Learning to See Peaks: Attention-Based Feature Extraction for Automated Chromatographic Peak Detection.

Area of Science:

• Analytical Chemistry
• Biopharmaceutical Analysis
• Chromatography

Background:

• Size-exclusion chromatography (SEC) is crucial for biopharmaceutical release and comparability assays.
• Peak detection in SEC faces challenges like signal overlap, baseline drift, and analyst variability, limiting reproducibility.
• Accurate peak interpretation relies on morphology and context, making automated methods desirable.

Purpose of the Study:

• To develop and evaluate a machine learning-based method for robust and reproducible peak detection in SEC.
• To improve the accuracy and consistency of peak identification in routine analytical workflows for large molecules.
• To provide an extensible framework for SEC peak analysis that minimizes analyst-dependent variability.

Main Methods:

• Development of Peak Feature Extractor 1 (PFE-1), a transformer-based model trained on simulated SEC chromatograms.
• Generation of synthetic chromatograms using a simulator statistically calibrated to real antibody SEC data.
• Evaluation of PFE-1 on synthetic benchmarks and a curated real SEC dataset using precision, recall, F1 scores, and intensity-weighted box loss.

Main Results:

• PFE-1 significantly outperforms convolutional and derivative-based baseline methods in peak detection accuracy.
• The model demonstrates superior performance, especially under challenging conditions with overlapping peaks and complex morphologies.
• On a real SEC benchmark, PFE-1 achieved the highest box-level agreement without requiring sample-specific inputs.

Conclusions:

• PFE-1 offers a reproducible and extensible solution for chromatographic peak detection in SEC.
• The developed framework enhances peak interpretation consistency for routine analytical workflows.
• Machine learning, specifically transformer models, shows great promise for overcoming limitations in SEC data analysis.