DNA methylation and hydroxymethylation quantification using vibrational spectroscopy

This study demonstrates that attenuated total reflectance Fourier transform infrared (ATR-FTIR) spectroscopy, combined with regression modeling, enables rapid, label-free, and non-destructive global quantification of DNA 5-methylcytosine and 5-hydroxymethylcytosine modifications, including in circulating tumor DNA samples.

The Gist

Imagine your DNA is a massive library of instruction manuals for building and running your body. Inside these manuals, there are tiny sticky notes attached to the letters. These sticky notes are called epigenetic marks. They don't change the letters themselves, but they tell the cell which instructions to read and which to ignore.

Two of the most important types of sticky notes are:

1. 5-mC (The "Methyl" Note): A simple little tag that usually tells a gene to "shut up" or stay silent.
2. 5-hmC (The "Hydroxymethyl" Note): A slightly more complex tag, often acting as a "pause" button or a sign that the gene is being prepared to be turned back on.

In diseases like cancer, these sticky notes get messed up. Genes that should be silent get turned on, and genes that should be active get silenced. To understand cancer, doctors need to count exactly how many of these notes are on the DNA.

The Problem: The Old Way is Clunky

Currently, counting these notes is like trying to read a book in the dark by using a sledgehammer. The standard methods involve:

• Chemical baths: Soaking DNA in harsh chemicals that can destroy the sample.
• Expensive machines: Giant, costly sequencers that only big labs can afford.
• Long wait times: It takes days to get a result.

The Solution: The "Vibrational Scanner"

This paper introduces a new, much simpler way to count these notes using a technique called ATR-FTIR spectroscopy.

Think of DNA molecules like tiny tuning forks. Every time you hit a tuning fork, it vibrates at a specific pitch.

• Unmethylated DNA (no sticky note) vibrates at one pitch.
• Methylated DNA (with the methyl note) vibrates at a slightly different pitch because the note adds a tiny bit of weight.
• Hydroxymethylated DNA (with the hydroxymethyl note) vibrates at a very different pitch because that note is heavier and shaped differently.

The researchers shine a special light (infrared light) on the DNA. The DNA absorbs the light and starts to "sing" (vibrate). By listening to the song, the machine can tell exactly what kind of sticky notes are attached.

The Magic of the "Smart Ear" (Machine Learning)

The song the DNA sings is complex. It's not just one note; it's a whole chord. To make sense of it, the researchers used Machine Learning (a type of computer brain).

1. Training the Brain: They fed the computer thousands of "songs" from DNA samples where they already knew the exact number of sticky notes. The computer learned the pattern: "Oh, when the song sounds like THIS, it means 50% methyl notes. When it sounds like THAT, it means 20% hydroxymethyl notes."
2. The Result: The computer became a master translator. It could look at a new, unknown DNA sample, listen to its song, and instantly say, "This sample has 45% methyl notes and 10% hydroxymethyl notes."

The "Hydroxymethyl" Surprise

The researchers found something interesting: The "Hydroxymethyl" note (5-hmC) is much louder and easier to hear than the "Methyl" note (5-mC).

• Analogy: Imagine the Methyl note is a whisper, and the Hydroxymethyl note is a shout. The computer could hear the shout perfectly, even in a noisy room. The whisper was harder to hear, especially when mixed with other sounds, but the computer still did a great job guessing it.

The Real-World Test: Listening to the Bloodstream

The ultimate test was to see if this works on ctDNA (circulating tumor DNA). This is DNA that cancer cells shed into your bloodstream. It's like finding a few torn pages from the instruction manual floating in a river.

• The Challenge: These torn pages are short, messy, and mixed with other debris. They sound different than the clean, long pages of the DNA they used for training.
• The Fix: The researchers taught the computer to "translate" the messy river sound into the clean library sound. They used a technique called Domain Adaptation (basically, a translator that says, "Okay, this river noise actually means the same thing as that library noise").
• The Win: It worked! The machine could accurately count the sticky notes on the messy, floating DNA fragments from the blood.

Why This Matters

This new method is:

• Fast: No more waiting days for results.
• Cheap: No expensive chemicals or giant machines.
• Gentle: It doesn't destroy the DNA, so you can use the sample for other tests later.
• Simple: You just drop a tiny drop of DNA on a crystal, press a button, and get an answer.

In summary: The researchers built a "vibrational microphone" that listens to the unique song of DNA. By teaching a computer to recognize the notes in that song, they created a fast, cheap, and gentle way to detect cancer markers in blood, potentially changing how we diagnose and monitor cancer in the future.

Technical Summary

1. Problem Statement

Quantifying global levels of DNA cytosine modifications—specifically 5-methylcytosine (5-mC) and 5-hydroxymethylcytosine (5-hmC)—is critical for understanding cancer biology and developing epigenetic biomarkers. However, current established methods face significant limitations:

• Bisulfite Sequencing: Destructive to DNA, cannot distinguish 5-mC from 5-hmC without complex chemical steps, and requires expensive sequencing infrastructure.
• Enzymatic Methods (EM-seq) & Arrays: Still rely on costly sequencing or are restricted to predefined genomic regions (arrays) and often fail to resolve 5-hmC.
• Mass Spectrometry (LC-MS/MS): Requires DNA digestion and specialized, expensive instrumentation.
• Immunoassays (ELISA): Suffer from antibody cross-reactivity and variability.
• Gap: There is a lack of rapid, label-free, non-destructive, and cost-effective methods capable of simultaneously quantifying global 5-mC and 5-hmC, particularly for low-input clinical samples like circulating tumor DNA (ctDNA).

