Abstract

Obtaining information about the molecular structure from the mass spectra is one of the most pursued challenges in non-targeted analysis. The complete solution to the problem has not been found yet, therefore only partial information about the structure can be obtained from mass spectra, often in the form of various molecular fingerprints. One of the latest approaches for prediction of molecular fingerprints from electron ionization mass spectra is DeepEI, which suggested a suboptimal procedure based on using a separate neural network for each molecular fingerprint (more than 100 models in our work and 636 using the DeepEI method). More than that, after repeating the procedure described in the original article, we assumed that at least some of their models were most likely overfitted. We streamlined the original approach by predicting multiple types of molecular finger- prints with a single multi-output neural network. We developed a lightweight and performant architecture (called Lite model for brevity) with improved accuracy (0.91 vs 0.89), precision (0.86 vs 0.77), and recall (0.71 vs 0.70) compared to the DeepEI approach. Additionally, the Lite version of the model was more than 100 times faster than the DeepEI approach in training and inference.