Rapid Prediction of Electron-Ionization Mass Spectrometry Using Neural Networks

Abstract

When confronted with a substance of unknown identity, researchers often perform mass spectrometry on the sample and compare the observed spectrum to a library of previously collected spectra to identify the molecule. While popular, this approach will fail to identify molecules that are not in the existing library. In response, we propose to improve the library's coverage by augmenting it with synthetic spectra that are predicted from candidate molecules using machine learning. We contribute a lightweight neural network model that quickly predicts mass spectra for small molecules, averaging 5 ms per molecule with a recall-at-10 accuracy of 91.8%. Achieving high-accuracy predictions requires a novel neural network architecture that is designed to capture typical fragmentation patterns from electron ionization. We analyze the effects of our modeling innovations on library matching performance and compare our models to prior machine-learning-based work on spectrum prediction.

Figure 1

Library matching task. (a) Depiction of how query spectra are matched to a collection of reference spectra as performed by mass spectrometry software. (b) Query spectra are compared against a library comprising spectra from the NIST 2017 main library and spectra predicted by our model (outlined in blue). Spectral images adapted from NIST Webbook.

Figure 2

Spectral prediction with MLP forward model (a) and MLP bidirectional model (b). For both spectra plots, the true spectrum is shown in blue on top, while the predicted spectrum is shown inverted in red. Note that the spectrum predicted by the bidirectional model shows fewer stray peaks than the forward model, particularly for larger m/z values. These peaks are much easier to predict with the reverse prediction mode.

Figure 3

Molecular representations are passed into a multilayer perceptron to generate an initial output. This output is used to make a forward prediction starting at and and in reverse starting from and ending at . A sigmoid gating is applied to the inputs as shown in eq 5.

Figure 4

Performance of different model architectures. (a) Comparison of the recall@10 accuracy of the linear regression model and the multilayer perceptron model using the forward, reverse, and bidirectional architecture. (b) Performance of NEIMS at different recall levels, and its comparison against the performance of using the NIST main library itself.

Figure 5

Comparing the similarity between the predicted spectrum and the ground truth spectrum to the overall similarity between spectra for the same molecule.