Fig. 1

Sketch of the Siamese neural network architectures and training strategy behind MS2DeepScore. The Siamese network uses the same “base network” twice during training and prediction to convert a binned spectrum into a spectral embedding (200-dimensional vector). The network is trained on spectral pairs and the mean squared error between the cosine similarity of the two spectral embeddings and the actual structural similarity score (here: Tanimoto between RDKit 2048bit fingerprints). To increase the robustness of the model data augmentation techniques were used which includes moderate random changes of peak intensities as well as removal and addition of low intensity peaks