Confounded Oracles Misguide Biomolecular Design: Population-Structure Shortcuts in Antimicrobial Candidate Optimization

Predictive oracles are important tools for in-silico biomolecular design. The goal of optimizing an oracle is to make sure that it learns correct associations relevant to the biological target of interest. We examine how an underlying population-specific shortcut in AMR prediction is translated to optimization of antimicrobial-peptide latent space using a controlled semi-synthetic stress test parameterized by the resistance prevalence of 3,975 Mycobacterium tuberculosis records where the entire system is simulated, including genomic features, peptide latent factors, and drug-target efficacy. While CORAL minimizes the source-target covariance mismatch by $>99.6\%$, without loss of ability to perfectly decode the simulated population-associated shortcut, DANN erodes transferable information while both shortcut and domain identity remain recoverable. Under matched prior-regularized optimization, CORAL results in improved efficacy compared to a source-only oracle when source and target share an efficacy mechanism. But worse efficacy with changing target (difference-of-differences permutation test, $p = 0.0039$). We discuss these experiments as a controlled oracle-reliability stress test rather than realistic peptide-engineering validation.