Deep representation learning for Nuclear Magnetic Resonance spectral clustering

Abstract

Nuclear Magnetic Resonance (NMR) spectroscopy is essential for molecular structure elucidation, drug discovery, and biomedical research, as it enables the identification of spectral patterns and compound similarities. However, traditional clustering methods struggle with robustness against chemical shift variations, peak intensity fluctuations, and noise, limiting their effectiveness in complex NMR datasets. To address these challenges, we propose a novel framework that combines an attention mechanism with a bidirectional long short-term memory autoencoder. This approach extracts robust, low-dimensional representations of NMR spectra by integrating adaptive local feature extraction with deep representation learning. Our method enables effective clustering across diverse spectral regions while preserving critical chemical information. Comprehensive experiments on both synthetic and real-world datasets demonstrate that this framework significantly outperforms conventional techniques in clustering accuracy and representation quality. These findings highlight its practical utility for enhancing NMR spectral analysis and molecular characterization.