2nd DREAM×CACHE Target-2035 DEL-ML · PGK2 (P07205)
Finding real PGK2 binders in a 429K catalog, taught only by an 890-million-compound DNA-encoded library.
This is the full, unedited chronicle of the campaign — every method, reversal, and honest correction; the directives that steered it; the research behind it; and a per-compound case file for the molecules we've studied, in 2D and 3D. The central difficulty: the library we learn from and the compounds we must pick barely overlap (a domain classifier tells them apart at AUC 1.0). The story is how we're crossing that gap.
The arc
1 · The wall
Eight straight submissions — pure-ML, a reproduced organizer baseline, a validated structure re-ranker, an OOD-aware tailored re-rank — all scored 0 confirmed hits at the top-50 cut. Every one shared a flaw: it let the DEL be the teacher, and the DEL lives in a chemical space the answers don't.
2 · The map
An AI-scientist research loop found our exact task already solved and published (Wellnitz 2025, ~15% on the sister target WDR91) — so zero was a pipeline bug. And a per-compound case-report engine showed our picks were out-of-distribution from the library.
3 · The inversion
The turn that worked: stop letting the DEL teach. Ask instead what PGK2 is evolved to bind — mine its canonical binders and pocket-motif neighbours across every public source, build a DEL-independent prior, and demote the DEL to a 15% auxiliary. Pocket-gated so off-fold noise (which collapses it to chance) never enters.
4 · The first hit
The 9th submission — the first that isn't DEL-primary — scored 1 confirmed hit, 1 series on val (p=0.052). Modest and borderline, but after eight zeros it's the first crack in the wall, and the first proof the mechanism-first inversion surfaces a real binder.
5 · The method transfers
On the blind test set — completely novel scaffolds — the DEL-independent method ranks true hits above non-hits at ROC-AUC 0.649, where DEL-only ML is ≈0.5 (random). The inversion generalizes across the gap. The residual wall is unchanged: global ranking transfers, but landing a hit in the brutal top-50-of-184K cut (N_hits still 0) is the unsolved conversion problem — the one a real PGK2 dataset would crack.