Training Deep Learning Neural Networks for Predicting CCS Using the METLIN-CCS Dataset

Collision Cross Section is the probability of a collision taking place between an ion and a buffer gas. Unlike retention times, which are strongly affected by the experimental setup, collision cross section values tend to remain consistent across experiments. This makes them a useful piece of information orthogonal to m/z and retention times during the metabolite annotation process. Due to the high cost in time and money for the experimental measurement of collision cross section values, their computational prediction from a compound’s chemical formula provides valuable support for metabolite annotation.The METLIN-CCS dataset, containing 61,863 experimentally measured collision cross section values, has recently been published. This represents a significant increase in size compared to previously available datasets. In this work, we use this dataset to train several machine learning models (deep learning neural networks, gradient boosting and an ensemble) to predict the collision cross section of a compound. Several representations of the compounds will be tested: fingerprints, descriptors, and both fingerprints and descriptors combined. These features were calculated with alvaDesc™. We employed Bayesian hyperparameter optimization for tuning the machine learning models. The best results were obtained with a highly regularized deep learning neural network trained with fingerprint and descriptors, obtaining a mean relative error of 3.20 ± 0.05 % in a 5-fold cross-validation.