All the reviewers agree that this work is empirically solid (beating quite strong baselines), the area of application (a representation for molecules) is important/significant, and the developed method is novel, interesting, and theoretically analyzed. It would be a welcome addition to the NeurIPS program.