Fig. 2: Spectral prediction performance for various GNN architectures.
From: FIORA: Local neighborhood-based prediction of compound mass spectra from single fragmentation events
The median cosine similarity is evaluated on the validation split, with error bars denoting the 95% confidence intervals. Note that the interquartile ranges of the data are considerably broader, as shown in Supplementary Fig. 24. Cosine similarity reaches its peak between a network depth of 4 and 6 layers before it falls off again. The graph convolution (GCN) and relational graph convolutional networks (RGCN) perform better than the attention based mechanisms (GAT and Transformer).
