Machine Learning for Peaks Detection in Nuclear Magnetic Resonance Spectra

Abstract

This research addresses the challenge of accurately detecting and automating the peak picking process in both pure and mixture Nuclear Magnetic Resonance (NMR) spectra. Peak picking is a crucial step in NMR analysis, but manual methods are often time-consuming and prone to errors, particularly in complex mixture spectra. Recent advancements in machine learning provide an opportunity to automate this process, improving both efficiency and accuracy; however, many of those methods focus more on one type of peak than the other and still require pre-processing steps. A machine learning system was developed to automate the detection and extraction of peaks in both pure and mixture NMR spectra, and it was systematically evaluated against several established techniques. The approach tackles key issues, including enhancing the detection of small peaks, and overlapping peaks, and managing the limited availability of labeled training data by generating synthetic datasets. Despite being trained on synthetic data, the model demonstrated strong performance on real NMR spectra, effectively automating peak detection. The model employs well-established machine learning techniques for object detection and segmentation, achieving 97% accuracy on synthetic data with no missed detections and few false positives, and 92% accuracy on real data. These results, compared to existing methods, suggest that the automated system can improve the accuracy and efficiency of peak picking in both pure and mixture NMR spectra, providing a valuable tool for researchers and practitioners in the field.