Bayesian Estimation of Mosaic Loss of Chromosome Y from Bulk RNA Sequencing Data

This study introduces a Bayesian framework that successfully estimates mosaic loss of chromosome Y (LOY) from bulk male RNA-seq data by modeling reduced Y-linked gene expression, achieving strong correlation with DNA-based measurements for large LOY events while highlighting the method's limitations due to transcriptional confounding and its nature as a probabilistic rather than direct substitute for DNA-based assays.

The Gist

Imagine your body is a massive library, and every cell in your body is a book. In men, most of these books have a specific chapter called "Chromosome Y." However, as men age, some of these books start losing that chapter entirely. This is called Mosaic Loss of Chromosome Y (LOY). Usually, scientists check for this missing chapter by looking directly at the DNA (the text of the book). But what if you only have a summary of the book's content (RNA data) and no access to the original text?

This paper introduces a clever new method to guess how many books are missing that chapter, using only the summaries.

The Problem: Reading the Summary Instead of the Book

Think of DNA as the original manuscript and RNA as a photocopy of the most popular pages. Scientists often have huge collections of these "photocopies" (RNA-seq data) from men, but they don't have the original manuscripts (DNA) to check if the "Chromosome Y" chapter is missing. They need a way to estimate the missing chapter just by looking at the photocopies.

The Solution: A "Smart Detective" Algorithm

The researchers built a Bayesian framework, which you can think of as a super-smart detective. Instead of just counting words, this detective looks at the "Chromosome Y" pages in the photocopy and asks:

• "Is this page missing more words than it should be?"
• "Could this be because the person is older?"
• "Is it just a fluke of how the book was copied?"

The detective compares the "Chromosome Y" pages against other pages that shouldn't be missing (the "control" pages). If the Y-pages are significantly quieter or shorter than the others, the detective concludes that some cells in that person's body are missing the Y-chromosome entirely.

How Well Did the Detective Work?

The team tested this detective on 377 men where they did have the original manuscripts (DNA data) to compare against.

• The Score: The detective's guesses matched the reality about 68% of the time (a correlation of 0.678).
• The Accuracy: On average, the guess was off by less than 2% of the total cells.
• Big Misses vs. Small Misses: The detective was excellent at spotting when a lot of the library was missing the chapter (over 20% of cells). It was like spotting a missing whole section of the library. However, it struggled to rank who had a tiny amount of missing chapters versus who had none at all. It's easier to see a huge hole in a wall than a single missing brick.

The "Confusing Noise" Warning

The researchers also tested this method on a different group of men (blood samples) where they couldn't check the original manuscripts. They found that while the estimates generally went up as the men got older (which makes sense), something else was messing with the signal.

They discovered that if you "wake up" the immune cells in a lab (ex vivo stimulation), the "Chromosome Y" pages in the photocopy suddenly look different, even if the original manuscript hasn't changed. This is like if a loud noise in the library made the photocopy machine skip a few lines, making it look like a chapter was missing when it wasn't. This means the method is sensitive to how the cells are behaving, not just what DNA they have.

The Bottom Line

This paper shows that you can use RNA data (the summaries) to get a probabilistic guess about missing Y-chromosomes. It's a very good tool for spotting big losses, but it shouldn't be treated as a perfect replacement for checking the original DNA. Think of it as a highly educated estimate based on the shadows the missing chapters cast, rather than a direct photograph of the missing pages themselves.

Technical Summary

Technical Summary: Bayesian Estimation of Mosaic Loss of Chromosome Y from Bulk RNA Sequencing Data

Problem Statement
Mosaic loss of chromosome Y (LOY) is a prevalent age-associated somatic alteration in men, traditionally quantified using DNA-based assays. However, a significant limitation exists in many large-scale research cohorts: while they possess bulk RNA sequencing (RNA-seq) data, they lack matched DNA-based LOY measurements. This gap prevents researchers from investigating the relationship between LOY and gene expression or other transcriptomic phenotypes in these existing datasets. The core challenge addressed is how to reliably infer the fraction of cells with LOY using only bulk RNA-seq data, where the signal is confounded by transcriptional noise, age-related expression changes, and the absence of a direct genomic readout.

