Evaluating a Lightweight Neural Network for Aggregate Isotope Distribution Prediction

Abstract

Abstract Isotope distribution prediction is an important part of mass spectrometry data analysis. A variety of strategies have been developed, including brute-force polynomial methods and Fourier-transform (FT) convolution methods. Here, we present a novel neural network (NN) approach to isotope prediction. These NN-based tools are distributed in a new package, IsoGen, alongside FT-based tools. We show that NNs perform as well as existing approaches when predicting isotope distributions for molecules with natural isotope ratios. We then demonstrate the capability of NNs for transfer learning, a method by which existing models that have been trained to perform one task can be reused as a starting point to train models for a second task. Here, an NN that has been trained to predict isotope distributions for molecules with natural isotope ratios can be retrained to predict distributions for molecules with perturbed isotope ratios. We also demonstrate that these training distributions do not need to be produced theoretically but can be extracted from experimental data in which the underlying isotopic composition is not known. Overall, NNs provide a robust method for isotope prediction that can be extended to applications in which the isotope distributions can be measured but not easily predefined.