Continuous research · Protein Data Bank
Three benches. Every week.
Graded by reality.
Each Wednesday the Protein Data Bank publishes new experimental structures: free ground truth. We grade three preregisteredEvery bench's predictions, thresholds, and controls are committed to version control before the run. Results that miss publish at the same prominence as results that hit. benches against it: which new structures will the archive itself later revise, where prediction diverges from experiment, and which binding pockets are genuinely novel. All results publish here, hits and misses alike.
Week of … Sample data
Every number below carries its provenance: temporal cutoff …All models are fit only on structures released before this date and frozen. Grading happens exclusively on structures released after it. The models have never seen what they're graded on., harness revision …. Snapshots are immutable; corrections ship as a new dated cut.
New entries graded
…
structures in this week's PDB release
Novel pockets flagged
…
target cell: novel structure, known sequence
Ligand-evidenced
…
tier 2+: a real ligand bound in the pocket
Archive mirrored
…
entries in the local mirror the benches run on
This week's detected pockets in the learned embedding's 2-D projection. The small cluster is the point: pockets whose geometry is novel while their sequence family is known: new chemistry on familiar proteins. Hover any point; every value is also in the table view.
Calibrated probability that each newly deposited structure will later be substantively revised by the archive itself. Most entries score low; the right tail is the watch list. Control to beat: ranking by resolution alone.
Highest-risk entries this week
| Entry | Method | Resolution | Score |
|---|
Where deposited experiments disagree with the corresponding predicted models, residue by residue, focused on interfacesResidues where two protein chains meet. Interfaces are where predictions earn or lose their keep, and where a confident-but-wrong model costs the most downstream.. Control to beat: the predictor's own confidence score.
Coming soon.
Ground truth arrives every Wednesday.
The Protein Data Bank publishes its weekly release on a fixed schedule. That makes it a rare thing in machine learning: a continuous supply of untouched, time-stamped ground truth. Our discipline is built around it: models fit on the past, frozen, committed, then graded on a future they have never seen. The same verified-completion posture as our agent work: the system being evaluated never gets to grade itself.
Pocket-novelty backtest
28.5%
of target-cell pockets are ligand-bound, against a 12.4% base rate. The preregistered bar was 24.8% (double the base rate); the descriptor baseline managed only 17.9%
Pre-cutoff pocket catalog
15,385
pockets detected and embedded from pre-cutoff structures: the frozen reference the novelty call is made against
Family ligand history
18,339
(family, ligand) pairs across 499 sequence families: what "novel for this family" is judged against
Want the longer version, with oracles, evidence tiers, and the temporal split? It's in the FAQ. Want this level of measurement discipline pointed at your own systems? That's the audit.