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

Abstract

Motivation. Predicted-structure databases have made it routine to ask what fraction of a proteome adopts a fold with no known relative. Reported novelty rates vary widely between studies, and it is unclear how much of that variation is biological rather than methodological. A novelty call is a negative result, so it inherits every limitation of the search that failed to find a match. Results. We identify three independent and correctable sources of inflation, quantified across 135,523 proteins from 15 alveolate genomes. Model confidence: apparent novel-fold rate falls monotonically with mean pLDDT and filtering changes the estimate by a median factor of 1.83, in every genome examined. Reference-database scope: searching only AlphaFold/Swiss-Prot missed 105 of 316 P. falciparum proteins that have a solved match in the PDB. Reference-database representation: when the query organism is itself in the reference set, queries match their own entries and apparent novelty measures curation status rather than structure. Applied together, the estimate moves from 54.3% to 20.7%. Re-analysing the largest published survey of structural novelty, the same gradient is present across 2.6 million AlphaFold Database clusters, and the reported dark fraction moved 10.7 percentage points between database releases with the method held fixed. We provide a reporting checklist. Availability. Analysis code at https://github.com/ktosmanprotein/structural-novelty-bias under an MIT licence; derived data at doi:10.5281/zenodo.23045417. All reported values regenerate from the source tables. Preprint. Not peer reviewed.