Meg-Apsu
Some enzymes move a hydrogen by letting it tunnel through the energy barrier rather than over it. Whether an enzyme does this cannot be read off its name; it depends on the geometry of its active site. MEG-APSU takes an enzyme's structure, runs it through an open-quantum-system model, and places it on one side of a line: quantum-critical, or classical. It is the engine that TERRA uses to scan species, and its honest worth is exactly as good as that line is drawn.
From the active-site geometry, MEG-APSU builds a small open-quantum-system (Lindblad) model of the hydrogen-transfer step and reads out how much of the rate the tunnel carries. A threshold on that quantity decides the label. It is written in Rust, from structure to classification.
Honest Accuracy
A classifier is only as trustworthy as the test it has not seen. MEG-APSU is measured on a frozen set of 40 proteins, split so that no protein family appears in both training and test. On that set, version 1.3 is right 92.5 % of the time. The simplest sensible rule — guess from the cofactor alone — gets 75 %. So the quantum model earns its keep, but it is not magic.
What Changed, and Why the Old 100 % Was Wrong
An earlier version reported 100 % accuracy. It was not real. The decision threshold had been tuned on the very proteins used to test it, and several heme proteins — myoglobin, haemoglobin, cytochrome c — were being called quantum-critical when they are not. Version 1.3 adds a filter for metal-catalytic sites and separates the development set from the test set properly. The honest number fell from 100 % to 92.5 %, and that fall is the point: a model that cannot be wrong on paper cannot be trusted in the world.
Applied to Drug Targets
Run across 115 FDA-target enzymes, MEG-APSU labels about 35.7 % of them quantum-critical. That figure should be read through the accuracy above: it is a classifier's estimate at 92.5 % per-enzyme reliability, not a verified count. It is a map of where to look, not a verdict on each target.
What It Does Not Claim
Three proteins are still misread on the test set — flavodoxin, haemerythrin, bacterioferritin — and the aconitase label is uncertain. The model's prediction of the kinetic isotope effect is weak (ρ = 0.19): it sorts enzymes into the right bin better than it predicts their exact numbers. And 40 proteins is a small test; the confidence interval says so plainly. The next version is meant to re-check the labels against direct tunnelling evidence and to grow the test set.
Preprint: bioRxiv 2026/713696 · Zenodo 10.5281/zenodo.19208904 · Code: github.com/sectio-aurea-q/meg-apsu