Knowledge Guided Geometric Editing for Unsupervised Drug Design
Deep learning models have been widely used in automatic drug design. Current deep approaches always represent and generate candidate molecules as a 1D string or a 2D graph, which rely on large measurement data from lab experiments for training. However, many disease targets in particular newly discovered ones do not have such data available. In this paper, we propose \method, which incorporates physicochemical knowledge into deep models, leading to unsupervised drug design. Specifically, \method directly models drug molecules in the geometric~(3D) space and performs geometric editing with the knowledge guidance by self-training and simulated annealing in a purely training data free fashion. Our experimental results demonstrate that GEKO outperforms baselines on all 12 targets with and without prior drug-target measurement data.
PDF Abstract