ChromAlignNet: Deep Learning Peak Alignment

• ChromAlignNet is a deep learning framework that employs a modular Siamese network to align GC–MS peaks without relying on reference chromatograms.
• Its specialized encoders for mass spectrum, peak shape, and chromatographic segments achieve high performance, with AUC up to 1.0 on simple datasets.
• The approach outperforms traditional rule-based methods, offering robust adaptability and scalability for analyzing complex GC–MS data.

ChromAlignNet is a deep learning framework devised for the alignment of peaks in Gas Chromatography-Mass Spectrometry (GC–MS) data. Accurate peak alignment is essential for quantitative and comparative analysis in GC–MS, as retention times (RT) of identical analytes may drift across chromatograms due to nonlinear and complex factors. ChromAlignNet proposes a solution rooted in modern neural network approaches, designed to address challenges where existing rule-based algorithms are subject to parameter sensitivity and limited scalability with data complexity. The model has demonstrated high quantitative performance, notably achieving Area Under the Curve (AUC) values approaching 1.0 for simple data sets and around 0.85 for very complex cases.

1. Network Architecture

ChromAlignNet utilizes a modular Siamese neural network structure that compares pairs of chromatographic peaks to determine if they originate from the same compound. The architecture consists of three parallel encoders, each responsible for one modality of chromatographic data:

• Mass Spectrum Encoder: Processes mass spectral data at the peak's maximum intensity using a fully connected network (two hidden layers of 64 neurons each, dropout, followed by a final 10-neuron layer), resulting in a compact 10-dimensional feature vector.
• Peak Profile Encoder: Encodes the peak's shape via a sequence of bidirectional LSTM layers with dropout, culminating in another 10-dimensional feature vector.
• Chromatogram Segment Encoder: Analyzes a segment of the chromatogram centered on the peak (±1.5 minutes) through two parallel stacks of 1D convolutional networks with varying depth and pooling, each yielding a 10-dimensional representation.

Each encoder processes its respective input for two peaks. The output vectors, $f(x_1)$ and $f(x_2)$, are compared via absolute difference $d = |f(x_1) - f(x_2)|$, producing a concatenated 30-dimensional vector. The absolute RT difference is appended, the result is passed through a dropout layer (rate 20%), a fully connected 64-neuron ReLU layer, and a single sigmoid neuron outputs a matching probability.

Encoder	Input Modality	Output Dimensionality
Mass Spectrum	Mass spectrum at apex	10
Peak Profile	Intensity over peak	10
Chromatogram Segment	±1.5 min chromatogram	10

This structure ensures processing symmetry and modular feature learning from each data aspect, supporting generalizability and robustness across chromatographic data sets of varying complexity.

2. Training and Quantitative Evaluation

ChromAlignNet is trained on manually curated peak pairs from various air and breath sample GC–MS data sets. The training strategy balances "positive pairs" (peaks from the same compound) with randomly subsampled "negative pairs" (different compounds). Peak selection is performed within specific m/z channels and retention windows. Training employs GPUs or CPU clusters, with epochs ranging up to 50 and durations from 2 to 10 hours depending on the set size. Validation splits (typically 20%) are used to mitigate overfitting.

Performance evaluation uses ROC curves with AUC as the primary metric. ChromAlignNet achieves near-perfect AUC ($\sim1$) on simple ambient air data sets and AUC values between 0.85 and 0.95 on complex breath data. For clustering, hierarchical average linkage is applied to pairwise probabilities, with an RT difference cutoff to form aligned peak clusters. In particularly challenging data, true positive (TP) rates approach 0.96, substantially exceeding rates from traditional algorithms.

3. Comparison with Rule-Based Methods

Traditional peak alignment strategies such as Correlation Optimized Warping (COW) and GCalignR rely on selected reference chromatograms and manually tuned shift parameters. Their alignment performance is sensitive to these parameters and often struggles with complex retention time shifts. ChromAlignNet's approach requires no reference chromatogram and minimal parameterization (only an RT cutoff during prediction). The model leverages mass spectra, peak shape, and chromatographic context automatically.

In comparative tests:

Data Set	ChromAlignNet TP Rate	COW TP Rate	GCalignR TP Rate
Field Breath	~0.96	0.56	0.48

On simple sets, TP and false positive (FP) rates are comparable, whereas on complex sets, ChromAlignNet substantially outperforms symbolic methods. A noted limitation is a tendency to merge adjacent peaks in cluster formation, driven by residual false positives in high-complexity data.

4. Generalization and Adaptability

ChromAlignNet's architecture is agnostic regarding input modality, aside from basic chromatographic structure. Adaptation to other modalities such as liquid chromatography (LC) and nuclear magnetic resonance (NMR) spectroscopy is feasible by retraining on the new data while preserving the encoder structure. Adjustments may involve input preprocessing (scaling, window choices) rather than architectural redesign. The approach thus provides a generalizable framework for peak alignment tasks in analytical chemistry beyond GC–MS.

5. Implementation Details and Computational Requirements

The implementation leverages Python 3.6 with Keras for neural network components. Hierarchical clustering utilizes SciPy, while evaluation employs scikit-learn routines. Instrument XML to CSV conversion and peak detection processes are implemented in MATLAB, with codebases made publicly available:

• Core network and clustering: https://github.com/mili7522/ChromAlignNet
• Peak detection: https://github.com/rosalind-wang/GCPeakDetection

Training is performed on GPU clusters (up to 4 Nvidia P100 GPUs) and CPU clusters for batch operations. Prediction for individual chromatograms is feasible on standard desktop processors.

6. Prospects for Improvement

Acknowledged areas for advancement include the peak profile encoder, which presents training difficulties due to variable peak length and profile. Future research may involve pretraining as an autoencoder or reformulating with CNN-based sequence encoders. For the mass spectrum encoder, selective input or attention to informative m/z values is suggested to enhance specificity. High FP rates in complex sets motivate exploration of alternative clustering schemes and loss functions (e.g., triplet loss). Network enhancements such as expanded filter counts and mitigation of overfitting via more refined dropout strategies are under consideration.

A plausible implication is that with such improvements, ChromAlignNet may further reduce manual intervention and increase reliability for biomarker discovery in heterogeneous and noisy chromatographic data.

7. Summary and Significance

ChromAlignNet represents a principled, deep learning-enabled methodology for chromatographic peak alignment. By combining discrete feature encoding branches within a unified Siamese structure and dispensing with the need for reference chromatograms or parameter tuning, it offers robust performance across diverse GC–MS data sets. The framework’s generalizability, computational efficiency, and extensibility suggest its utility as a foundational alignment solution for various modalities of analytical chemistry data, with practical implications for large-scale biomarker identification and comparative studies.