External positives
59retained
ConoServer
Conopeptide candidates with source annotations; 59 leakage-audited records form the formal Stage 2 cohort.
Revision-period evaluation
Post-hoc robustness evaluation across four external sources
These cohorts were assembled and evaluated during manuscript revision; this is not prospective blind validation. Stage 1 discrimination and positive-only retained-label ranking are reported separately, with observed specificity and data gaps intact.
Status
Completed with limitations
Post-hoc | frozen 02 Aug 2026
- 629
- formal Stage 1 rows
- 87
- external positives
- 542
- operational comparators
- 0.8972
- primary AUROC
- 0.7428
- primary AUPRC
Source evidence
Four services, four bounded roles
Source graphics identify the original databases and do not imply endorsement.
ArachnoServer
OOD challenge positives
28retained
ArachnoServer
High-confidence spider-toxin candidates of at most 50 aa; 28 records remain a separate challenge cohort.
Assay-scoped comparators
521records
Hemolytik 2.0
Bioactive peptides retained under non-hemolytic assay evidence, used as operational rather than universal negatives.
Reference comparators
21records
IUPHAR/BPS GtoPdb
Twenty therapeutic-peptide records plus one disulfide-rich proxy retained with explicit evidence scope.
Cohort construction
Assembled during revision, frozen before inference
The positive audit began with 114 candidates: 76 ConoServer Tier A and 38 ArachnoServer high-confidence records. Accession, exact-sequence, de-X sequence, MMseqs90 cluster, and all-vs-all pairwise checks retained 87 independent records and excluded 27 records with internal-overlap evidence.
The formal Stage 1 cohort combines those 87 positives with 542 evidence-bounded comparator records. The separate 87 + 87 length-matched subset is a sensitivity analysis, not the headline cohort.
Formal Stage 1
629 rows
- ConoServer formal
- 59
- ArachnoServer OOD
- 28
- Bioactive/non-hemolytic
- 521
- Therapeutic peptides
- 20
- Disulfide-rich proxy
- 1
Stage 1 / discrimination
Primary result and robustness check
Metrics apply to the constructed comparator cohort, not universal toxicity prevalence.
| Frozen model | Role | n | AUROC (95% CI) | AUPRC (95% CI) | Sensitivity | Specificity |
|---|---|---|---|---|---|---|
| Primary Full ensemble | Primary | 629 | 0.89720.8549-0.9338 | 0.74280.6557-0.8189 | 0.9540 | 0.4686 |
| Robustness Full ensemble | Robustness | 629 | 0.87210.8243-0.9115 | 0.72560.6444-0.8021 | 0.9195 | 0.5498 |
Primary intervals use test-sample percentile bootstrap (B = 10,000; seed 20260805). The robustness-ensemble audit uses its frozen stratified bootstrap (B = 2,000; seed 20261801).
- Length-matched audit
- 174 rows (87 + 87)
- Primary matched AUROC
- 0.8869
- Primary matched AUPRC
- 0.9055
- Close non-toxic homologues
- n = 0 / FPR not applicable
Stage 2 / retained-label ranking
Source-separated positive-only evaluation
Primary Full ensemble; query-only A3M fallback, not database homolog-search MSA.
Revision cohort
ConoServer n = 59
- Hit@1
- 0.6271
- Hit@3
- 0.8136
- MRR
- 0.7432
- micro-AP*
- 0.6357
OOD / challenge
ArachnoServer n = 28
- Hit@1
- 0.4643
- Hit@3
- 0.7143
- MRR
- 0.6168
- micro-AP*
- 0.4460
* Micro-average precision is a retained-vs-unannotated diagnostic. Unretained labels are not verified biological negatives.
Open evidence
Aggregate report and integrity ledger
Public files contain aggregate metrics and frozen hashes. Row-level source-derived sequences are not mirrored; obtain them from the original services under their current citation and reuse terms.