GraphEBM: Towards Permutation Invariant and Multi-Objective Molecular Graph Generation

Although significant progress has been made in molecular graph generation recently, permutation invariance and multi-objective generation remain to be important but challenging goals to achieve. In this work, we propose GraphEBM, a molecular graph generation method via energy-based models (EBMs), as an exploratory work to perform permutation invariant and multi-objective molecule generation. Particularly, thanks to the flexibility of EBMs and our parameterized permutation-invariant energy function, our GraphEBM can define a permutation invariant distribution over molecular graphs. We learn the energy function by contrastive divergence and generate samples by Langevin dynamics. In addition, to generate molecules with a specific desirable property, we propose a simple yet effective learning strategy, which pushes down energies with flexible degrees according to the properties of corresponding molecules. Further, we explore to use our GraphEBM for generating molecules towards multiple objectives via compositional generation, which is practically desired in drug discovery. We conduct comprehensive experiments on random, single-objective, and multi-objective molecule generation tasks. The results demonstrate our method is effective.