Table 1.
Global and local structure quality of adenine-binding proteins from the SOIPPA dataset.
Figure 1.
Effects of target structure distortions on the quality of local alignments of ATP-binding sites.
MCC is Matthew's correlation coefficient calculated against the reference alignments constructed using target crystal structures.
Figure 2.
Prediction of aligned residue pairs using machine learning for SAH-binding proteins from the SOIPPA dataset.
The correlation between the actual pairwise Cα-Cα distances upon the reference alignment of binding sites and those predicted by SVR is shown for (A) crystal structures, (B) high-, and (C) moderate-quality protein models, respectively. (D) The ROC plot for the prediction of equivalent residue pairs using SVC; CS – crystal structures, HQ – high-quality, MQ – moderate-quality models, R – random prediction.
Table 2.
Accuracy of inter-residue distance prediction for adenine-binding proteins from the SOIPPA dataset.
Figure 3.
Construction of sequence order-independent binding site alignments by eMatchSite.
Two target proteins are ATP-dependent DNA ligase (PDB-ID: 1a0iA, yellow) and histamine N-methyltransferase (PDB-ID: 2aotA, red). Left (A–D) and right (E–H) panels show the alignment of binding sites in the crystal structures and protein models, respectively. (A, E) Matrices of pairwise Cα-Cα distances between two binding sites predicted by SVR. Residue indexes are shown in the first column and row. Sets of residue pairs that have the smallest Cα-Cα distances identified by the Kuhn-Munkres algorithm are highlighted in green. (B, F) Sequence order-independent alignments of two binding sites constructed from residue pairs that have the smallest Cα-Cα distances; arrows indicate equivalent pairs. (C, G) Protein structures are superposed according to the local alignment of their binding sites; binding residues and predicted pocket centers are shown as solid sticks and balls, respectively. (D, H) Relative orientation of binding ligands upon the local alignment of target binding sites; ATP in 1a0iA and S-adenosyl-L-homocysteine in 2aotA are shown as solid and transparent sticks, respectively.
Figure 4.
Performance of eMatchSite, PocketMatch and SiteEngine on the SOIPPA dataset of adenine-binding proteins.
The accuracy of local alignment predictors is compared to that using global sequence and structure alignments for (A) crystal target structures, (B) high-, and (C) moderate-quality protein models. TPR and FPR are the true and false positive rates, respectively; gray area corresponds to a random prediction.
Table 3.
Comparison of sequence order-independent binding site alignments constructed by SOIPPA and eMatchSite for adenine-binding proteins.
Table 4.
Comparison of sequence order-independent binding site alignments constructed by SiteEngine and eMatchSite for adenine-binding proteins from the SOIPPA dataset.
Figure 5.
Performance comparison for eMatchSite, PocketMatch, SiteEngine and sup-CK.
Binding site matching is conducted using the (A–C) Kahraman and (D–F) Homogeneous datasets. The accuracy of local alignment predictors is compared to that using global sequence and structure alignments for (A, D) crystal target structures, (B, E) high-, and (C, F) moderate-quality protein models. TPR and FPR are the true and false positive rates, respectively; gray area corresponds to a random prediction.
Figure 6.
Performance of eMatchSite, PocketMatch and SiteEngine on the Steroid dataset.
The accuracy of local alignment predictors is compared to that using global sequence and structure alignments for (A) crystal target structures, (B) high-, and (C) moderate-quality protein models. TPR and FPR are the true and false positive rates, respectively; gray area corresponds to a random prediction.