Methods & research
What we're trying, why, and what the evidence says — each lever tested on the WDR91 ground-truth yardstick before it touches PGK2.
The winning recipe (Wellnitz 2025)
Our exact task — train on a DEL, screen a disjoint Enamine library — is solved and published: LightGBM on FCFP6, an NTC-corrected, disynthon-aggregated, unfiltered enrichment label, a decorrelated structure re-rank on the full top slice, and diverse-series selection. 7/48 (~15%) confirmed. Our #1 bug was a filtered label that traded away the recall that catches real binders.
The structure arm (Boltz-2)
An open AF3-class protein+ligand co-folder with an affinity head, deployed on the NEU Explorer GPU cluster. On WDR91 ground truth it separates the 19 confirmed binders from DEL-enriched non-binders at AUC 0.667 — a domain-gap-immune signal, because it folds the actual molecule into the pocket rather than relying on library resemblance.
The conversion, validated
AUC ≠ top-50. A model at AUC 0.80 still recovers ~0 hits at a brutal cut. The fix — proven on WDR91 — is the ensemble: rank-averaging the trained model with the Boltz re-rank recovers 11/19 (vs 10 model-alone, 7 Boltz-alone), because they catch different hits. Diversity selection is nearly free (18/19 hits have distinct scaffolds).
The data deep-dive
A per-compound case-report engine (see the compound studies) revealed our picks were out-of-distribution from the library — the mechanistic cause of the zeros. Adjacent data: PGK2 is externally data-empty, but its paralog PGK1 has 50 potent actives usable as a structure-calibration + similarity prior. Orthogonal: JUMP Cell Painting as a domain-gap-immune morphology surrogate.
Cross-domain research
25 permanent 'trick another field's method into DEL use' angles — positive-unlabeled learning, quant-finance rare-event detection, single-cell count denoising, astronomy photon-counting, ecology presence-only models, extreme-value theory, topological data analysis, and more — each tested against the WDR91 yardstick.
The autonomous loop
A Meta-Queen (AI-scientist strategist) picks a research frontier each day, forms a falsifiable hypothesis, writes a directive, and delegates to worker agents; an Innovation-Queen mines hackathons for transferable insight (87 catalogued). Both run indefinitely on cloud crons and commit to the repo.