MolGAN is an implicit generative model for small molecular graphs that avoids the need for expensive graph matching or node ordering heuristics. It uses generative adversarial networks (GANs) to generate molecular graphs directly, and combines this with a reinforcement learning objective to encourage the generation of molecules with desired chemical properties. The model is trained on the QM9 chemical database and is capable of generating nearly 100% valid compounds. MolGAN outperforms recent string-based (SMILES) and likelihood-based methods in terms of validity and novelty, although it is susceptible to mode collapse. The model uses a permutation-invariant discriminator and reward network based on graph convolution layers to operate directly on graph-structured data. It is trained using improved WGAN and a reinforcement learning objective, with the generator learning to match the empirical distribution and the discriminator learning to distinguish between generated and real molecules. The model is evaluated on various metrics including validity, novelty, and uniqueness, and is found to generate highly valid molecules with high diversity. MolGAN is faster to train than other models and is more efficient in generating chemically valid compounds. The model is susceptible to mode collapse, which can be mitigated through early stopping. The results show that MolGAN achieves higher validity and novelty scores compared to other models, and is a promising approach for generating small molecular graphs.MolGAN is an implicit generative model for small molecular graphs that avoids the need for expensive graph matching or node ordering heuristics. It uses generative adversarial networks (GANs) to generate molecular graphs directly, and combines this with a reinforcement learning objective to encourage the generation of molecules with desired chemical properties. The model is trained on the QM9 chemical database and is capable of generating nearly 100% valid compounds. MolGAN outperforms recent string-based (SMILES) and likelihood-based methods in terms of validity and novelty, although it is susceptible to mode collapse. The model uses a permutation-invariant discriminator and reward network based on graph convolution layers to operate directly on graph-structured data. It is trained using improved WGAN and a reinforcement learning objective, with the generator learning to match the empirical distribution and the discriminator learning to distinguish between generated and real molecules. The model is evaluated on various metrics including validity, novelty, and uniqueness, and is found to generate highly valid molecules with high diversity. MolGAN is faster to train than other models and is more efficient in generating chemically valid compounds. The model is susceptible to mode collapse, which can be mitigated through early stopping. The results show that MolGAN achieves higher validity and novelty scores compared to other models, and is a promising approach for generating small molecular graphs.