Abstract
The application of machine learning (ML) to -omics research is growing at an exponential rate owing to the increasing availability of large amounts of data for model training. Specifically, in metabolomics, ML has enabled the prediction of tandem mass spectrometry and retention time data. More recently, due to the advent of ion mobility, new ML models have been introduced for collision cross-section (CCS) prediction, but those have been trained with different and relatively small data sets covering a few thousands of small molecules, which hampers their systematic comparison. Here, we compared four existing ML-based CCS prediction models and their capacity to predict CCS values using the recently introduced METLIN-CCS data set. We also compared them with simple linear models and with ML models that used fingerprints as regressors. We analyzed the role of structural diversity of the data on which the ML models are trained with and explored the practical application of these models for metabolite annotation using CCS values. Results showed a limited capability of the existing models to achieve the necessary accuracy to be adopted for routine metabolomics analysis. We showed that for a particular molecule, this accuracy could only be improved when models were trained with a large number of structurally similar counterparts. Therefore, we suggest that current annotation capabilities will only be significantly altered with models trained with heterogeneous data sets composed of large homogeneous hubs of structurally similar molecules to those being predicted.