tauFisher accurately predicts circadian time from a single sample of bulk and single-cell transcriptomic data

Junyan Duan, Michelle N. Ngo, Satya Swaroop Karri, Lam C. Tsoi, Jóhann E. Guðjónsson, Babak Shahbaba, John Lowengrub, Bogi AndersenbioRxiv 2023 · DOI

TL;DR

This paper introduces tauFisher, a new computational tool that accurately predicts the internal circadian time from a single transcriptomic sample. This is crucial for advancing circadian medicine and understanding biological rhythms, especially in single-cell data where time information is often missing. tauFisher demonstrates impressive accuracy across various data types and platforms, outperforming existing methods.

What Problem Does This Paper Address?

The circadian clock, our internal 24-hour biological timer, regulates a vast array of physiological processes. Disruptions to this clock are linked to numerous diseases, making the ability to determine a tissue's or patient's internal circadian time essential for circadian medicine – tailoring treatments to optimize their effectiveness and minimize side effects. However, most transcriptomic datasets lack this critical time information, hindering research and clinical applications. Existing methods for inferring circadian time from transcriptomic data often require complete time-series data, are platform-specific, or are computationally intensive. This paper aims to bridge this gap by providing a robust, platform-agnostic tool that can accurately predict circadian time from a single sample, even for single-cell RNA sequencing (scRNA-seq) data.

Methods

tauFisher employs a multi-step pipeline:

  1. Identifying Diurnal Genes: It first identifies genes that exhibit a 24-hour rhythmic expression pattern using methods like JTK_Cycle or Lomb-Scargle.
  2. Functional Data Analysis (FDA): To handle varying sampling intervals and reduce noise, discrete gene expression data points are converted into continuous functions using Fourier basis expansion.
  3. Within-Sample Normalization: This novel step calculates the difference in expression between pairs of predictor genes at each time point and then scales these differences to a [0, 1] range. This makes the data unit-less and allows for direct comparison between training and testing datasets without re-training.
  4. Principal Component Analysis (PCA): The normalized differences are then projected into a lower-dimensional space using PCA.
  5. Multinomial Regression: The first two principal components are used as covariates in a multinomial logistic regression model to predict the circadian time.

A key innovation is the within-sample normalization step, which allows tauFisher to make accurate predictions from a single sample without needing a complete time series for testing.

Datasets

The study utilized several publicly available datasets for benchmarking and validation, including:

  • Bulk RNA-seq and Microarray Datasets: A diverse set of datasets from various species (human, mouse), tissues (blood, skin, SCN, kidney, liver), and sequencing platforms (Affymetrix, Illumina) were used to benchmark tauFisher against existing methods (ZeitZeiger, TimeSignatR).
  • Single-Cell RNA-seq (scRNA-seq) Datasets:
    • Mouse SCN scRNA-seq data (GSE117295) was used to test tauFisher's ability to predict circadian time on pseudobulk data derived from scRNA-seq.
  • Newly generated mouse dermal skin scRNA-seq data was also used for similar testing.

Key Insights and Ideas

  • Circadian Heterogeneity in Immune Cells: The study revealed that dermal immune cells exhibit greater circadian phase heterogeneity compared to dermal fibroblasts. This suggests that while the collective rhythm in immune cells appears dampened, it's due to individual cells being out of sync rather than a uniformly weak clock.
  • tauFisher's Cross-Platform and Cross-Data Type Robustness: tauFisher demonstrates remarkable consistency in predicting circadian time whether trained on bulk data and tested on bulk data, or trained on bulk data and tested on scRNA-seq data. This highlights its generalizability and robustness across different assay technologies and experimental setups.
  • Importance of Within-Sample Normalization: This novel step is crucial for tauFisher's ability to accurately predict circadian time from a single sample, effectively removing batch effects and scaling data to a common reference without re-training.
  • Divergent Rhythmic Pathways in Skin Cells: Dermal fibroblasts and immune cells, despite sharing some core clock genes, exhibit distinct rhythmic pathways. Fibroblasts are enriched for metabolism-related rhythmic processes, while immune cells show enrichment in immune response pathways.

Key Results

  • Superior Accuracy: tauFisher consistently outperformed ZeitZeiger and TimeSignatR in accuracy and RMSE across all eight benchmark bulk RNA-seq and microarray datasets. It achieved an accuracy of 0.875 with an RMSE of 2.669 in cross-platform bulk predictions.
  • Accurate scRNA-seq Prediction: tauFisher accurately predicted circadian time for scRNA-seq data, achieving an accuracy of 0.833 for SCN cells and 0.778 for dermal skin cells.
  • Computational Efficiency: tauFisher's reliance on a few key features makes training and testing computationally efficient.
  • Platform Agnostic: Once trained, the same tauFisher predictor can be used for datasets from the same tissue, regardless of the platform.

Limitations and Future Work

  • Cell-Type Specificity for scRNA-seq: While tauFisher can predict circadian time for scRNA-seq data, the current study was limited to pseudobulk analysis due to insufficient cell counts for specific immune cell subtypes to generate reliable individual cell-type pseudobulk data. Future work could explore methods to address this.
  • Understanding Dampened Rhythms: The study hypothesizes that circadian heterogeneity contributes to dampened rhythms in immune cells. Further research could investigate whether stimulating synchronization of these heterogeneous clocks is more beneficial than stimulating individual cell clocks in certain disease contexts.
  • Integration with Other Methods: The authors suggest that combining tauFisher with batch-effect correction methods could facilitate cleaner integration of diverse datasets and improve transcriptomic data analysis.

The development of tauFisher represents a significant step forward in our ability to leverage transcriptomic data for circadian research and the emerging field of circadian medicine.

Summary updated

Abstract

Abstract As the circadian clock regulates fundamental biological processes, disrupted clocks are often observed in patients and diseased tissues. Determining the circadian time of the patient or the tissue of focus is essential in circadian medicine and research. Here we present tau-Fisher, a computational pipeline that accurately predicts circadian time from a single transcriptomic sample by finding correlations between rhythmic genes within the sample. We demonstrate tauFisher’s out-standing performance in both bulk and single-cell transcriptomic data collected from multiple tissue types and experimental settings. Application of tauFisher at a cell-type level in a single-cell RNA-seq dataset collected from mouse dermal skin implies that greater circadian phase heterogeneity may explain the dampened rhythm of collective core clock gene expression in dermal immune cells compared to dermal fibroblasts. Given its robustness and generalizability across assay platforms, experimental setups, and tissue types, as well as its potential application in single-cell RNA-seq data analysis, tauFisher is a promising tool that facilitates circadian medicine and research.
BibTeX
@misc{duan2023taufisher,
  author = {Junyan Duan and Michelle N. Ngo and Satya Swaroop Karri and Lam C. Tsoi and Jóhann E. Guðjónsson and Babak Shahbaba and John Lowengrub and Bogi Andersen},
  title = {{tauFisher accurately predicts circadian time from a single sample of bulk and single-cell transcriptomic data}},
  year = {2023},
  doi = {10.1101/2023.04.04.535473},
  howpublished = {bioRxiv},
}
Figure