Edit-Aware Generative Molecular Graph Autocompletion for Scaffold Input

Sheng Hu, Ichigaku Takigawa, Chuan Xiao

Abstract: We present a novel molecular graph generation method by auto-completing a privileged scaffold which represents a core graph substructure step-by-step. We propose a generative GNN model thus providing the ability to generate unseen molecular graphs outside the given training set. An edit-aware graph autocomplete- tion paradigm that follows the “substructure-by-substructure” process is designed to complete the scaffold queries in multiple substructure adopt operations and allow meaningful edit operation to show the user’s intention. Such operations enable the involvement of user decisions when interacting with a generative user-centered AI system, which differentiates our work from existing single-run generation paradigms. We also propose a scaf- fold trie for fast training pair augmentation or changing training models in real-time. Moreover, we design a top-k ranking func- tion which considers the preferences on popularity and diversity for different applications, such as query compositions for graph database and drug discovery respectively. Such techniques en- able human experts to synergistically interact with the generative models grounded on large data.