Figure 2.

The Ensemble Spectra Prediction (ESP) model has two phases. Phase 1. Molecular encoding using GNN and MLP followed by spectra prediction enhancement using label-mixing to capture spectral dependencies and using multi-tasking on predicting LDA topic distribution. Phase 2. Training ensemble scoring model, we compare the rank of spectra prediction from GNN and MLP encoding on candidate ranking problem for query spectra in the training set to predict a score on weighing the two models. Note that ESP focus on solving the three subproblems of metabolite annotation: GNN and MLP models in Phase 1 focus on (1) molecular encoding; label-mixing and multi-tasking on LDA topic distribution prediction in Phase 1 focus on (2) spectra encoding; ensemble scoring based on ranks of target molecule from two models in Phase 2 focus on (3) candidate ranking.