Evaluating Conditional Resistance in Synthetic Microbiology Profiles

Matching single-drug resistance rates does not ensure preservation of conditional resistance relationships in synthetic microbiology profiles. We adapt conditional susceptibility comparisons into a protocol with fixed organism–drug targets, signed-gap errors against held-out data, explicit missing estimates and conditioning-group counts. In a controlled MIMIC-IV shuffle, mortality AUROC and the multidrug-resistance proxy remain nearly unchanged while gap error rises 5.3-fold. Transformer-generated profiles show attenuated mean absolute gaps: 26.35 versus 29.78 percentage points in held-out MIMIC-IV, and 29.44 versus 35.87 in MIMIC-III. These are descriptive differences under the evaluated representations. All 377 MIMIC-IV targets are calculable, but only 233 have at least 20 profiles in every generated and held-out conditioning group across three sampling seeds. The protocol makes relationship attenuation and sparse estimates visible alongside aggregate evaluation.