Fig. 5: Model limitations.

a, Antibody prediction quality increases with the number of model seeds. The quality of top-ranked, low-homology antibody–antigen interface predictions as a function of the number of seeds. Each datapoint shows the mean over 1,000 random samples (with replacement) of seeds to rank over, out of 1,200 seeds. Confidence intervals are 95% bootstraps over 10,000 resamples of cluster scores at each datapoint. Samples per interface are ranked by protein–protein ipTM. Significance tests were performed using by two-sided Wilcoxon signed-rank tests. n = 65 clusters. Exact P values were as follows: 2.0 × 10⁻⁵ (percentage correct) and P = 0.009 (percentage very high accuracy). b, Prediction (coloured) and ground-truth (grey) structures of Thermotoga maritima α-glucuronidase and beta-d-glucuronic acid—a target from the PoseBusters set (PDB: 7CTM). AF3 predicts alpha-d-glucuronic acid; the differing chiral centre is indicated by an asterisk. The prediction shown is top-ranked by ligand–protein ipTM and with a chirality and clash penalty. c, Conformation coverage is limited. Ground-truth structures (grey) of cereblon in open (apo, PDB: 8CVP; left) and closed (holo mezigdomide-bound, PDB: 8D7U; right) conformations. Predictions (blue) of both apo (with 10 overlaid samples) and holo structures are in the closed conformation. The dashed lines indicate the distance between the N-terminal Lon protease-like and C-terminal thalidomide-binding domain. d, A nuclear pore complex with 1,854 unresolved residues (PDB: 7F60). The ground truth (left) and predictions from AlphaFold-Multimer v.2.3 (middle) and AF3 (right) are shown. e, Prediction of a trinucleosome with overlapping DNA (pink) and protein (blue) chains (PDB: 7PEU); highlighted are overlapping protein chains B and J and self-overlapping DNA chain AA. Unless otherwise stated, predictions are top-ranked by our global complex ranking metric with chiral mismatch and steric clash penalties (Supplementary Methods 5.9.1).