Deterministic layer (2a)#
- Lipinski's rule of five as the basic drug-likeness filter.
- A heuristic entry indicator based on the eNTRy rules, the predictive compound accumulation rules proposed for broad-spectrum antibiotics.
Learned accumulation and efflux model (2b)#
The accumulation model is trained the same way as in the source publication: on LC/MS measurements in an E. coli strain lacking the TolC pump, using 37 descriptors and gradient boosting. It reaches ROC AUC 0.845 on a random split and 0.834 on a scaffold split.
Why this is an annotation, not a cut-off#
Stage 2b does not act as a hard gate. The full set of accumulation predictions is passed to the third stage, which preserves the quality signal for compounds that a 0.95 threshold would remove and lets us assess that threshold retrospectively. Enforcing 0.95 would pass 1.1% of the library; the model's operating point of 0.833 would pass 9.2%.
Calibration against experimental data does not justify any threshold. The predicted retention probability separates neither enzymatic activity (ROC AUC 0.44, the opposite direction) nor whole-cell activity (0.59). Gating here only reduces the number of compounds without improving agreement between the ranking and experiment. The stage therefore works as an annotation and as input to the kinetic model.
The permeability and efflux mappings have never been calibrated against real intracellular accumulation data, because such data does not exist at this scale. This is the conceptual differentiator of the system and it remains an unconfirmed hypothesis.