Data for: Three correctable biases inflate estimates of structural novelty from predicted-structure databases

Derived data supporting the manuscript Three correctable biases inflate estimates of structural novelty from predicted-structure databases. Contents: per-protein structural-novelty calls for 135,523 proteins across 15 eukaryotic proteomes, with best Foldseek target, TM-score, mean pLDDT and novelty verdict; PDB100 re-verification of all 8,011 confidently novel hypothetical proteins in the panel, with per-protein reassignments and own-genus flags; US-align re-scoring of the 137 Plasmodium falciparum candidates; self-match counts for the five-organism comparison; and the AlphaFold Database dark-cluster re-analysis for releases v3 and v6. Proteome accessions and taxon identifiers were verified against the UniProt REST API on 22 September 2026 and are pinned rather than resolved at runtime. Foldseek 8.ef4e960 was used for forward searches against AlphaFold/Swiss-Prot and 10.941cd33 for PDB100 (release 2025-01-01). The raw Foldseek alignment output (20.9 million rows, ~1.6 GB) is not deposited; it regenerates from the analysis code and the pinned panel, and is available from the author on request. Analysis code: https://github.com/ktosmanprotein/structural-novelty-bias (MIT). The AlphaFold Database cluster data re-analysed here were published by Barrio-Hernandez et al., Nature 2023, under CC-BY 4.0.