A baseline identity collision in the released ProteinTalks data loader
This technical note documents a source-specific baseline-identity error in the public ProteinTalks data loader at PTV-1 commit 999bf23ce4195d904ab3e00bc6b9ec733b8bae4c. Substring matching selects a baseline from the overlapping MDA-MB-453(ATCC) context for the separately labelled MDA-MB-453 context. In the audited public release, the minimal correction changes 94 of 1,530 complete trajectories and separates 79 cross-context pairs that otherwise have identical complete model inputs despite different targets. Seven separated pairs have conflicting observed binary screen labels. Four within-context collision groups remain unresolved. The archive contains audit code, a minimal patch, derived affected-condition and collision tables, figures, and aggregate verification receipts. Original research measurement arrays and complete upstream source are acquired separately from pinned public references. Native-loader replay, separately implemented numerical checks, and three synthetic real-loader regression tests verify the correction. A calculated empirical squared-error fitting floor is not a trained-model performance result. Whether the audited source affected the original paper's published benchmarks has not been established. The proposed correction is submitted in upstream pull request #10. OpenAI Codex agents assisted with source review, code, calculations and drafting; independent checks mean separately implemented computational checks.
Authors
- Ihor Kendiukhov (ORCID: https://orcid.org/0000-0001-5342-1499)
Institutions
- University of Tübingen (DE)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-15
- DOI
- https://doi.org/10.5281/zenodo.24248268
- Primary Topic
- Scientific Computing and Data Management
- Type
- article
- Field-Weighted Citation Impact
- 0.00