PETase ANnotation and Triage System

Sorted by identity to the nearest enzyme whose PET activity was actually measured — not by the profile bitscore, and not by a learned classifier. A trained head was tried and abandoned on evidence: across enzyme families it performs at chance and its coefficients point in opposite directions between families, and within a family it cannot be shown to beat this. Retrieval needs no training and cannot learn a family-specific rule that reverses on a family it has not seen.

The confidence column is the honest part. The signal was measured as a function of sequence identity: weak at 70%, unmeasurable below 50%. So 16 candidates are in range, 39 marginal, and 384 are ranked but carry no demonstrated validity at their distance from anything measured.

And the tension worth naming rather than hiding: ranking by similarity rewards candidates that resemble what is already known, which is the opposite of discovery and precisely the bias this project set out to correct. The low-identity candidates are the interesting ones and also the ones nothing here can score. That is not a bug in the ranking — it is the state of the evidence, and the fix is measurement on new lineages rather than a cleverer model.