Improved quantitation in data-independent acquisition proteomics via retention time boundary imputation

This paper introduces Nettle, an open-source tool that improves quantitation in data-independent acquisition proteomics by imputing peptide retention time boundaries to integrate chromatographic signals, thereby reducing missing data limitations and enhancing accuracy compared to traditional methods.

The Gist

Imagine you are trying to listen to a choir of thousands of singers (proteins) in a massive, echoing hall. Your job is to record how loud each singer is to understand the song they are singing. This is what scientists do in proteomics: they use a machine called a mass spectrometer to "listen" to proteins in a sample.

However, there's a problem. Sometimes, a singer is too quiet, or the hall is too noisy, and your recorder misses them entirely. In the data, these missing singers show up as blank spots or "missing values."

The Old Way: Guessing the Silence

Traditionally, when scientists found these blank spots, they had two bad options:

1. Throw them away: If a singer was missing too often, they'd just ignore that singer completely. But this is like ignoring the quietest members of the choir, who might actually be singing the most important part of the song.
2. Plug in a guess: They would use math to guess what the missing volume should have been. Imagine looking at the other singers and saying, "Well, the guy next to him is loud, so this quiet guy must be loud too." The problem is, these guesses are often wrong. They can create fake patterns (like thinking two singers are harmonizing when they aren't) or dilute the real signal with noise. It's like filling in a blank spot on a map with a drawing of a mountain that isn't actually there.

The New Way: Nettle (The Time-Traveling Detective)

This paper introduces a new tool called Nettle. Instead of guessing the volume (the missing number), Nettle guesses the time the singer was supposed to be there.

Here is the analogy:
Imagine you are watching a parade. You know a specific float (a peptide) is supposed to pass by at 2:00 PM. But on one day, your camera glitched, and you missed it.

• The Old Way: You guess, "Maybe it was 2:05 PM, and maybe it was 50 feet wide." You just make up a number and move on.
• The Nettle Way: Nettle looks at the other days the parade happened. It sees that on Tuesday, the float passed at 2:02 PM. On Wednesday, it passed at 1:58 PM. It realizes, "Ah, there's a pattern! The float usually passes between 1:55 and 2:05."

Nettle doesn't guess the loudness; it calculates the start and end time (the "retention time boundaries") when that float should have been visible. Once it knows the exact time window, it goes back to the raw video footage of that specific day, finds that tiny window, and measures the actual signal that was there, even if it was very faint.

Why This is a Big Deal

The authors tested Nettle on four different scenarios, and it worked like magic:

1. It found more singers: In a study of Alzheimer's disease brain tissue, Nettle found many more "differentially abundant" peptides (singers who were louder in sick brains than healthy ones) than the old methods. It found specific markers for the disease that were previously invisible because they were too quiet to be "heard" without the right timing.
2. It saw the faintest whispers: In a test where they diluted a sample until it was almost invisible (like listening to a whisper from a mile away), Nettle could still measure the signal. It extended the range of what the machine could detect, allowing scientists to see proteins that were previously too scarce to count.
3. It fixed the "glitchy camera": Sometimes, the parade schedule shifts slightly (the machine drifts). Nettle noticed that on one specific day, the float arrived 5 minutes early. Instead of getting confused, Nettle adjusted its time window to catch the float at the new time, ensuring the measurement was accurate.
4. It helped predict radiation: In a test involving mice exposed to radiation, the scientists wanted to predict the dose based on protein levels. Because Nettle could measure the faint proteins more accurately, the computer model became much better at guessing the radiation dose, reducing errors significantly.

The Bottom Line

Think of Nettle as a smart, time-traveling editor. Instead of filling in a blank page with a random guess, it looks at the context, figures out exactly when the missing information should have been there, and then goes back to the source to extract the real, measured data.

This means scientists can trust their data more, find more important biological clues (like early signs of disease), and study proteins that are too rare to be seen with traditional methods. It turns "missing data" into "measured data" by being smarter about when to look.

Technical Summary

1. Problem Statement

In Data-Independent Acquisition (DIA) proteomics, missing values (peptides present in the sample but lacking quantitative assignments) are a major bottleneck. This missingness arises from poor ionization, co-elution, or inability to confidently assign spectra.

• Current Limitations: Traditional handling methods involve either:
  1. Removing proteins with high missingness rates, which reduces statistical power and biases against low-abundance proteins.
  2. "Plug-in" Imputation: Replacing missing values with statistical estimates (e.g., low-value Gaussian draws, K-Nearest Neighbors, MissForest, or Deep Learning models).
• Drawbacks of Plug-in Imputation: These methods introduce spurious correlations, dilute biological signals, fail to account for uncertainty, and often perform poorly when missingness is "Missing Not At Random" (MNAR)—a common scenario where low-abundance peptides are more likely to be missing. Furthermore, they cannot recover quantitative ratios when a peptide is completely missing in one experimental group.

