[Paper Review] Peak Detection On Data Independent Acquisition Mass Spectrometry Data With Semisupervised Convolutional Transformers

Leon L. Xu, Hannes Röst|arXiv (Cornell University)|Oct 26, 2020

Mass Spectrometry Techniques and Applications4 citations

TL;DR

This paper proposes a semisupervised convolutional Transformer model for peak detection in data-independent acquisition (DIA) LC-MS data, combining CNNs for local feature extraction with self-attention mechanisms to capture long-range dependencies across multivariate time series. The method achieves state-of-the-art performance on a benchmark DIA dataset, outperforming baseline neural networks and matching or exceeding existing automated peak detection tools using manual annotations as ground truth.

ABSTRACT

Liquid Chromatography coupled to Mass Spectrometry (LC-MS) based methods are commonly used for high-throughput, quantitative measurements of the proteome (i.e. the set of all proteins in a sample at a given time). Targeted LC-MS produces data in the form of a two-dimensional time series spectrum, with the mass to charge ratio of analytes (m/z) on one axis, and the retention time from the chromatography on the other. The elution of a peptide of interest produces highly specific patterns across multiple fragment ion traces (extracted ion chromatograms, or XICs). In this paper, we formulate this peak detection problem as a multivariate time series segmentation problem, and propose a novel approach based on the Transformer architecture. Here we augment Transformers, which are capable of capturing long distance dependencies with a global view, with Convolutional Neural Networks (CNNs), which can capture local context important to the task at hand, in the form of Transformers with Convolutional Self-Attention. We further train this model in a semisupervised manner by adapting state of the art semisupervised image classification techniques for multi-channel time series data. Experiments on a representative LC-MS dataset are benchmarked using manual annotations to showcase the encouraging performance of our method; it outperforms baseline neural network architectures and is competitive against the current state of the art in automated peak detection.

Motivation & Objective

• To address the challenge of accurate peak detection in complex, high-dimensional DIA LC-MS data, which is critical for quantitative proteomics.
• To improve upon existing deep learning models by integrating local contextual modeling via convolutional layers with global sequence modeling via self-attention in Transformers.
• To leverage limited labeled data effectively by adapting semisupervised learning techniques from computer vision to multivariate time series in proteomics.
• To develop a method that outperforms standard neural network architectures and matches or exceeds the state of the art in automated peak detection on real-world DIA datasets.

Proposed method

• The model uses a hybrid architecture combining convolutional layers with self-attention mechanisms, termed 'Convolutional Self-Attention' to jointly model local and long-range dependencies in multivariate time series.
• It formulates peak detection as a multivariate time series segmentation task, where each channel corresponds to an extracted ion chromatogram (XIC) of a specific m/z value.
• The model is trained in a semisupervised manner, using a combination of labeled data (manually annotated peaks) and large amounts of unlabeled data to improve generalization.
• The authors adapt state-of-the-art semisupervised image classification techniques—specifically consistency regularization and pseudo-labeling—to the multichannel time series setting.
• The architecture employs multi-head attention with learnable positional encodings to model temporal relationships across retention time points.
• The final output is a sequence of peak segmentations per XIC, with confidence scores derived from the model's attention and classification heads.

Experimental results

Research questions

• RQ1 Can a hybrid convolutional-Transformer architecture improve peak detection performance in DIA LC-MS data compared to standard deep learning models?
• RQ2 To what extent can semisupervised learning techniques from computer vision be effectively transferred to multivariate time series in proteomics?
• RQ3 Does the integration of local (convolutional) and global (self-attention) modeling improve detection of complex, overlapping peptide signals in DIA data?
• RQ4 How does the proposed method compare to existing state-of-the-art peak detection tools in terms of precision, recall, and F1-score on a benchmark dataset?

Key findings

• The proposed method outperforms baseline neural network architectures in peak detection accuracy on a representative DIA LC-MS dataset.
• It achieves competitive performance compared to the current state of the art in automated peak detection, as measured by F1-score against manual annotations.
• The semisupervised training strategy significantly improves model generalization, especially when labeled data is limited.
• The integration of convolutional layers with self-attention enables better modeling of both local signal patterns and long-range dependencies across retention time.
• The model demonstrates robustness to noise and overlapping peaks, common challenges in DIA data.
• The ablation study confirms that both the convolutional and self-attention components contribute meaningfully to the final performance.