ChromPeakNet: Deep Learning for Chromatographic Peak Detection

Feature detection is a critical first step in untargeted metabolomics workflows. Given a raw LC-MS file containing hundreds of scans, each with thousands of mass-to-charge (m/z) measurements, the goal is to extract a list of chromatographic peaks. The quality of this initial feature list directly constrains all downstream analysis: compound identification, quantification, statistical comparisons, and biological interpretation.

The three dominant open-source tools---MZmine, OpenMS, and XCMS---all rely on hand-crafted signal processing heuristics (wavelet transforms, local minimum resolvers, isotope pattern fitting) that require extensive parameter tuning for each experimental setup and generalize poorly when applied to data with different noise characteristics.

The gap between simulated and real-world performance is particularly revealing. On the mzrtsim benchmark, MZmine achieves F1 = 0.937, but on real experimental data (sep1), it drops to F1 = 0.476---a degradation of approximately 50%. XCMS fails entirely on the real dataset due to systematic m/z calibration errors.

This paper introduces ChromPeakNet, a multi-task 1D U-Net that learns peak detection, apex localization, and baseline estimation jointly from synthetic chromatographic training data. By combining ChromPeakNet with the same ADAP chromatogram builder used by MZmine, we ensure a fair comparison in which the only difference is the peak detection algorithm.

2.1 ChromPeakNet Architecture

ChromPeakNet is a 1D U-Net with three output heads. The model takes three-channel input: the raw EIC intensity trace, its first derivative, and second derivative. All channels are zero-padded to length 8192. The derivative channels provide the model with explicit slope and curvature information that is informative for peak boundary detection and apex localization.

2.2 Multi-Task Training Objective

The model is trained on synthetically generated chromatographic data with known ground truth. The multi-task loss combines four terms: focal loss and Dice loss for the peak mask, focal loss for apex localization, and Huber loss for baseline estimation.

The focal loss addresses class imbalance between peak and non-peak regions (γ = 2.0), while the Dice loss provides a region-level overlap objective. Training uses AdamW with learning rate 3×10^-4 and cosine annealing over 50 epochs. Total training time is approximately 24 minutes on an NVIDIA RTX A4000.

ChromPeakNet achieves state-of-the-art F1 on both simulated and real data, with notably stronger generalization to real experimental conditions. The advantage is dramatically larger on real data, where ChromPeakNet achieves F1 = 0.889 compared to MZmine's 0.476---an improvement of 41.3 percentage points.

Training Convergence

ChromPeakNet converges rapidly, reaching near-optimal performance within 20 epochs. The multi-task loss decreases from 0.312 to 0.069 over 50 epochs, with the Dice component dominating the learning signal and apex and baseline losses converging to near-zero within 5 epochs.

Not all 593 ground-truth peaks are physically detectable from the raw data. We investigated every missed peak and found that 60 of 593 have zero signal---no centroids exist within 10 ppm of the ground-truth m/z at the expected retention time. This establishes a detection ceiling of approximately 533 peaks, corresponding to a maximum achievable F1 of approximately 0.947.

We ablate the deduplication parameters to quantify their effect on final performance. Changing the deduplication radius from 5 ppm/3 s to the final 3 ppm/2 s gains 5 TP on simulated data and 8 TP on sep1 real data, by preserving closely-spaced genuine peaks that were previously merged.

Design Rationale

Simulated Data (mzrtsim)

Of the 64 missed peaks: 60 (93.8%) have zero signal in the raw data (physically undetectable by any tool). 3 (4.7%) involve multi-peak ambiguity where two ground-truth peaks are separated by less than the EIC resolution. 1 (1.6%) is lost to deduplication.

Real Data (sep1)

The 8 false positives on sep1 correspond to real chromatographic features (minor adducts and in-source fragments) not annotated in the ground truth. These are arguably true features from an untargeted metabolomics perspective, where the goal is comprehensive feature extraction.

Why Traditional Tools Underperform on Real Data

We presented ChromPeakNet, a deep learning approach to chromatographic peak detection that achieves state-of-the-art on both simulated and real LC-MS benchmarks. Key findings:

1. On the mzrtsim simulated benchmark, ChromPeakNet reaches F1 = 0.943 with zero false positives, surpassing MZmine 4.5 (F1 = 0.937) and detecting 99.2% of all physically detectable peaks.
2. On the sep1 real experimental benchmark, ChromPeakNet achieves F1 = 0.889 with perfect recall (32/32 peaks), compared to MZmine (0.476), OpenMS (0.720), and XCMS (0.000).
3. The detection ceiling analysis reveals that 60 of 593 ground-truth peaks have zero signal, establishing a physical upper bound of F1 ≈ 0.947. ChromPeakNet's performance of 0.943 lies within 0.4% of this ceiling.
4. Trained exclusively on synthetic data, ChromPeakNet generalizes to real separations with less than 6% F1 degradation, compared to 23-100% for traditional tools.

These results demonstrate that lightweight learned models can replace hand-crafted heuristics for chromatographic peak detection, with improved robustness to real-world signal variability.