Fig 1.
Potent hERG blockers exhibit preferential structural similarity.
(a) Schematic diagram of chemical structure network analysis. The chemical-club coefficient (ChC) measures the density of connections (structurally similar pairs) among query compounds (red nodes) above a given threshold of hERG inhibition (top, red bars). (b) ChC calculation plotted for the MLSMR library for randomized (blue, mean +/- 3 standard deviations for randomized datasets) and observed (red) activity for 10 μM data, where compound adjacency is judged by a Tanimoto Coefficient > 0.7 for FCFP_6 circular fingerprints.
Fig 2.
Continuity of structure variants between blockers and nonblockers in the MLSMR and previous hERG datasets.
Each dataset is represented by a structure network as described in Fig. 1. Compound neighbors are classified as blocker (>50% hERG Inhibition at 10 μM) and nonblocker (<50%). The frequency of compounds with a given number of blocker and nonblocker neighbors in each dataset is plotted, with white cells representing empty data, and the origin representing singleton compounds with no neighbors. Grid points along the vertical axis (horizontal axis) represent compounds with a majority blocker neighbors (nonblocker neighbors). The region along the diagonal represents the transition zone where compounds possess mixed blocker and nonblocker neighbors. This transition zone is illustrated by three example neighborhoods containing blockers (red nodes) and nonblockers (light blue nodes).
Fig 3.
Ensemble modeling segregates compound populations by hERG liability.
(a) Schematic of ensemble voting procedure. MLSMR compounds (left) are divided in 5 test folds (bottom), each receiving binary votes from an ensemble of hERG classifiers (center right) developed using balanced batches of the 80% training set not included in each test fold (top right). (b) The average vote of hERG blocker over all models in the ensemble is recorded as a hERG Blocker Score (hBS). Compounds with high hBS (red histogram bar) have consistent blocker classification, while those with low hBS (dark blue histogram bar) have consistent nonblocker classification. Compounds with intermediate hBS score are denoted by light blue histogram bars. (c) Distribution of hERG inhibition (10 μM) for compounds with High (>0.95), Intermediate (0.05<hBS<0.95) and Low (<0.05) hBS.
Fig 4.
Structural similarity within and between six populations of compounds in the MLSMR assigned by activity-predictability classes.
(a) Summary network visualizing the relationships of six possible combinations of activity (blocker and nonblocker) and predictability (predictable, inconsistent and unpredictable) classes. Each node represents the population of compounds of the same activity-predictability assignment, with edge width representing relative structural similarity quantified by relative connection density (see Methods), within and between each population using the structure network defined in Fig 1. Node sizes for P-B, I-B and U-B represent enrichment of blockers among compounds with high, intermediate, and low hBS compared to the distribution of the entire dataset. Similarly, P-NB, I-NB and U-NB sizes represent enrichment of nonblockers among compounds with low, intermediate, and high hBS. (b) An example cluster that highlights connection patterns within and between P-B, U-NB and I-NB compounds. (c) An example cluster that highlights connection patterns within and between P-NB, U-B and I-NB compounds. (d) An example cluster in which inconsistent and unpredictable compounds are more pronounced. Networks are generated using Cytoscape 2.8.2.
Fig 5.
Novel structural determinants of hERG inhibitions.
(a) (Left) Classical charged hERG pharmacophore consisting of positively charged basic nitrogen (blue) and hydrophobic groups (red), demonstrated by cisapride (I), thioridazine (II) and astemizole (III). (Right) Distribution and density of LogP values for neutral and charged hERG blockers in D2644 collection (right) and MLSMR (left). (b) Density of chemical space mapped using largest and smallest BCUTTM charge descriptors for uncharged hERG blockers in D2644 (Left) and MLSMR (Center), and MLSMR library (Right). Red outlines denote enriched regions for neutral hERG blocker patterns in (c), (d). (c) Distribution of hERG inhibition for compounds containing prazosin fragment (red) compared to MLSMR library (blue). (d) As (c), for compounds containing illustrated triazatricyclo scaffold.
Fig 6.
Experimental evaluation of MLSMR-derived ensemble hERG modeling.
(a) Scatterplot and histogram distribution of predicted blocker numbers for ChemBridge DIVERSet 384-well plates. Experimentally evaluated plates for high (red) and low (light blue) predicted hERG inhibition are highlighted. (b) Correlation of predicted and experimentally observed blockers for eight test plates. (c) Pie graphs of true positive rate (recall) for high and low-risk plates at 10 μM concentration. Area is proportional to the number of experimentally observed blockers. Light color indicates false negatives, dark color true positives.