2. Methodology: Nettle

The authors propose Nettle, a novel approach that imputes Retention Time (RT) boundaries rather than quantitative values directly.

• Core Concept: Instead of guessing a missing intensity, Nettle predicts the start and end times of a peptide's chromatographic peak (RT boundaries) by borrowing information from related runs. Once these boundaries are imputed, the software integrates the Extracted Ion Current (XIC) within those specific windows to generate a real, measured quantitative value.
• Algorithm:
  1. Input: An "RT Matrix" where rows are transitions and columns are RT start/end times for each MS run.
  2. Imputation Engine: Nettle uses an ensemble of two machine learning methods to fill missing RT boundaries:
     • Weighted k-Nearest Neighbors (wKNN): Identifies similar runs based on Euclidean distance of RT profiles.
     • Soft-Impute: A low-rank matrix completion algorithm using iterative regularized singular value decomposition (SVD).
  3. Integration: The imputed RT boundaries are exported to Skyline, which performs the actual integration of the XIC signal within the predicted windows.
• Key Distinction: Unlike "Match-Between-Runs" (MBR) which transfers peptide identifications (requiring a detectable MS2 signal), Nettle provides integration boundaries regardless of MS2 detection. If the signal is absent, the integration yields a near-zero value, which is a valid measurement of absence rather than a missing value.

3. Key Contributions

• Novel Paradigm: Shifts the imputation target from quantitative intensity (statistical estimation) to chromatographic boundaries (physical measurement), resulting in "real" measured quantities.
• Open-Source Tool: Development of Nettle, a standalone Python tool available on GitHub, designed to integrate with standard DIA workflows (DIA-NN and Skyline).
• Handling Systematic Shifts: The method implicitly learns and corrects for systematic RT shifts between replicates without requiring explicit RT alignment.

4. Results

The authors evaluated Nettle on four distinct datasets:

A. Quantitative Accuracy (SMTG & MMCC Datasets)

• Accuracy: In a hold-out experiment using Alzheimer's disease (SMTG) data, Nettle achieved a significantly lower Mean Squared Error (MSE) compared to traditional plug-in methods (KNN, MissForest, Low-Value) under both MCAR and MNAR conditions.
• Dynamic Range: Using Matrix-Matched Calibration Curve (MMCC) yeast datasets, Nettle demonstrated the ability to quantify peptides at extreme dilutions (down to 1%) where traditional library searches failed to assign peaks.
• Differential Abundance (DA): In the MMCC serial dilution, Nettle identified an average of 10,105 additional DA peptides per condition compared to unimputed data, effectively recovering signals lost to missingness.

B. Biological Discovery (Alzheimer's Disease)

• New Biomarkers: Nettle identified significantly more DA peptides between AD subtypes (e.g., Familial AD vs. Sporadic AD) than unimputed data.
• Specific Markers: It successfully recovered quantitation for key AD markers (COL25A1, MAPT, MDK, SMOC1) that were missing in standard analysis.
• Validation: In a controlled experiment where 30% of data was artificially masked, Nettle preserved known biological differences in Amyloid-β peptides, whereas KNN plug-in imputation failed to detect significant differences between groups.

C. Linear Dynamic Range Extension

• Using the Mag-Net MMCC dataset, the authors performed a "turning point" analysis. Nettle lowered the Lower Limit of Quantitation (LLOQ), effectively extending the linear dynamic range for low-abundance peptides. This allows for reliable quantification at concentrations previously below the detection threshold.

D. Biodosimetry Application

• In a study estimating ionizing radiation exposure in mice skin tissue, Nettle improved the performance of a regression model.
• Metrics: Compared to KNN imputation, Nettle reduced the Median Average Error (MAE) of predicted radiation dose from 13.69 to 10.02 and increased the Pearson correlation coefficient from 0.55 to 0.76.

5. Significance

• Overcoming Missingness: Nettle provides a robust solution to the pervasive problem of missing values in DIA proteomics without introducing statistical artifacts.
• Enhanced Sensitivity: By recovering quantitations for low-abundance peptides, it significantly improves the detection of biomarkers and differential expression in complex biological samples.
• Practical Utility: The method is particularly valuable for studies with small sample sizes or precious clinical samples where losing data points due to missingness is not an option.
• Future Impact: The authors suggest Nettle will be crucial for scaling DIA to large cohorts, single-cell proteomics, and biomarker discovery in fields where traditional pipelines fail due to data sparsity.

In summary, Nettle represents a paradigm shift from statistical imputation of intensities to the imputation of physical chromatographic boundaries, yielding more accurate, biologically relevant, and statistically robust proteomic data.