UC Davis

Understanding metabolism is essential for all living organisms [1] . The science ofstudying all metabolites and lipids is metabolomics [1] . One of the most important aims in metabolomics is to identify all detectable structures; yet, the enormous chemical space of natural products and enzymes makes this aim a moving target. Mass spectrometry (MS/MS) is the most commonly used method to identify compounds due to its high sensitivity, selectivity and the information content of mass spectra that come from breaking chemical bonds [2] . However, compounds can only be identified if their mass spectra are included in MS/MS libraries. Compared to over 100 million compounds presented in the PubChem database, MS/MS libraries are small with less than 200,000 structures. Computational chemistry and cheminformatics may bridge this gap by predicting mass spectra from chemical structures (structure-to-spectrum), or by fitting experimental MS/MS information to lists of candidate structures (spectrum-to-structure). For structure-to-spectrum approaches, the (1) a rule-based heuristic methods work fast and efficient if rules can be learned and expressed for wide ranges of compound classes. A successful example are lipids, where MS/MS spectra are mainly governed by loss of head groups and loss of fatty acyl groups [3] . (2) Machine learning algorithm-based are promising but require a very large set of experimental MS/MS spectra for training and validation sets. Examples such as CFM-ID currently lack predictive accuracy [4] . (3) Quantum chemistry based methods, in principle, offer a solution by applying first-principles in physical chemistry to the problem of breaking chemical bonds and assessing the stability of product ions. While there are published examples and tools, these have not been widely tested or explored. Therefore, this dissertation presents a collection of computational chemistry work ranging from analyzing and evaluating quantum chemistry-based software, performing molecular dynamic simulations, investigating mechanistic details of fragmentation reactions, and developing a computational framework. This framework is now called CIDMD, a quantum chemistry-based structure-to-spectrum approach in computational metabolomics, and the thesis explores its analysis package. With thorough investigations and evaluations, this dissertation offers insights into modeling in silico mass spectra and better understanding towards a better approach and design in the future. Chapter 1 provides a brief introduction to quantum chemistry and describes the foundation of collision-induced dissociation process that governs CID-MS/MS spectra. The current state of MS/MS predictions for compound identification are summarized. Chapter 2 presents the application of the quantum chemistry electron-ionization package, the QCEIMS software developed by Dr. Stefan Grimme, Bonn/Germany. Chapter 2 describes the accuracy of QCEIMS software for predicting EIMS spectra of 80 purines and pyrimidines. Chapter 3 introduces the computational framework CIDMD (collision-induced-dissociation molecular dynamics). CIDMD is a quantum chemistry-based method combined with a structure-to- spectrum approach. Chapter 3 showcases the application of CIDMD to 12 molecules and evaluates the accuracy and information from CIDMD predicted theoretical mass spectra. Specifically, CIDMD predicted fragmentation reaction pathways yielded mechanistic details in MS/MS fragmentations that may lead to future improvements in the software. Importantly, we found that the initial structural protomer used as input into CIDMD calculations may largely affect the resulting CIDMD mass spectra. Consequently, this thesis ends by studying CIDMD fragmentation of protomers in chapter 4 in relation to corresponding experimental MS/MS spectra.