Closing the Loop: Evidence-Gated Pseudo-Labels for Active Learning in Molecular Property Optimization

Active learning (AL) can reduce experimental burden in molecular discovery, but its iterative cycle stalls when queried compounds cannot be immediately labeled by an experimental oracle. We present a weakly supervised framework that uses selectively admitted pseudo-labels as provisional feedback between sparse experimental rounds. The framework integrates lightweight multi-scale conditional modulation of whole-molecule and substructure representations, confidence--uncertainty paired acquisition through AutoMolDesigner, and Dirichlet-evidence-guided triple gating. At each iteration, a high-disagreement candidate is selected for exploration and a high-confidence candidate for exploitation. Their pseudo-labels are incorporated only when the predictions are stable, structurally supported by a frozen experimental reference set, and sufficiently decisive after evidence fusion; otherwise, the model abstains from self-training on these samples. This design reconciles the opposing objectives of uncertainty-driven acquisition and confidence-based self-training while limiting recursive pseudo-label confirmation bias.