Adversarial Learned Molecular Graph Inference and Generation

Recent methods for generating novel molecules use graph representations of and employ various forms convolutional neural networks inference. However, training requires solving an expensive isomorphism problem, which previous approaches do not address or solve only approximately. In this work, we propose ALMGIG, a likelihood-free adversarial learning framework inference de novo molecule generation that avoids explicitly computing reconstruction loss. Our approach extends generative by including cycle-consistency loss to implicitly enforce the property. To capture properties unique molecules, such as valence, extend Graph Isomorphism Network multi-graphs. quantify performance models, compute distance between distributions physicochemical with 1-Wasserstein distance. We demonstrate ALMGIG more accurately learns distribution over space than all baselines. Moreover, it can be utilized drug discovery efficiently searching using molecules' continuous latent representation. code is available at https://github.com/ai-med/almgig