A supervised Bayesian method for time (re)annotation of transcriptomics data

Elio Nushi, François P. Douillard, Katja Selby, Benjamin A. Blount, Oliver Pennington, Nigel P. Minton, Miia Lindström, Antti HonkelaNAR 2025 · DOI

TL;DR

This paper introduces a novel Bayesian method using Gaussian Processes to accurately assign time annotations to historical transcriptomics data (like microarrays) that lack precise timestamps. By leveraging a well-annotated RNA-Seq time series as a reference, this approach significantly improves the biological interpretation of older datasets, leading to better identification of gene expression changes and a clearer understanding of cellular processes.

What Problem Does This Paper Address?

Many valuable biological datasets, particularly older ones like microarrays, suffer from imprecise or missing time annotations. This is often because collecting high-quality time-series data is expensive and time-consuming. However, without accurate temporal information, it's challenging to understand dynamic biological processes like cell cycles or developmental pathways, and to reliably identify differentially expressed genes. This paper tackles the problem of "re-annotating" these historical datasets with accurate time points by using a more comprehensive, modern time-series dataset as a reference.

Methods

The core of the proposed method is a Bayesian approach utilizing Gaussian Process (GP) regression. Here's a breakdown:

  • Reference Data Modeling: For each gene in the reference RNA-Seq time series, a GP model is trained. GPs are powerful because they can model complex, non-linear temporal patterns and importantly, provide a measure of uncertainty in their predictions. The kernel of the GP is chosen to capture temporal dynamics.
  • Time Annotation: For a target experiment (e.g., a microarray sample) with genome-wide gene expression values but no time annotation, the method calculates the likelihood of that sample belonging to each time point in the reference data. This is done by combining the likelihoods across all genes, effectively finding the time point that best explains the entire gene expression profile of the sample.
  • Quality Control: The authors implement several quality control steps to ensure reliable annotations. This includes filtering genes that don't fit the GP model well, comparing GP fits to simpler time-invariant models, and using leave-one-out cross-validation to assess predictive accuracy.
  • Comparison: The proposed GP method is compared against a baseline k-nearest neighbors (kNN) approach, which also uses the reference time series.

Datasets

The study uses two main datasets from Clostridium botulinum:

  • Reference Data: Longitudinal RNA-Seq samples collected from three replicate cultures over 2-12 hours, with optical density measurements to track growth phases. This dataset is used to train the GP models.
  • Target Data: 80 legacy microarray samples collected at different growth stages, originally annotated only by broad growth phase (e.g., early-exponential, stationary) but lacking precise time points.

Key Insights and Ideas

  • Bridging the Temporal Gap: The most significant insight is that even with different technologies (RNA-Seq vs. microarrays) and potentially different strains, a robust temporal recalibration is possible by leveraging a well-characterized reference time series.
  • GP's Predictive Power: Gaussian Processes are shown to be highly effective for modeling complex temporal gene expression dynamics and, crucially, for interpolating and predicting time points even when the reference data is sparse.
  • Uncertainty Matters: The Bayesian nature of GPs allows for quantifying the uncertainty associated with each time annotation, providing a more nuanced understanding of the reliability of the assigned time points.
  • Beyond Clock Time: The method not only assigns precise time points but also maps them to biologically meaningful "reference growth phases" (RGPs), which are more informative than the original, broad legacy growth phase (LGP) annotations.

Key Results

  • Improved Annotation Accuracy: The GP method achieved high accuracy, with predictions for sparse RNA-Seq data being within a 30-minute margin of the true time. Even with simulated noisy data (up to 30% of genes perturbed), 97% of annotations were within 0.5 hours.
  • Enhanced Biological Interpretation: Re-annotating legacy microarray data significantly improved the description of growth phases, as evidenced by Principal Component Analysis (PCA). This led to a clearer separation of samples belonging to different growth stages.
  • Increased Differential Gene Detection: The improved temporal annotations resulted in a larger number of differentially expressed genes detected when comparing successive growth phases, indicating a more accurate biological picture.
  • Qualitative Superiority: Compared to the kNN baseline, the GP method provided a higher resolution in describing the data and a clearer separation between samples in different growth phases.
  • Platform Agnostic: The method demonstrated its applicability across different transcriptomics platforms (RNA-Seq and microarrays) and even with related bacterial strains.

Limitations and Future Work

  • Reference Data Requirement: The method fundamentally relies on the availability of a well-annotated reference time series for the organism or a closely related one.
  • Gene Compatibility Filter: Genes whose expression patterns are highly inconsistent between the reference and target data, or those with very low signal-to-noise ratios, are filtered out. While this improves robustness, it means not all genes contribute to the annotation.
  • Future Directions: The authors suggest that this approach could be extended to other omics data types (proteomics, metabolomics) and to any data where time is a key driver of variation. Further exploration of how miss-labeled samples affect downstream analyses is also mentioned.

Summary updated

Abstract

Transcriptomics experiments are often conducted to capture changes in gene expression over time. However, time annotations may be missing, imprecise, or not reflect the same physiological state of the bacterial culture between different experiments. Assigning accurate time points to these experiments using a reference time course is therefore crucial for identifying differentially expressed genes, and understanding gene regulatory networks for elucidating the studied organism's physiology and life cycle. This important task, which could enhance the biological interpretation of the transcriptomics experiments, has not been previously addressed. In this work, we propose a novel method to solve the challenge of realigning transcriptomics experiments based on a reference time course. Our method is based on a Bayesian approach that uses Gaussian process regression modeling. We show a use case of applying our method for assigning time annotations in legacy microarray samples of the bacterium <i>Clostridium botulinum</i>, which were solely annotated based on the growth phase at the time when the culture aliquots were sampled, utilizing recently collected RNA-Seq time series data comprising multiple replicates as a reference. The method significantly improved the description of the growth phases of the microarray data compared to the original annotations by clearly delineating the microarray samples belonging to different growth phases, as demonstrated by principal component analysis. Consequently, a larger number of differentially expressed genes was detected when comparing experiments belonging to successive growth phases. We compare this innovative approach with a baseline method that uses k-nearest neighbor algorithm and show that our method offers a higher resolution in the description of the data by exposing smaller time changes between samples. We also test the performance of the method on sparse RNA-Seq time series (i.e. sampled every second hour). All the predictions for the samples were within a 30-min margin of their true time.
BibTeX
@inproceedings{nushi2025supervised,
  author = {Elio Nushi and François P. Douillard and Katja Selby and Benjamin A. Blount and Oliver Pennington and Nigel P. Minton and Miia Lindström and Antti Honkela},
  title = {{A supervised Bayesian method for time (re)annotation of transcriptomics data}},
  booktitle = {NAR},
  year = {2025},
  doi = {10.1093/nargab/lqaf203},
}
Figure