# Selectivity benchmark

Predicted wild-type margin versus **measured** selectivity, pChEMBL(mutant) − pChEMBL(wild type), for compounds assayed on both. The control that matters is whether a one-bit "has a warhead" rule does just as well.

| Allele | n | Selective (≥3×) | Margin AUC | Margin ρ | 'Has a warhead' AUC | Verdict |
| --- | --- | --- | --- | --- | --- | --- |
| **L858R** | 247 | 90 | **0.497** | +0.100 | 0.486 | no predictive value |
| **L858R,T790M** | 489 | 264 | **0.613** | +0.227 | 0.593 | below the 'molecular weight' control |
| **L858R,T790M,C797S** | 70 | 39 | **0.444** | -0.033 | 0.429 | no predictive value |
| **T790M** | 97 | 22 | **0.603** | +0.186 | 0.521 | beats every control |

- **L858R**: AUC 0.497 against a 90/247 selective base rate is indistinguishable from chance: the margin does not predict measured mutant selectivity. The margin takes only 2 distinct values here and 51% of compounds share one of them, so it is a categorical verdict rather than a ranking; an AUC over it is mostly measuring ties.
- **L858R,T790M**: AUC 0.613 against 0.645. The margin takes only 4 distinct values here and 69% of compounds share one of them, so it is a categorical verdict rather than a ranking; an AUC over it is mostly measuring ties.
- **L858R,T790M,C797S**: AUC 0.444 against a 39/70 selective base rate is indistinguishable from chance: the margin does not predict measured mutant selectivity. The margin takes only 3 distinct values here and 74% of compounds share one of them, so it is a categorical verdict rather than a ranking; an AUC over it is mostly measuring ties.
- **T790M**: AUC 0.603, above the best control 'molecular weight' at 0.522. The margin carries selectivity information those do not. The margin takes only 4 distinct values here and 41% of compounds share one of them, so it is a categorical verdict rather than a ranking; an AUC over it is mostly measuring ties.