A discrete protein subset drives structure prediction discordance in orphan proteins
Structure and disorder predictors are increasingly used as decision-grade tools in protein engineering and in the analysis of newly emerged proteins, yet how the current state-of-the-art behaves on sequences outside the well-charted evolutionary space remains poorly characterised. We previously reported that AlphaFold2 confidence and the disorder predictor flDPnn produced discordant predictions for naturally evolved de novo Drosophila proteins and for shuffled sequences. Here, we revisit the comparison with AlphaFold3 and the best-performing disorder predictor PUNCH2 on the same sequence sets together with conserved Drosophila proteins and intrinsically disordered proteins. The discordance persists: pLDDT correlates positively with PUNCH2 disorder in random and de novo proteins and negatively with β-strand fraction, opposite to the conserved and disordered baselines. A class-specific, score-defined driver subset jointly captures the unusual high-pLDDT, high-disorder, low-strand combination and contains 24.5% of de novo, 29.4% of random, 5.1% of conserved, and 1.3% of disordered proteins. Removing this subset normalises the correlations. A held-out classifier trained on architectural and compositional features that were not used in the driver definition recovers the subset, with helix and coil fraction, sequence length, entropy and hydropathy as the strongest predictors. The discordance is therefore not a sequence-class artefact but a localised, compositionally identifiable phenotype that current predictors handle in a non-canonical way - a concrete failure mode that protein designers and other working on sequences remote in sequence space should be aware of when relying on predictor outputs.