Mining structural information in gas chromatography-mass spectrometry data for analytical-descriptor-based quantitative structure–activity relationship

Abstract

Recently, a new approach to quantitative structure–activity relationship (QSAR) has been proposed, which employs machine learning techniques and uses analytical signals from the full scan of mass spectra as input. Unlike traditional QSAR, this approach does not need exhaustive structural determination to assess numerous unknown compounds. The new approach assumes that a mass spectral pattern reflects the structure of a target chemical. However, despite the remarkable performance of this method, the relationship between the spectrum and the structure is complex and its interpretation is a challenge to the further development of QSAR based on analytical signals. This study explored whether gas chromatography-mass spectrometry (GC-MS) data contain meaningful structural information that is advantageous for QSAR prediction by comparing it with the traditional molecular descriptor used in QSAR prediction. Chemical groups were assigned to each chemical linked to the GC-MS data and molecular descriptor dataset to investigate their relationships. Then, data clustering was performed by t-distributed stochastic neighbor embedding on the GC-MS data (i.e., analytical descriptor) and on four molecular descriptors: ECFP6, topological descriptor in CDK, MACCS key, and PubChem fingerprint. Although the chemicals represented by the analytical descriptor were not clearly clustered according to the chemical class, most clusters were formed by chemicals with similar spectrum patterns. An additional investigation suggested that the analytical and molecular descriptors preserved structural information in different ways. The predictive performance of QSAR based on analytical and molecular descriptors was evaluated in terms of molecular weight, log Ko−w, boiling point, melting point, vapor pressure, water solubility, and two oral toxicities in rats and mice. The analytical- and molecular-descriptor-based models performed comparably. The influential variables in the analytical-descriptor-based model were further investigated by comparing analytical-descriptor-based and linear regression models using simple indicators of the mass spectrum. In general, the analytical-descriptor-based approach predicted the physicochemical properties and toxicities of structurally unknown chemicals that the molecular-descriptor-based one could not. These results suggest that the new approach is valuable for evaluating unknown chemicals in many scenarios.