Enhancing Peak Assignment in CNMR Spectroscopy: A Novel Approach Using Multimodal Alignment

Abstract

Nuclear magnetic resonance (NMR) spectroscopy is pivotal in unraveling molecular structures and dynamic behaviors. Although machine learning models show promise in NMR spectral prediction, challenges persist in peak assignment, a crucial step in molecular structure determination. Addressing this, our paper presents a pioneering approach, multimodal alignment correlating CNMR spectral peaks (presented in a sequence data format) with their corresponding atoms in molecular structures (presented in graph data format). This solution establishes correspondences across two heterogeneous modalities: molecular graph and spectral sequence. It employs a dual-coordinated contrastive learning architecture featuring three key modules: a molecular-level alignment module, an atomic-level alignment module, and a communication channel. Our approach yields exceptional results, boasting a peak-to-atom match rate exceeding 90% for exact matches. Additionally, it achieves a remarkable accuracy of over 95% in assigning CNMR spectra to molecules, thus making a significant contribution to isomer recognition.