Fig. 2
From: Drug discovery and mechanism prediction with explainable graph neural networks
Illustration of feature extraction procedure of atom 2 in the Butyramide molecule. In the initial stage, chemical features including number of non-hydrogen neibours, valency, atomic number, etc., are extracted and hashed to compute the initial identifier of each atom. In the update stage, starting from radius of 1, the bond orders and identifiers of the surrounding atoms (atom 3, 6 and 1) are combined and concatenated with the iteration number and the initial identifier of the focused atom (atom 2). The hash function is applied again on the concatenated feature to form the new identifier. Finally, in the reduction stage, each of the identifiers of atom 2 generated in the update stage are converted to a 64-bit binary vector and concatenated to form the final atom feature. In our implementation, we combined the binary vectors of radius 0, 1, 2 and 3, which forms a the final feature vector of length \(4 \times 64 = 256\).
