NMR Spectrum Classification of Pharmaceutical Compounds Using Convolutional Neural Network (CNN)

Abstract

Nuclear Magnetic Resonance (NMR) spectroscopy is a crucial technique for identifying the structure of pharmaceutical compounds. However, manual interpretation of NMR spectra is time-consuming and requires expert knowledge. This study aims to develop an automated classification model using Convolutional Neural Network (CNN) applied to 1H-NMR spectra converted into 2D image representations. The dataset includes 200 spectra from five compound classes (alkaloids, flavonoids, steroids, antibiotics, and amino acids), sourced from public databases (NMRShiftDB and PubChem). The CNN model consists of two convolutional and max-pooling layers and was evaluated using multiclass classification metrics. The results showed an accuracy of 92.3% and an average F1-score of 90.9%, outperforming baseline models such as KNN and Random Forest. ANOVA analysis revealed a statistically significant difference between the models (p < 0.05). This study demonstrates that CNN is effective for rapid and automated classification of pharmaceutical compounds based on spectral patterns.