A Phase-0 experiment asking whether ATP synthase's ring architecture falls out of ancient protein fragments and constraint satisfaction alone — and the first result in the project that wasn't discovered after the fact.
GSYN is a go/no-go experiment, not a finished result: can the F₁ ring architecture of ATP synthase be derived, rather than assumed, from a small alphabet of ancient protein fragments and constraint satisfaction over graphs — with the solver never told what it's looking for? Three roles stay strictly separated: Foldseek and a CP-SAT solver do the actual deterministic work, a local model only starts, watches, and kills jobs and summarizes logs, and interpreting the results is mine, working only from the exported exchange table, never the raw pipeline internals.
The rule that matters most: the solver is never given the target. No target structure, no target-derived constraint, no early stop on a match. If a six-membered ring shows up, it has to fall out of the graph on its own.
Most of what's run so far confirms things after the fact — the pipeline runs, and it happens to match known biology (1BMF and 2GXA both did this). The most recent run, on 1E0J, is different in kind: the closed 6-cycle prediction was pre-registered — written down, committed — before the fragment pipeline ever saw the structure. The pipeline then reproduced that exact edge set, six for six, with zero deviation and no chord. That's the first genuinely blind prediction in the project, not a retrospective one.
Across all four ring targets tested so far (1BMF, 1E79, 2GXA, 1E0J): pooled p = 24/25 = 0.960 (95% CI [0.796, 0.999]) that a real physical contact between fragments shows up as a level-2 graph edge. One located miss stands, unchanged by this run: 1E79's C–F interface, 575 atom-atom contacts, zero fragment-level detection — a real limitation, not swept under anything.
Still Phase 0. Still a yes/no question about whether this representation can carry the signal at all, not yet a claim about the biology.