Ryder: Epigenome normalization using a two-tier model and internal reference regions

Ryder is a flexible Python package that normalizes diverse epigenomic sequencing data by leveraging stable internal reference regions to correct technical variability, thereby enhancing the accuracy of genome-wide signal alignment and the detection of genuine biological changes.

The Gist

Imagine you are trying to compare the noise levels in two different concert halls. One hall is empty, and the other is packed with people. If you just turn up the volume on your microphone in the empty hall to match the loudness of the packed hall, you might think the empty hall is just as noisy. But that's wrong! The empty hall is actually quiet; you just turned the microphone gain up too high.

This is exactly the problem scientists face when studying epigenetics (the "software" that tells our DNA how to run). They use powerful machines to read the "volume" of genes and proteins. But sometimes, the machine itself gets a little wonky, or the sample preparation varies slightly, making one sample look louder or quieter than it really is. This "technical noise" can hide the real biological story.

Enter Ryder, a new tool created by scientists at the NIH to fix this problem. Here is how it works, using some everyday analogies:

The Problem: The "Broken Ruler"

Imagine you are measuring the height of plants in two different gardens.

• Garden A has a ruler that is slightly stretched out.
• Garden B has a ruler that is slightly squished.

If you measure a plant in Garden A and get "10 inches," and a plant in Garden B and get "10 inches," you might think they are the same height. But because the rulers are broken, one plant is actually 12 inches and the other is 8 inches.

In the past, scientists tried to fix this by bringing in a Spike-in Control. This is like bringing a standard, factory-made "10-inch stick" into both gardens to calibrate your rulers.

• The Catch: If you drop the stick in Garden A and it lands in mud, but in Garden B it lands on a rock, your measurement is still wrong. Also, if you bring in too many sticks, you run out of space to measure the actual plants. If you bring in too few, the stick doesn't help much. It's a delicate, often messy balancing act.

The Solution: Ryder and the "Internal Landmarks"

Instead of bringing in an outside stick, Ryder looks for landmarks inside the gardens that shouldn't change.

Think of the genome (your DNA) as a massive city. Most of the city changes constantly (buildings go up, traffic shifts). But there are certain landmarks that are permanent: the city hall, the main library, or a specific, ancient oak tree. No matter what happens in the rest of the city, the City Hall stays exactly where it is.

In the human body, a protein called CTCF acts like these permanent landmarks. It binds to specific spots on the DNA in almost every cell type, and it rarely moves.

• Ryder's Strategy: It says, "Let's ignore the noisy, changing parts of the city for a moment. Let's look at the City Hall (CTCF sites). If the signal at the City Hall looks different between my two samples, I know my 'ruler' is broken, not the city."

How Ryder Works (The Two-Tier Approach)

Ryder is smart because it doesn't just fix the whole picture at once. It uses a two-step process:

1. Step 1: Fix the Background Noise (The Static)
   Imagine there is a constant hum of static in your radio. Ryder first figures out how loud that static is in each sample and turns it down to the same level. It uses the "City Hall" spots to measure the static.

2. Step 2: Align the Signal (The Music)
   Once the static is fixed, Ryder looks at the actual music (the biological signals). It stretches or shrinks the volume of the music so that the "City Hall" spots match perfectly between the two samples.

By doing this separately for the "static" and the "music," Ryder avoids the mistakes of older methods that tried to do everything with one simple math trick.

Why This Matters: Finding the Real Story

The paper shows that without Ryder, scientists might miss important changes or see things that aren't there.

• The BRG1 Example: The scientists studied a protein called BRG1, which helps keep DNA accessible (like opening a window to let fresh air in). When they removed BRG1, they expected the "windows" to close.
  • Old Methods: Sometimes, the "static" (technical noise) was so loud that it looked like the windows were opening wider, or the change was too subtle to see.
  • With Ryder: By using the "City Hall" landmarks to calibrate the data, Ryder revealed a clear, dose-dependent truth: As BRG1 was removed, the windows (enhancers) slowly but surely closed. This is a real biological effect that was previously hidden by the noise.

The Bottom Line

Ryder is like a super-smart auto-tune for biological data. Instead of relying on an external reference that might get lost or contaminated, it uses the DNA's own "permanent landmarks" to ensure that when scientists compare two samples, they are comparing apples to apples, not apples to oranges.

It allows researchers to see the true biological signals clearly, ensuring that discoveries about how our genes work are based on reality, not on a glitch in the machine.

Technical Summary

1. Problem Statement

Sequencing-based epigenomic profiling methods (e.g., ChIP-seq, ATAC-seq, DNase-seq, CUT&RUN, MNase-seq) are essential for understanding gene regulation but suffer from significant technical variability. Sources of noise include sample quality, library preparation, sequencing depth, and batch effects.

• Limitations of Current Methods:
  • Spike-in Controls: While widely used, they rely on the assumption that exogenous chromatin behaves identically to endogenous chromatin. This is often unverified and prone to failure due to titration errors or inconsistent application (as seen in recent BRG1 studies). They often provide only a global scaling factor, failing to capture localized variations.
  • Computational Methods: Existing tools (e.g., MAnorm, S3norm, IGN) often rely on assumptions that may not hold, such as invariant binding in shared peaks (which may actually vary quantitatively) or the existence of perfectly matched RNA-seq data for normalization. They struggle with global chromatin shifts (e.g., during aging or drug treatment) or broadly spread signals (e.g., H3K27me3).

