1

Supplementary Materials

Table B.The performances of the best representative binding hypotheses generated for-D-glucosidase inhibitors.

Runa / Hypothesesb / Pharmacophoric Features in Generated Hypotheses / Total
Cost / Cost of null
hypothesis / Residual
Costc / Rd / Global Re
1 / 1f / 3xHBA, Hbic / 90.6 / 174.6 / 84.0 / 0.93 / 0.75
3 / HBA, 2xHBD, Hbic / 95.3 / 174.6 / 79.3 / 0.90 / 0.72
10g / 2xHBA, Hbic, RingArom / 98.5 / 174.6 / 76.1 / 0.91 / 0.77
2 / 3 / 2xHBA, HBD, Hbic / 93.0 / 174.6 / 81.6 / 0.92 / 0.74
5 / 3xHBA, Hbic / 93.7 / 174.6 / 80.9 / 0.92 / 0.75
8 / HBA, 2xHBD, Hbic / 98.3 / 174.6 / 76.3 / 0.89 / 0.76
3 / 4 / HBA, HBD, Hbic, RingArom / 98.3 / 174.6 / 76.3 / 0.92 / 0.76
5 / HBA, HBD, Hbic, RingArom / 98.6 / 174.6 / 76.0 / 0.91 / 0.76
6 / 2xHBD, Hbic, RingArom / 98.7 / 174.6 / 75.9 / 0.90 / 0.77
4 / 4g / 2xHBD, Hbic, RingArom / 98.2 / 174.6 / 76.4 / 0.92 / 0.70
5 / 3xHBA, Hbic / 98.5 / 174.6 / 76.1 / 0.90 / 0.77
9g / HBA, HBD, Hbic, RingArom / 99.5 / 174.6 / 75.1 / 0.90 / 0.73
5 / 3 / 2xHBA, HBD, Hbic / 111.0 / 349.8 / 238.8 / 0.93 / 0.74
8 / HBA, HBD, Hbic, RingArom / 116.5 / 349.8 / 233.3 / 0.92 / 0.73
10 / 2xHBA, Hbic, RingArom / 117.2 / 349.8 / 232.6 / 0.92 / 0.74
6 / 2 / HBA, 2xHBD, Hbic / 118.7 / 349.8 / 231.1 / 0.91 / 0.73
6 / HBA, HBD, Hbic, RingArom / 122.0 / 349.8 / 227.8 / 0.91 / 0.75
7 / 2xHBA, Hbic, RingArom / 123.0 / 349.8 / 226.8 / 0.92 / 0.75
7 / 5 / HBA, 2xHBD, Hbic / 115.6 / 349.8 / 234.2 / 0.93 / 0.74
7 / 2xHBA, HBD, Hbic / 116.1 / 349.8 / 233.7 / 0.92 / 0.75
8 / HBA, 2xHBD, Hbic / 116.2 / 349.8 / 233.6 / 0.92 / 0.74
8 / 2 / HBA, HBD, Hbic, RingArom / 102.0 / 349.8 / 247.8 / 0.95 / 0.71
8 / HBA, HBD, Hbic, RingArom / 122.2 / 349.8 / 227.6 / 0.92 / 0.81
9 / HBA, 2xHBD, Hbic / 122.9 / 349.8 / 226.9 / 0.91 / 0.71

aCorrespond to runs in Table A in Supplementary Materials.

bBest models from their respective clusters, as judged based on F-statistics.

cThe difference between the total cost and the cost of the corresponding null hypotheses.

dThe correlation coefficients between bioactivity estimates and bioactivities of corresponding training set compounds.

eFisher statistics calculated based on the linear regression between the fit values of collected inhibitors (1-41, Fig. 1 and Table 1) against pharmacophore hypothesis (employing the "best fit" option and eq. (4) and their respective β-D-glucosidase bioactivities).

fRank of each hypothesis in each particular run by CATALYST.

gBolded pharmacophores emerged in the best QSAR equations (bolded).

Receiver Operating Characteristic (ROC) Curve Analysis

Selected pharmacophore models (i.e., Hypo10/1, Hypo4/4, Hypo9/4and their shape-complemented versions) were validated by assessing their abilities to selectively capture diverse β-D-glucosidase active compounds from a large testing list of actives and decoys.

