biomedcentral.com/1471-2105/16/S4/SPage 10 ofshould be expected in
Tp://www.biomedcentral.com/1471-2105/16/S4/SPage 10 ofshould be expected in typical molecular docking and virtual screening campaigns. For each protein family in that experiment's test set, there is at least one protein family in the training set of our proposed NN and RF SFs, but the two sets share no protein-ligand complexes when these pairs of compounds are considered as whole biological units. We describe here a more stringent experiment to assess the generalization of the NN and RF SFs when they are applied to score ligands for novel drug targets. In this experiment, we evaluate the BA predictive accuracy of the NN SFs on four protein families not present in their training set. These protein families are the most frequent in our data and include 112 HIV protease, 73 trypsin, 44 carbonic anhydrase, and 38 thrombin complexes. A test set for each of these protein families wasconstructed by sampling all complexes formed by that protein from the training (Pr ) and the test (Cr ) sets. The training complexes corresponding to each of these four test sets are the remaining protein-ligand pairs in Pr. For each protein family, we fitted the proposed NN and RF models to the corresponding independent training complexes and validated them on the test set complexes that are formed between that type of protein and a unique set of co-crystallized ligands. The prediction accuracy of our proposed models and the top four conventional scoring functions on complexes formed by the four protein types are shown in Table 3. Examining the upper portion of the table for the four families where the test and training sets are disjoint for the NN and RF SFs, we notice that the predictiveTable 3 Comparison of the scoring powers of BsN-Score, BgN-Score, SNN-Score, Random Forests (RF), and the four top performing conventional 1-phenyl-4-(4,4,5,5-tetramethyl-1,3,2-dioxaborolan-2-yl)-1H-pyrazole SFs PubMed ID:https://www.ncbi.nlm.nih.gov/pubmed/14704548 on four protein-family-specific tests setsHIV protease (N = 112) Scoring function X-Score::HPScore BsN::XARG RF::XARG BgN-Score::XARG SYBYL::ChemScore DrugScore::PairSurf DS::PMF04 SNN-Score::X RF::XARG BsN-Score::XARG BgN-Score::XARG SNN-Score::X Scoring function DS::PLP2 SYBYL::G-Score SYBYL::ChemScore BsN-Score::X SNN-Score::X SYBYL::PMF-Score BgN-Score::XA RF::XARG BsN-Score::XARG RF::XARG BgN-Score::XARG SNN-Score::X1Trypsin (N = 73)RpRsSD 1.54 1.56 1.519 1.58 1.58 1.59 1.61 1.64 0.44 0.64 1.02 1.RMSE 1.509 1.705 1.719 1.860 2.255 0.588 0.710 1.024 1.D N Y Y Y U U U Y N N N NScoring function SYBYL::ChemScore DS::Ludi2 X-Score::HSScore DS::PLP2 BgN-Score::XAR RF::XAR BsN-Score::AR SNN-Score::X BsN-Score::XARG RF::XARG BgN-Score::XARG SNN-Score::XRp1 0.829 0.823 0.817 0.797 0.776 0.774 0.766 0.735 0.937 0.934 0.892 0.829 Rp1 0.756 0.722 0.699 0.697 0.667 0.667 0.666 0.651 0.913 0.910 0.858 0.Rs2 0.773 0.791 0.824 0.774 0.719 0.753 0.709 0.672 0.920 0.08 0.848 0.789 Rs2 0.704 0.726 0.637 0.693 0.672 0.626 0.586 0.622 0.938 0.934 0.876 0.SD3 0.95 0.96 0.97 1.02 1.06 1.07 1.08 1.14 0.59 0.60 0.76 0.940 SD3 1.38 1.48 1.58 1.52 1.58 1.58 1.58 1.61 0.86 0.86 1.08 1.37 -RMSE4 U U N U Y Y Y Y N N N ND0.341 0.290 0.289 0.287 PubMed ID:https://www.ncbi.nlm.nih.gov/pubmed/12113769 0.255 0.225 0.183 0.039 0.964 0.918 0.848 0.748 Rp1 0.800 0.706 0.699 0.674 0.631 0.627 0.625 5-Fluoro-3-nitropyridin-2(1H)-one 0.601 0.948 0.910 0.884 0.0.339 0.230 0.219 0.209 0.228 0.170 0.200 0.048 0.975 0.922 0.808 0.716 Rs2 0.772 0.646 0.631 0.434 0.451 0.618 0.423 0.374 0.921 0.860 0.766 0.1.401 1.070 1.133 1.119 1.209 0.678 0.657 0.805 0.957 RMSE4 1.433 1.552 1.603 1.674 1.737 1.155 1.125 1.320 1.Carbonic anhydrase (N = 44) SD3 0.84 0.99 1.00 1.03 1.