There is a critical need for a flexible normalization framework that can distinguish genuine biological signals from technical artifacts without relying on external spike-ins that may be unsuitable for specific assays.

2. Methodology: The Ryder Framework

Ryder is a Python package designed for cross-sample normalization and differential feature detection. Its core innovation is a two-tier normalization strategy anchored by stable internal reference regions (e.g., invariant CTCF binding sites).

Workflow Components:

The pipeline consists of two main scripts:

1. paw.py (Normalization): Rescales signals and corrects background noise.
2. patrol.py (Differential Analysis): Identifies variable features using fold-change or Poisson statistics.

The Two-Tier Normalization Algorithm:

Ryder assumes that constitutive CTCF binding sites (conserved across cell types) serve as stable internal references. The process involves five steps:

1. Outlier Removal: Uses Mahalanobis distance on M-A transformed log₂ signals to identify and remove outlier reference sites that do not behave invariantly.
2. Region Definition: Defines "signal" regions (peaks) and "background" regions (adjacent to references).
3. Parameter Estimation:
   • Calculates scaling factors for background ($sf_{bkg}$) and reference signal regions ($sf_{sig}$).
   • Estimates linear transformation parameters ($\alpha$ and $\beta$) via z-score transformation or linear fitting between target and reference signals in log₂ space:
     $$y' = \alpha y + \beta$$
     Where $\alpha = \sigma_x / \sigma_y$ and $\beta = \bar{x} - (\sigma_x / \sigma_y)\bar{y}$.
4. Classification: Genomic bins are classified as background or signal based on a noise cutoff derived from the intersection of their log₂ distributions.
5. Application:
   • Background bins: Scaled linearly using $sf_{bkg}$.
   • Signal bins: Undergo a two-step process: linear scaling using $sf_{sig}$, followed by log₂ transformation adjusted by $\alpha$ and $\beta$, then exponentiation.

Flexibility: The tool supports user-defined reference regions (e.g., TSSs) and can optionally integrate spike-in controls if desired, though it prioritizes internal references.

3. Key Contributions

• Internal Reference Strategy: Shifts the paradigm from external spike-ins to invariant internal genomic sites (specifically CTCF), which are inherently subjected to the same experimental conditions as the target regions.
• Two-Tier Modeling: Unlike single-factor global scaling, Ryder separately models and corrects background noise and signal intensity, ensuring accurate signal-to-noise ratios across samples.
• Assay Agnosticism: Validated across a wide range of assays: DNase-seq, CUT&RUN, ATAC-seq, MNase-seq, and ChIP-seq.
• Robustness to Global Shifts: Specifically designed to handle scenarios where global chromatin states change (e.g., drug inhibition, cellular aging), a scenario where many existing methods fail.

4. Key Results & Validation

The authors validated Ryder using multiple datasets and compared it against standard methods (RPM, spike-in ratios, and other computational tools).

• GATA3 Knockout (DNase-seq):
  • Observation: Unnormalized data showed a global signal increase in KO samples, obscuring specific biological changes.
  • Result: Ryder corrected the global bias, revealing a specific reduction in chromatin accessibility at GATA3-bound enhancers (e.g., the Ctla4 enhancer) that was previously masked.
• BRG1 Depletion (DNase-seq & CUT&RUN):
  • Context: BRG1 is a subunit of the BAF complex essential for enhancer accessibility.
  • Comparison: Standard RPM and spike-in methods failed to distinguish between global background shifts and specific binding loss. In one case, spike-in data was deemed unusable by the original authors.
  • Result: Ryder successfully identified a dose-dependent loss of BRG1 binding at enhancers and a corresponding dose-dependent decrease in chromatin accessibility. It correctly showed that promoter accessibility remained largely unchanged until extensive degradation, a subtle trend missed by other methods.
• MNase-seq (Nucleosome Occupancy):
  • Using only a background scaling factor derived from invariant CTCF sites, Ryder accurately detected increased nucleosome occupancy at enhancer centers upon BRG1 inhibition, consistent with the loss of nucleosome-free regions.
• ChIP-seq (Global Histone Modifications):
  • EZH2 Inhibition (H3K27me3): Correctly quantified the global decrease in H3K27me3 without creating the artifactual "increase" in stable H3K4me3 often seen with total-read normalization.
  • HDAC Inhibition (H3K9ac): Confirmed the expected global increase in H3K9ac levels.

5. Significance

• Improved Biological Interpretation: By effectively separating technical noise from biological signal, Ryder enables the detection of subtle, quantitative biological changes (e.g., dose-response relationships) that are often lost in standard normalization.
• Mitigation of Spike-in Risks: The paper highlights the risks of relying on spike-ins (titration errors, mismatched dynamics) and offers a robust, self-contained alternative using internal references.
• Reproducibility: The method provides a transparent, testable assumption (stability of CTCF sites) that can be verified within the dataset itself, unlike the opaque assumptions of some spike-in protocols.
• Accessibility: As a freely available Python package, Ryder lowers the barrier for researchers to perform high-quality, normalized epigenomic analyses across diverse experimental designs.

In conclusion, Ryder represents a significant advancement in epigenomic data analysis, offering a flexible, robust, and assumption-transparent framework that outperforms existing methods in scenarios involving global chromatin shifts and technical variability.