The testing list was prepared as described by Verdonk and co-workers [56, 57].Briefly, decoy compounds were selected based on three basic one-dimensional (1D) properties that allow the assessment of distance (D) between two molecules (e.g., i and j): (1) the number of hydrogen-bond donors (NumHBD); (2) number of hydrogen-bond acceptors (NumHBA) and (3) count of nonpolar atoms (NP, defined as the summation of Cl, F, Br, I, S and C atoms in a particular molecule). For each active compound in the test set, the distance to the nearest other active compound is assessed by their Euclidean Distance (eq. (A)):

The minimum distances are then averaged over all active compounds (Dmin). Subsequently, for each active compound in the test set, around 36 decoys were randomly chosen from the ZINC database [58]. The decoys were selected in such a way that they did not exceed Dmin distance from their corresponding active compound.

To diversify active members in the list, we excluded any active compound having zero distance () from other active compound(s) in the test set. Active testing compounds were defined as those possessing β-D-glucosidaseaffinities 0.044.0 μM. The test set included 8 active compounds and 288 decoys.

The test set (296 compounds) was screened by each particular pharmacophore employing the "Best flexible search" option implemented in CATALYST, while the conformational spaces of the compounds were generated employing the "Fast conformation generation option" implemented in CATALYST. Compounds missing one or more features were discarded from the hit list. In-silico hits were scored employing their fit values as calculated by eq. (4).

The ROC curve analysis describes the sensitivity (Se or true positive rate, eq. (B)) for any possible change in the number of selected compounds (n) as a function of (1-Sp). Sp is defined as specificity or true negative rate (eq. (C)) [57, 59].

where, TP is the number of active compounds captured by the virtual screening method (true positives), FN is the number of active compounds discarded by the virtual screening method, TN is the number of discarded decoys (presumably inactives), while FP is the number of captured decoys (presumably inactives).

If all molecules scored by a virtual screening (VS) protocol with sufficient discriminatory power are ranked according to their score (i.e., fit values), starting with the best-scored molecule and ending with the molecule that got the lowest score, most of the actives will have a higher score than the decoys. Since some of the actives will be scored lower than decoys, an overlap between the distribution of active molecules and decoys will occur, which will lead to the prediction of false positives and false negatives [57, 59]. A ROC curve is plotted by setting the score of the active molecule as the first threshold. Afterwards, the number of decoys within this cutoff is counted and the corresponding Se and Sp pair is calculated. This calculation is repeated for the active molecule with the second highest score and so forth, until the scores of all actives are considered as selection thresholds.

The ROC curve representing ideal distributions, where no overlap between the scores of active molecules and decoys exists, proceeds from the origin to the upper-left corner until all the actives are retrieved and Se reaches the value of 1. In contrast to that, the ROC curve for a set of actives and decoys with randomly distributed scores tends towards the Se = 1-Sp line asymptotically with increasing number of actives and decoys [57]. The success of a particular virtual screening workflow can be judged from the following criteria:

1)Area under the ROC curve (AUC).In an optimal ROC curve an AUC value of 1 is obtained; however, random distributions cause an AUC value of 0.5 [57, 59].

2)Overall Accuracy (ACC): describes the percentage of correctly classified molecules by the screening protocol (eq. (D)). Testing compounds are assigned a binary score value of zero (compound not captured) or one (compound captured) [57, 60, 61].

where, N is the total number of compounds in the testing database, A is the number of true actives in the testing database.

3)Overall specificity (SPC): describes the percentage of discarded inactives by the particular virtual screening workflow. Inactive test compounds are assigned a binary score value of zero (compound not captured) or one (compound captured)[57, 60, 61].

4)Overall True Positive Rate (TPR or overall sensitivity): describes the fraction percentage of captured actives from the total number of actives. Active test compounds are assigned a binary score value of zero (compound not captured) or one (compound captured).

5)Overall False Negative Rate (FNR or overall percentage of discarded actives): describes the fraction percentage of active compounds discarded by the virtual screening method. Discarded active test compounds are assigned a binary score value of zero (compound not captured) or one (compound captured).