Extended Data Fig. 6: Additional details for the GNN update blocks.

(A) The bond update block concatenates the features of the source atom and the target atom together with bond’s starting features and passes them through two dense layers. The output of this network is summed with the input bond features in a residual fashion. (B) The atom update block separately concatenates the atom’s features with the features of each bond targeting the given atom and passes them through two dense layers. The output of these layers is summed over the incoming bonds and further passed through a series of dense layers to form an updated atom state. This updated state is summed with the previous atom feature vector in a residual fashion. (C) The global update block concatenates all bonds and atoms in the molecule and performs a multi-head dot product attention operation. This operation extracts a single feature vector for the entire molecule in a manner that is in variant to the ordering of atoms and bonds in the graph.