Methodology
The authors developed a Bayesian framework designed to estimate the LOY fraction from male bulk RNA-seq data. The core logic of the model relies on detecting reduced expression of Y-linked genes relative to an expected baseline. This baseline is established by adjusting for:

• Covariates: Age and other expression-related variables.
• Control Genes: Autosomal and X-linked genes used to normalize for global expression shifts.

Two specific modeling approaches were implemented and compared:

1. Fast Empirical Bayes Estimator: A primary model that calculates point estimates of the LOY fraction.
2. Mixture/PCA Hierarchical Bayesian Sensitivity Model: A more complex model incorporating Principal Component Analysis (PCA) to capture latent structure, providing interpretable posterior quantities.

The models were trained and validated using 377 male samples from the Genotype-Tissue Expression (GTEx) project, for which independent, DNA-based LOY measurements were available. The validation strategy included:

• Correlation Analysis: Assessing the relationship between estimated and measured LOY.
• Error Metrics: Calculating Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and empirical coverage.
• Sensitivity Analyses: "Leave-one-Y-gene-out" tests to verify signal distribution and prior-sensitivity analyses to evaluate the impact of shrinkage priors on calibration.
• External Validation: Testing the model on an external whole-blood RNA-seq dataset (without measured LOY) to observe age trends and the effects of ex vivo immune stimulation.

Key Contributions and Results
The study demonstrates that bulk RNA-seq data contains usable information regarding LOY, though with specific limitations regarding precision.

• Signal Validation: Individual Y-linked genes showed negative correlations with independently measured DNA-based LOY, confirming that LOY manifests as a shared depletion signal in the transcriptome.
• Performance Metrics: The primary fast empirical Bayes estimator achieved a Pearson correlation of 0.678 with measured LOY. It yielded a Mean Absolute Error of 1.79% and a Root Mean Squared Error of 3.72%. Crucially, the model achieved 95.2% empirical coverage of the measured LOY values.
• Event Detection Capability: Performance was highly stratified by the magnitude of the LOY event. The model excelled at identifying large LOY events, achieving an Area Under the Curve (AUC) of 0.964 for samples with measured LOY greater than 20%. Conversely, the model's ability to finely rank or distinguish between samples with low levels of LOY remained uncertain.
• Model Comparison: The more complex mixture/PCA hierarchical Bayesian model provided similar validation performance and offered interpretable posterior quantities but did not improve the point estimation accuracy over the faster empirical Bayes approach.
• Robustness and Confounding: Sensitivity analyses confirmed that the LOY signal is distributed across multiple Y-linked transcripts rather than relying on a single gene. However, the external validation revealed a critical limitation: ex vivo immune stimulation in blood samples shifted RNA-derived LOY estimates and reduced the expression of multiple Y-linked transcripts. This indicates that transcriptional confounding can mimic or obscure LOY signals.

Significance and Claims
The paper concludes that while bulk RNA-seq can serve as a source of information for estimating LOY, particularly for detecting large-scale mosaic loss events, it is not a direct substitute for DNA-based mosaicism measurement. The authors position the RNA-derived LOY estimate as a probabilistic transcriptome-based estimate.

The significance of this work lies in enabling the analysis of LOY in the vast number of existing cohorts that possess RNA-seq data but lack matched DNA assays. However, the authors maintain a modest stance, emphasizing that these estimates must be interpreted with caution due to transcriptional confounding factors (such as immune stimulation) and the inherent difficulty in resolving low-level mosaicism. The framework provides a statistical tool to leverage existing data while clearly delineating the boundaries of its accuracy and applicability.