a, TM-score distribution for AlphaFold evaluated on a held-out set of template-filtered, long PDB chains (n = 151 chains). Includes recent PDB proteins with more than 800 resolved residues and best 50% coverage template below 30% identity. b, Correlation between full chain TM-score and pTM on the same set (n = 151 chains), Pearson’s r = 0.84. The ground truth and predicted structure are shown for the most over-optimistic outlier (PDB: 6OFS, chain A). c, pTM distribution on a subset of the human proteome that we expect to be enriched for structurally novel multidomain proteins (n = 1,165 chains). Human proteome predictions comprise more than 600 confident residues (more than 50% coverage) and no proteins with 50% coverage templates. d, Four of the top hits from the set shown in c, filtering by pTM > 0.8 and sorting by number of confident residues. Proteins are labelled by their UniProt accession. For clarity, regions with pLDDT < 50 are hidden, as are isolated smaller regions that were left after this cropping.