Advancing Molecular Structure Elucidation for Mass Spectrum Interpretation via Cepstral Analysis

In recent years, artificial intelligence has significantly advanced drug discovery through developments in molecular property prediction, generative models for de novo molecular design, and molecular optimization algorithms. However, a substantial gap remains between computational design and experimental realization, as proposed molecules must be synthesized and experimentally validated to verify their structures. To achieve reliable molecular validation, mass spectrometry has become one of the most widely adopted analytical techniques, yet the interpretation of mass spectra still relies heavily on expert chemists and extensive manual analysis. This reliance creates a major bottleneck in the molecular design-to-validation pipeline. In this work, we propose a dual-representation framework that leverages cepstral analysis to enhance the molecular structure elucidation in AI-driven mass spectrum interpretation. By transforming the original mass spectrum into its cepstrum representation through a cepstral analysis technique, our approach provides complementary structural information in the quefrency domain that captures periodic patterns and envelope characteristics. We concatenate both representations as dual inputs to a transformerbased architecture, enriching the feature space for learning the complex structure-spectrum relationship. Experimental results on the MassSpecGym benchmark demonstrate that our multirepresentation strategy significantly outperforms conventional single-representation methods, highlighting the importance of data representation for mass spectrum interpretation.