What NOBBLE measures

NOBBLE generates and scores peptide poses against curated targeted-protein-degradation machinery using Boltz-2 ternary co-folding, and asks one question of every pair: does the presence of the glue or degrader change the predicted interface, and does that change separate experimentally degraded substrates from experimentally non-degraded ones assayed in the same screen?

delta_iptm = iptm_min_ternary - iptm_min_binary

Measured on the receptor-to-peptide chain pair, not the global ipTM: on a ternary complex the global figure averages in the receptor-to-ligand interface, which is the easy one, so a glue sitting perfectly in its pocket while the degron fails to dock would raise it.

The confidence caveat, stated plainly

The Thera-SAbDab benchmark of Boltz-2 across 519 therapeutic antibodies paired with cognate and non-cognate antigens found that its confidence metrics gave only weak, though above-chance, discrimination: ROC-AUC between 0.5 and 0.60, with ipTM-min carrying the strongest signal and multivariate modelling adding essentially nothing. Peptide-protein should fare better than antibody-antigen, but this application does not present ipTM as a verdict. It is reported as enrichment, alongside the geometry.

That is not a theoretical caution. On the benchmark panel's first pair, a pose that genuinely bridged receptor to peptide scored ipTM 0.885 and one that did not scored 0.877; the highest ipTM in the whole ternary set, 0.916, did not bridge at all. The bridge percentile separated them; ipTM did not.

Scoring signals and how far each is trusted

Lysine reach: why there is no score

The specification asks for BINMAN's reach-window fit, described as already calibrated against 976 observed diGly sites. It is not calibrated. BINMAN attempted the fit, measured it and declined to ship it, and these are the numbers behind that decision:

All at or below chance, and the Cb-Cb direction is slightly inverted: ubiquitylated lysines sit farther from the ligand (18.73 A median) than unmodified ones (14.89 A). The explanation is mechanistic rather than technical, which is why refitting is unlikely to help: the ligand site is where a substrate is recruited, not where ubiquitin is transferred, and the E2 that transfers it sits elsewhere on the cullin scaffold and is not in the trimmed construct at all. NOBBLE therefore reports lysine geometry as a measurement and emits no reach score.

Receptor constructs and the token budget

Every construct is defined by a named InterPro entry, never by a residue range, and its boundaries are resolved at build time. Two are declared exceptions to the 400-token tier-1 cap, both because the receptor does not fold or function without its partners.

Data sources

NOBBLE re-acquires nothing. Every source below is read from what BINMAN already curated and pinned. Licence terms carry over unchanged: PROTAC-DB is internal-use only and never redistributed, UbiBrowser ships as derived counts only, and BioLiP2 annotations are not redistributed. The published repository bundles no third-party dataset.

Reference bridge distribution: 285,996 experimentally observed bridges, which is what every bridge percentile on these pages is measured against.

The composite, and why it may be empty

Not fitted. not_fitted: 6 positive(s) and 0 negative(s) scored, below the floor of 8 labelled pairs. Spec 4.8 fits on the benchmark panel and the Sievers matched set; the Sievers arm runs after the panel. A composite fitted on fewer pairs than it has components would be fitting noise

Spec 4.8 fits the composite weights on the benchmark panel and the Sievers matched set and never hand-sets them, so until that fit has data every pair's nobble_score is null and the components are reported individually. A ranked column looks authoritative whatever its provenance, which is the reason for refusing rather than defaulting. The sliders on the Campaign tab explore a weighting; they do not publish one.

Validation

Three layers, all gating. A campaign that fails any of them publishes its failure rather than its hits. A poor ROC is not a gate failure; reporting no ROC, or reporting one without the matched negatives, is.

1. Pose recovery

Superposed on the receptor, peptide backbone atoms only. Every record carries its own anchor size and anchor RMSD, because a pose RMSD means nothing without evidence that the frame it was measured in was sound.
2 benchmark pair(s) have no recovery number yet: their ternary jobs have not completed

2. Time split

not computed not_computed: no benchmark deposition was released after the 2021-09-30 training cutoff, so this campaign has no blind set. The retrospective number stands alone and must not be read as prospective performance

The seven panel members are all canonical, long-published ternaries, which is exactly why they are good positive controls and exactly why they cannot estimate prospective performance. A blind number needs post-cutoff depositions from mgtbind_ternary that the campaign does not currently run.

3. Matched-control ROC

not computed not_computed: 5 positive(s) and 0 matched negative(s) scored. The Sievers arm runs after the benchmark panel; a curve drawn from benchmark pairs alone would have no matched negatives and spec 5 forbids reporting one

Measured throughput

The campaign sizes itself from this measurement, not from the specification's estimate. Note that the figure is dominated by whichever jobs have run so far: pairformer cost is superlinear in sequence length, so a rate measured on small constructs is optimistic for large ones.

Provenance

The repository records every judgement made during the build, with the alternatives, the reason and how to reverse it, including the places where the build specification turned out to be wrong about its own data: DECISIONS.md for decisions, FINDINGS.md for results, BUILD_LOG.md for the event log and BUILD_SPEC.md for the specification itself.

github.com/bellcheddar/NOBBLE · built on BoltzMaker and BINMAN