Innovative methods for data analysis in analytical chemistry using Bayesian statistics and machine learning

In analytical chemistry, rapid advancement in instrumentation, especially in high resolution mass-spectrometry is making a significant contribution for further developments of the field. As such, in separation science, nowadays, several hyphenated techniques have proven to be the state-of-the-art techniques for compound identification and characterization. Furthermore, these techniques have found several application areas including biomarker discovery, forensic investigation, food-health research and many others, mainly because of their high sensitivity, selectivity, and scope of analysis. As these techniques advance, the amount of data generated by the high resolution instruments has also increased tremendously, requiring a more sophisticated data analysis techniques to cope with the ‘big data’, and that can simultaneously extract and utilize the information. Taking into account the benefit of instrument coupling (i.e. LC-HRMS) for better separation and characterization of compounds has already been proven to be the most efficient approach, in the years to come, it can be speculated that even more sophisticated techniques such as liquid chromatography coupled to ion mobility spectrometry, and high resolution mass spectrometer (LC-IMS-HRMS) will be the dominating techniques, expanding the dimensionality and the complexity of the data produced. Thus, more than ever, it’s crucial for data analysis techniques to advance a lot faster to manage the computational challenges, and so as to harness the full potential of these promising analytical techniques for routine analysis.
The project of this thesis was a joint effort of industry partners involved in forensic, food-safety and material science, and had a common goal and interest in the development of robust data analysis techniques for compound screening. As such, the main aim of this thesis was directed towards improving the conventional data analysis techniques in regards to peak detection, and compound identification for both targeted and untargeted screening. Given this aim, novel algorithms based on probabilistic machine learning, demonstrated to be a better alternative in comparison to conventional threshold-based approaches were developed, and validated.