Construction#
Each quantity is first passed through a desirability function onto a common dimensionless scale, then blocks are summed with weights. The objective function is defined by one explicit schema and we report the effective weights.
Score v2#
In variant v2 the objective has two active blocks:
- 0.75 affinity desirability,
- 0.25 QED desirability.
The molecular dynamics block acts purely as a gate, with an effective weight of zero. Compared to earlier variants, score v2:
- removes the penetration block from the aggregate as uncalibrated, keeping it as an annotation;
- neutralises the secondary KD gate and resistance label, which was a deterministic threshold on the same affinity signal and zeroed 78.35% of records;
- omits the docking block (absent from the campaign data) and the systems biology block (indistinguishable in rank from affinity desirability).
The variant was selected by analysing 16 configurations over 1,003 records with measurements, and its stability was checked with 200 weight perturbations of ±20%. When a block is unavailable its weight disappears and the remaining weights are renormalised.
Calibration against experimental data#
| Method | Spearman rho | ROC AUC | PR-AUC |
|---|---|---|---|
| Score v2 | 0.705 | 0.875 | 0.931 |
| Affinity only | 0.697 | 0.858 | 0.924 |
On a shared cohort of 1,003 records with a pActivity measurement, score v2 outperforms affinity alone on all three measures. This is a retrospective validation of a variant chosen on that same cohort, not a result on an untouched test set.
On the whole-cell endpoint, growth inhibition of E. coli, the same change lowers ROC AUC to 0.483. No single objective function dominates on both available endpoints. Full comparison: DHFR campaign results.