TL;DR
This paper introduces MOSAIC, a novel computational framework that integrates multi-omics data (like RNA and protein levels) to more accurately identify circadian rhythms. By jointly modeling data from less noisy sources (e.g., RNA) with noisier sources (e.g., proteins), MOSAIC significantly improves the detection of rhythmic patterns, especially in proteomic datasets, and also accounts for non-circadian trends.
What Problem Does This Paper Address?
Circadian rhythms, our internal ~24-hour biological clocks, regulate a vast array of physiological processes. Identifying these rhythms typically involves analyzing time-series data from a single "omics" type, such as transcriptomics (RNA levels) or proteomics (protein levels). However, a key challenge arises because RNA rhythms don't always translate to protein rhythms, suggesting complex post-transcriptional regulation. Furthermore, proteomic data is notoriously noisy, making it difficult to distinguish true biological rhythms from experimental artifacts. Existing methods often analyze each omics type independently, potentially missing rhythms in noisy data or misattributing noise to regulatory mechanisms. This paper aims to bridge this gap by developing a method that leverages the lower noise of transcriptomic data to enhance the detection of rhythms in noisier proteomic data, while also accounting for non-oscillatory biological trends.
Methods
MOSAIC employs a four-stage workflow:
- Model Selection: For each gene in each omics dataset, MOSAIC fits several models: linear, exponential (non-oscillatory), and ECHO (Extended Circadian Harmonic Oscillator, which accounts for changing amplitudes). It then selects the best-fitting model for each gene using the Bayesian Information Criterion (BIC).
- Joint Modeling: This is the core innovation. If a gene shows rhythmic patterns in one omics type but not the other (e.g., rhythmic RNA but not protein), MOSAIC jointly models both datasets. This allows the less noisy data to inform the model fitting for the noisier data, specifically by sharing the period parameter while allowing other parameters to vary independently. This stage is crucial for rescuing rhythms masked by noise.
- Model Comparison: The performance of the joint model is compared against the independently fitted models using BIC to ensure that joint modeling actually provides an improvement.
- Goodness of Fit Evaluation: The quality of the final model fits is assessed using p-values derived from Kendall's tau rank correlation, adjusted for multiple hypothesis testing.
Datasets
The authors evaluated MOSAIC using:
- Synthetic Datasets: These were generated to mimic real biological data, with varying resolutions (2, 4, 6 hours), replicates (1, 2, 3), and noise levels (low, medium, high), specifically simulating higher noise in the "proteomic" data. This allows for a controlled assessment of MOSAIC's performance.
- Neurospora crassa Data: Publicly available transcriptomic and proteomic time-series data from Neurospora crassa was used to demonstrate MOSAIC's efficacy on real biological samples.
Key Insights and Ideas
- Noise as a Mask: The paper strongly suggests that many "arrhythmic" proteins might actually be rhythmic, but their oscillations are obscured by experimental noise in proteomic measurements.
- Synergy of Multi-Omics: Jointly modeling data from different omics layers isn't just about combining information; it's about leveraging the strengths of each layer (e.g., low noise, high signal) to overcome the weaknesses of others.
- Beyond Oscillations: Recognizing that not all biological regulation is rhythmic is crucial. MOSAIC's inclusion of non-oscillatory models provides a more complete picture of cellular control mechanisms.
- Differential Regulation: The analysis of Neurospora crassa data revealed significant differences in how the circadian clock regulates RNA versus protein levels, with RNA showing broader and more diverse oscillatory control.
Key Results
- Improved Rhythm Detection: MOSAIC accurately recovers circadian rhythms at higher rates in proteomic data compared to existing methods, and also improves detection in transcriptomic data.
- Synthetic Data Performance: MOSAIC showed robust performance across various synthetic data conditions, with joint modeling rescuing up to 45 oscillating proteins in challenging (high noise, low resolution) scenarios.
- Comparison to Existing Methods: When reduced to a rhythm-detection-only mode, MOSAIC outperformed ECHO, JTK_CYCLE, and MetaCycle in terms of F1-scores and true positive rates, especially under noisy conditions, though with a slightly higher false discovery rate in the worst-case scenarios.
- Neurospora crassa Findings: In the Neurospora crassa dataset, MOSAIC identified over 300 novel oscillatory proteins and 43 oscillatory transcripts. The analysis also highlighted that 18.5% of the proteome oscillated independently of the transcriptome, underscoring the role of post-transcriptional regulation.
- Broader Regulatory Landscape: MOSAIC revealed that the circadian clock broadly regulates RNA compared to protein, with oscillatory transcripts and proteins involved in a more diverse range of biological processes.
Limitations and Future Work
- False Discovery Rate: While MOSAIC excels at identifying true positives, its False Discovery Rate (FDR) can increase under very challenging conditions (e.g., low resolution, single replicate). The authors recommend adhering to best practices for data collection to mitigate this.
- Nonlinear Least Squares: The paper acknowledges the nonconvexity of nonlinear least squares fitting, which necessitates careful selection of starting points to ensure optimal solutions.
- Future Directions: The authors suggest that MOSAIC's multi-omics workflow could be extended to other omics types (e.g., phosphoproteomics, metabolomics) and other oscillatory processes (e.g., cell cycle), making it a versatile tool for broader biological rhythm analysis.
Summary updated