Enabling Data-Efficient Machine Learning through Multi-Target Training Set Design via Greedy Neighborhood Coverage

Abstract Machine learning is increasingly used to guide asymmetric catalysis, but each enantioselectivity measurement remains experimentally expensive. Prior work established that predictive accuracy in these datasets is dominated by reactions that are structurally similar to the target, motivating neighborhood-restricted models, such as radius-based random forests (RaRF). However, realistic screening campaigns require predictions for many related targets simultaneously (e.g., 20–50 reactions in a scope), and selecting training data independently for each target can be redundant because target neighborhoods overlap. Here, we frame multi-target training set design as a neighborhood multi-coverage problem, where the goal is to select a compact shared set of experiments such that each target retains at least one neighbor within a given similarity radius in reaction fingerprint space. We introduce MT-RaRF, a greedy neighborhood-coverage algorithm that prioritizes candidate experiments that satisfy the remaining local-neighbor needs of underserved targets while exploiting overlap among neighborhoods. Across several previously curated asymmetric catalysis datasets, MT-RaRF can use only a fraction of the experiments required by independent RaRF while maintaining comparable mean absolute errors. Systematic comparisons across similarity radii, training budgets, and target-set distributions establish practical guidelines for prospective use: MT-RaRF is most effective when locality is enforced, experimental budgets are limited, and target panels contain overlapping neighborhoods. MT-RaRF therefore makes locality-driven models practical and affordable for multi-target screening in asymmetric catalysis.