Fig. 6

Sketch of MS2DeepScore running in Monte-Carlo dropout modus. By keeping the dropout layers switched on, model predictions will essentially be sampled from random variations of the respective neural network. With a dropout rate of 0.2, a random selection of 20% of all nodes in the central dense layer(s), see Fig. 1, will be silenced. From n resulting variations of the created spectrum embeddings an array of n * n scores can be computed. Finally, the array of ensemble scores is used to calculate a single median score together with the interquartile range as a measure of model uncertainty