Accuracy of contact prediction based on the overall coevolution score ().

The threshold to remove OTUs with short branches and the number of remaining OTUs that are used for each protein here are listed in Table 2. The numbers of contacts and of sites, and their ratio are listed. Protein structures used to calculate contact residue pairs are listed in Table 1. Neighboring residue pairs within 5 residues () along a peptide chain are excluded in the evaluation of prediction accuracy. TP and FP are the numbers of true and false positives, and their sum is equal to the number of predicted contacts; only predictions for and are listed. DI means the prediction based on the direct information (DI) score published in [16]. PPV stands for a positive predictive value; i.e., . Better values are typed in a bold font. MDPNT stands for the mean Euclidean distance from predicted site pairs to the nearest true contact in the 2-dimensional sequence-position space [16]. Better values are typed in a bold font. MDTNP stands for the mean Euclidean distance from every true contact to the nearest predicted site pair in the 2-dimensional sequence-position space [16]. Better values are typed in a bold font.‡Filters that are based on a secondary structure prediction and cysteine pairs, and were applied to DI in [16], are applied to both DI and . For DI, an additional filter [16] based on residue conservation is also used.‡‡Only the conservation filter is used for DI but no filter is used for .