Validation

The endpoint

CYP2D6 turns tamoxifen into endoxifen, the metabolite that does the work. A variant that lowers CYP2D6 activity lowers endoxifen formation. The test is whether the model, given only the variant, predicts that shift on variants it never saw in training.

The firewall

Anything involving tamoxifen or endoxifen is kept out of the training data entirely. If it leaked in, the test would be circular and the number would mean nothing. The held-out set lives behind a hard separation from the corpus, checked before the model is frozen.

Pre-registration

The analysis was written down and frozen before any patient data: the model, the corpus, the primary comparison, the success criterion, all under a hash on the frozen protocol. The primary question is whether a continuous activity prediction beats the discrete per-allele score guidelines use today. Fixing this in advance is what stops the result from being tuned after the fact.

Honest state

It is not concluded. On the part of the range the model can represent, the signal is coherent and the confidence interval excludes zero. On the full set it is weaker, pulled down by gain-of-function alleles a multiplicative model cannot reach by construction. The outcome is genuinely open. That is why the protocol stays frozen instead of being reported early.

The corpus is also noise-limited. The protocol says so rather than glossing over it. Laboratories measuring the same CYP2D6 variants disagree by a factor of 2.4. Measured against that ceiling the model reaches about a quarter of what is reachable. The binding constraint right now is replicate measurements, not more substrates.

Where this sits

Predicting a single variant's activity shift is the near-term, bounded question. It is where the honest signal is. Predicting per-drug substrate specificity is the harder one, and nobody has solved it. The methodology explains why it measures null on the current features. The null is characterized, not hidden.