Uncertainty- and Coverage-Aware SpaceHASTEN for Mining Unenumerated Chemical Spaces

Nonenumerated make-on-demand chemical spaces contain billions of accessible molecules, but the optimal way to navigate such a space in a structure-based virtual screening campaign is an open question. Greedy acquisition naturally concentrates on high-scoring local chemistry, favoring immediate yield over broader exploration. We introduce Uncertainty- and Coverage-Aware SpaceHASTEN (UC-SpaceHASTEN), in which a guarded controller coordinates where to search the library and which newly discovered molecules to dock. Calibrated predictions favor promising candidates, while coverage rewards discourage repeatedly sampling the same chemistry. UC-SpaceHASTEN substantially broadened explored chemistry and improved several diversity metrics at the cost of finding fewer docking-score hits. These findings reveal a controllable yield-breadth trade-off and show that explicit coverage objectives can produce more diverse virtual-hit portfolios from nonenumerated chemical spaces.