UC Riverside

The identification of unknown small molecules from tandem mass spectrometry (MS/MS)data is a fundamental challenge in the chemical and biological sciences, as a significant portion of spectra from untargeted experiments remain unidentified. While computational techniques like analog search can suggest structurally similar compounds, they fail to specify the exact location of chemical modifications, leaving a critical gap in the annotation process. To surmount this issue, this dissertation investigates the problem of structural annotation of small molecules using tandem mass spectrometry (MS/MS) data, with a particular focus on leveraging information from known similar molecules known as analogs and presents a cohesive suite of computational tools developed to address this challenge, progressing from targeted localization to full structure generation.First, ModiFinder is introduced, a novel algorithm that localizes structural modification sites by aligning the MS/MS spectra of a known molecule and its unknown analog. By systematically analyzing shifted and unshifted fragment peaks, ModiFinder generates a likelihood score for each atom in the known structure being the modification site, successfully outperforming baseline methods. Then, the robustness and applicability of this approach are significantly enhanced by incorporating richer data from multiple collision energies and mass spectrometry adducts. This multispectrum approach expands ModiFinder's utility to a wider range of compounds by providing complementary fragmentation information.Finally, the more ambitious challenge of de novo structure elucidation is addressed by developing ModiStruct, a state-of-the-art generative model. This model reframes the problem as a diffusion-driven molecular inpainting task, where it reconstructs the complete molecular graph of an unknown compound conditioned on a known substructure, its molecular formula, and their respective MS/MS spectra. ModiStruct achieves state-of-the-art accuracy on public benchmarks, validating the power of using structural priors within a generative framework. Collectively, this work provides a powerful set of tools that advance the automated identification of small molecules, from pinpointing specific chemical changes to generating entire novel structures. These contributions help to bridge the gap between MS/MS data and chemical knowledge, with broad applications in metabolomics, drug discovery, and natural products research.