Bandit-Guided Optimization Accelerates Drug Discovery in Chemical Space

Identifying promising candidates from vast chemical libraries, which can contain billions or trillions of compounds, presents a significant challenge in fields like drug discovery due to the high cost of evaluating each candidate. While surrogate-based optimization can improve efficiency by reducing the number of expensive evaluations, the sheer size of modern molecular libraries makes even the surrogate inference process computationally intensive. To address this, researchers have developed BOBa (Bandit-guided Optimization for Big chemical spaces), a novel framework that bypasses the need for full-library surrogate inference. BOBa achieves this by adaptively distributing computational resources across different partitions of the chemical action space. It treats these partitions as "arms" in a multi-armed bandit problem, allowing the system to concentrate inference and evaluations on the most empirically promising sections while still ensuring principled exploration of the entire space. Experiments conducted on real-world synthesis-on-demand libraries demonstrated that combining optimism-under-uncertainty bandits with intelligent action space partitioning is crucial for effective allocation of both inference and evaluation resources. The findings highlight a flexible trade-off between screening performance and the computational cost of surrogate inference, offering a practical pathway for optimizing current large libraries and enabling virtual screening for ultra-large chemical spaces.