The organism has a well-described metabolic network and calibration data available, while posing the permeability and efflux problem characteristic of Gram-negative bacteria.
Scale of the campaign#
The campaign narrows by roughly five orders of magnitude between generation and the compounds we actually inspected:
Library quality#
The analysed sub-library of 53,740 compounds contains 33,876 unique Bemis-Murcko scaffolds (85.4% singletons) with a mean internal diversity of 0.8693, so there is no concentration on a single chemotype. 91.8% of the de novo compounds sit below Tanimoto 0.4 to known DHFR actives. In ECFP4 space, only 3.34% of de novo compounds fall inside the applicability domain defined around measured active ligands.
Calibration of score v2#
Variants are compared on a shared cohort of 1,003 records with a pActivity measurement and 155 Bemis-Murcko scaffolds (860 records after excluding the grey zone):
- Score v2: Spearman rho = 0.705 [0.667; 0.741], ROC AUC 0.875 [0.849; 0.899], PR-AUC 0.931.
- Affinity alone: rho = 0.697, AUC 0.858, PR-AUC 0.924.
Adding QED at 0.75 / 0.25 lifts all three measures to the score v2 values; adding penetration back in drops AUC to 0.695. This is a retrospective validation of a variant selected on the same cohort, not a result on an untouched test set.
The whole-cell endpoint#
View as a table
| Method | Enzymatic | Whole-cell |
|---|---|---|
| Score v2 | 0.875 | 0.483 |
| Affinity only | 0.858 | 0.653 |
On growth-inhibition labels for E. coli from the Stokes screen, score v2 reaches ROC AUC 0.483 (n = 2,116 with 67 actives), against 0.653 for affinity alone and 0.554 for a random ranking. The drop comes from QED: here, high QED is significantly associated with a lack of activity. The conclusion is that no single objective function dominates on both available endpoints.
Benchmark of nine scoring methods#
All methods were evaluated out-of-fold in 10-fold GroupKFold cross-validation over 155 Bemis-Murcko scaffolds, on the same 1,003 records:
- Score v2 has the highest rho (0.705) and PR-AUC (0.931) in the set.
- ECFP4 with XGBoost has a higher point ROC AUC (0.893), but the paired AUC difference against v2 is not significant after Holm correction (p = 0.432).
- The advantage of v2 in rho over XGBoost is significant (p = 0.0048).
The benchmark therefore shows an advantage that depends on the criterion, not an unambiguous superiority.
Candidates for further validation#
The maximum PROTO-NOOS score in the de novo pool (0.6177) is clearly lower than the maximum in the ChEMBL set (0.9997); no de novo compound exceeds the 0.9 threshold. Nine of the 15 top-scoring de novo compounds contain a carboxylic acid motif (60%, against 5.72% in the library), consistent with the preference of the penetration axis. We propose a target portfolio in three layers: known DHFR chemotypes as positive controls, structurally distinct de novo compounds after removing PAINS and Brenk alerts, and a separate series of allosteric hypotheses.
The candidates are computational hypotheses that require synthesis plus biochemical and microbiological validation. Without a wet lab (MIC, IC50, cytotoxicity) the reported AUC values remain retrospective validation, not a preregistered holdout. The campaign covered a single target, with no selectivity panel.