How the (data) rich get richer: machine learning as a key driver of MS informatics

Computational proteomics has evolved extensively over the past decade, introducing the first widely successful machine learning approach with the Percolator algorithm for post-processing of identification data, and slowly expanding the capabilities of predictive algorithms for analyte behaviour during liquid chromatography separation and fragmentation. Despite clear advances in these areas, however, the adoption of such predictions in identifications pipelines was very slow to non-existent. This was primarily due to the already quite capable traditional identification algorithms, especially when complemented with Percolator post-processing. The advent of data independent acquisition (DIA) helped to renew interest in these approaches, however, as this approach’s reliance on experimentally-derived spectrum and chromatogram libraries was inherently hampered by the limitations in the data dependent acquisition (DDA) methods used to acquire these libraries. By creating purely predicted libraries, this limitation of having to rely solely on previously observed and identified signals could be overcome. Meanwhile, renewed interest was also sparked in DDA proteomics, as the field had started to diversify into more complex analyses such as metaproteomics, proteogenomics, immunopeptidomics, and open modification searches, all of which revealed inherent limitations in existing, traditional search engines. As it turns out, machine learning based predictions have shown to be highly effective at solving the issues encountered in DIA and complex DDA proteomics approaches, and the production of highly performant algorithms has soared as a consequence. Interestingly, the availability of vast amounts of public data have also enabled a breakthrough in the types of machine learning algorithms that can now be employed on proteomics data, which has seen a surge in the application of complex neural networks (so-called deep learning approaches) in the field. When provided with enough data, these deep learning algorithms deliver extremely good predictions, which will in turn fuel a more sensitive and more robust interpretation of already acquired data. I will here present our latest, cutting-edge developments in machine learning based algorithms for computational proteomics, as well as our applications of these algorithms to the complete re-analysis of all publicly available human proteomics data, to DIA data interpretation, and to COVID-19 assay optimisation.