부산대학교 도서관

We developed Distilled Graph Attention Policy Network (DGAPN), areinforcement learning model to generate novel graph-structured chemicalrepresentations that optimize user-defined objectives by efficiently navigatinga physically constrained domain. The framework is examined on the task ofgenerating molecules that are designed to bind, noncovalently, to functionalsites of SARS-CoV-2 proteins. We present a spatial Graph Attention (sGAT)mechanism that leverages self-attention over both node and edge attributes aswell as encoding the spatial structure -- this capability is of considerableinterest in synthetic biology and drug discovery. An attentional policy networkis introduced to learn the decision rules for a dynamic, fragment-basedchemical environment, and state-of-the-art policy gradient techniques areemployed to train the network with stability. Exploration is driven by thestochasticity of the action space design and the innovation reward bonuseslearned and proposed by random network distillation. In experiments, ourframework achieved outstanding results compared to state-of-the-art algorithms,while reducing the complexity of paths to chemical synthesis.