a, pLDDT distribution of the resolved parts of PDB sequences (n = 3,440,359 residues), the unresolved parts of PDB sequences (n = 589,079 residues) and the human proteome (n = 10,537,122 residues). b, Performance of pLDDT and the experimentally resolved head of AlphaFold as disorder predictors on the CAID Disprot-PDB benchmark dataset (n = 178,124 residues). c, An example low-confidence prediction aligned to the corresponding PDB submission (7KPX chain C)66. The globular domain is well-predicted but the extended interface exhibits low pLDDT and is incorrect apart from some of the secondary structure. a.a., amino acid. d, A high ratio of heterotypic contacts is associated with a lower AlphaFold accuracy on the recent PDB dataset, restricted to proteins with fewer than 40% of residues with template identity above 30% (n = 3,007 chains) (Methods). The ratio of heterotypic contacts is defined as: heterotypic/(intra-chain + homomeric + heterotypic).