PURE: policy-guided unbiased REpresentations for structure-constrained molecular generation

Structure-constrained molecular generation (SCMG) generates novel molecules that are structurally similar to a given molecule and have optimized properties. Deep learning solutions for SCMG are limited in that they are predisposed towards existing knowledge, and they suffer from a natural impedance mismatch problem due to the discrete nature of molecules, while deep learning methods for SCMG often operate in continuous space. Moreover, many task-specific evaluation metrics used during training often bias the model towards a particular metric -"metric-leakage". To overcome these shortcomings, we propose Policy-guided Unbiased REpresentations (PURE) for SCMG that learns within a framework simulating molecular transformations for drug synthesis. PURE combines self-supervised learning with a policy-based reinforcement learning (RL) framework, thereby avoiding the need for external molecular metrics while learning high-quality representations that incorporate an inherent notion of similarity specific to the given task. Along with a semi-supervised training design, PURE utilizes template-based molecular simulations to better explore and navigate the discrete molecular search space. Despite the lack of metric biases, PURE achieves competitive or superior performance to state-of-the-art methods on multiple benchmarks. Our study emphasizes the importance of reevaluating current approaches for SCMG and developing strategies that naturally align with the problem. Finally, we illustrate how our methodology can be applied to combat drug resistance by identifying sorafenib-like compounds as a case study.