Fig. 3From: Force field-inspired molecular representation learning for property predictionVisualization of FFiNet. a Heat maps of the atom similarity matrix for the compound 2-(phenoxy)-1-phenylethanone. The atom similarity is obtained by calculating the Pearson correlation coefficient for the output state vectors of the final layer. The atoms in the molecule are automatically separated into two clusters after training. The part shown in the red square corresponds to the area highlighted in red in the molecule, and the part shown in the green square corresponds to the areas highlighted in green in the molecule. b The atomic contributions for the lipophilicity. The chlorine atom in the first molecule and the secondary amine atom in the second molecule get the most weight. The LogP of the molecule above and the molecule below are 2.68 and 0.38, respectively. c The attention weights of the atoms within 3-hop of the target atom. The target atom is denoted by a blue pentagram. Different attention heads can perceive the message of atoms at different hopsBack to article page