics},
  year = {2019},
  doi = {10.1186/s12859-019-3072-2},
}
Figure