2. Methodology

The authors propose a framework using Attenuated Total Reflectance Fourier Transform Infrared (ATR-FTIR) spectroscopy combined with machine learning to quantify these modifications.

• Sample Preparation:
  • Standards: Used 338 bp APC promoter DNA fragments generated via PCR with dCTP (unmethylated), 5-methyl-dCTP (5-mC), or 5-hydroxymethyl-dCTP (5-hmC).
  • Mixtures: Created binary (5-mC vs. 5-hmC) and ternary (unmethylated, 5-mC, 5-hmC) mixtures spanning 0–100% modification levels.
  • Clinical Proxy: Used Seraseq ctDNA reference materials (multiplexed, 7 genomic regions, 155–220 bp fragments) to simulate clinical samples.
• Spectroscopy:
  • Acquired spectra (750–4000 cm⁻¹) using an Agilent Cary 670 FTIR spectrometer.
  • Samples were air-dried on a diamond-coated ZnSe ATR crystal.
• Data Preprocessing:
  • Vector normalization, baseline correction, Savitzky-Golay smoothing, and second-order derivative transformation to enhance resolution.
  • Domain adaptation techniques were applied to align ctDNA spectra with synthetic DNA standards.
• Analytical Approaches:
  1. Univariate Analysis: Identified specific peak ratios (e.g., 1660/1060 cm⁻¹) correlated with modification levels.
  2. Multivariate Analysis: Employed Partial Least Squares Regression (PLSR) to model complex spectral data and predict modification percentages.
  3. Domain Adaptation: Used PCA to identify and remove "DNA-type" variance (differences between synthetic and ctDNA) before applying PLSR for ctDNA quantification.

3. Key Contributions

• First Demonstration of 5-hmC Quantification via FTIR: While previous studies focused on 5-mC, this work proves ATR-FTIR can distinguish and quantify 5-hmC, noting that the hydroxymethyl group produces more distinct spectral changes than the methyl group.
• Simultaneous Global Quantification: Developed a single-acquisition method to quantify both 5-mC and 5-hmC in complex mixtures without chemical conversion or labeling.
• Clinical Transferability (Proof-of-Concept): Successfully bridged the "domain gap" between synthetic DNA standards and real-world ctDNA using domain adaptation, enabling quantitative calibration for clinical samples.
• Identification of Spectral Biomarkers: Pinpointed specific vibrational modes sensitive to modification:
  • ~1660 cm⁻¹: Nucleobase carbonyl (C=O) stretching (increases with modification).
  • ~1060 cm⁻¹: Phosphate backbone (PO₂⁻) symmetric stretching (decreases with modification).
  • ~3200–3400 cm⁻¹: O-H/N-H stretching (specifically enhanced in 5-hmC due to hydrogen bonding).

4. Results

• Spectral Characterization:
  • PCA clearly separated unmethylated, 5-mC, and 5-hmC DNA. PC1 (85.1% variance) distinguished 5-hmC from the others, while PC2 (10.6% variance) separated 5-mC from unmethylated DNA.
  • 5-hmC showed greater spectral distinctiveness than 5-mC, particularly in the hydroxyl (O-H) and phosphate regions.
• Quantification Performance (Synthetic DNA):
  • Univariate (Peak Ratio 1660/1060): Achieved $R^2 = 0.97$ for both 5-mC and 5-hmC.
  • Multivariate (PLSR): Improved accuracy significantly.
    • 5-hmC: $R^2 = 0.99$, RMSE = 2.6%.
    • 5-mC: $R^2 = 0.97$, RMSE = 5.7%.
• Complex Mixtures:
  • In ternary mixtures (all three states present), 5-hmC remained highly quantifiable ($R^2 = 0.97$, RMSE = 5.1%).
  • 5-mC accuracy decreased ($R^2 = 0.90$, RMSE = 9.6%) due to the subtler spectral signature of the methyl group being masked by overlapping signals.
• ctDNA Application:
  • Direct application of synthetic models failed due to domain shift (fragment length, sequence diversity).
  • After domain adaptation (removing PC1 variance and aligning mean/std), the model achieved a calibration curve for ctDNA with $R^2 = 0.98$ and RMSE = 5.2% under cross-validation.

5. Significance

• Rapid and Non-Destructive: The method requires minimal sample preparation, no chemical conversion, and preserves DNA integrity, making it suitable for low-input clinical samples.
• Cost-Effective: Utilizes standard FTIR instrumentation, avoiding the high costs of sequencing or mass spectrometry.
• Clinical Potential: Establishes a proof-of-concept for using vibrational spectroscopy to monitor global epigenetic changes in ctDNA, a minimally invasive biomarker for cancer diagnosis and monitoring.
• Future Direction: The study suggests that while 5-hmC is easier to detect via FTIR, future work should focus on improving 5-mC sensitivity in complex mixtures, potentially through multimodal fusion with Raman spectroscopy or training models specifically on diverse ctDNA datasets.

In conclusion, this paper establishes ATR-FTIR spectroscopy as a viable, rapid, and label-free platform for the global quantification of DNA methylation and hydroxymethylation, successfully extending the technology from synthetic standards to clinically relevant ctDNA samples.