LOOPY: Guiding Generative Design with Locally Calibrated Predictor Uncertainty

Biological sequence design increasingly relies on property predictors to direct generation toward desired functions. Optimization shifts candidates beyond the data used to train and validate these predictors, risking systematic overprediction. Here, we introduce a method for measuring Local Oracle Optimism and Predictor uncertaintY (LOOPY) during biological sequence generation. LOOPY combines local calibration error, reference-data support, and uncertainty-aware candidate reweighting. Across measured plant enhancer libraries, we observe predictor optimism in four of five assay conditions. In generation experiments, score guidance reduces labeled-data support across all five DNA enhancer readouts and increases peptide support relative to both labeled and generator-training sequences. Penalizing local error preserves DNA support at a cost in predicted activity and has little effect on peptides because local errors vary minimally across candidates. In total, by detecting predictor overoptimization and tracking support throughout generation, LOOPY identifies where experimental validation is most necessary.