Fixed-Budget Dose Layouts in Merck 4 × 4 Combination Matrices
We compared target-specific eight-point dose layouts for four withheld interior responses in 171 held-out Merck 4 × 4 drug-pair matrices, with (1,1) designated as the primary target. For the projection predictor, near-minus-far mean absolute error was −0.05772 at the primary target (95% pair-bootstrap interval −0.07163 to −0.04441) and −0.07085 for the four-target pair average (−0.07899 to −0.06262). Overall projection errors for near, far-plus-anchor and fixed random layouts were 0.0412, 0.1120 and 0.0773; on the same near inputs, the order-based predictors had lower average errors than the tested local interpolators, while projection showed no clear added benefit over the raw witness. A post-hoc Gaussian-process comparator showed no clear average difference from projection on the Merck near inputs, with optimizer-bound warnings in 293 of 684 fits.
Authors
- Yan PAN (ORCID: https://orcid.org/0009-0002-8835-3603)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.5281/zenodo.24314440
- Primary Topic
- Computational Drug Discovery Methods
- Type
- preprint