Table 2.
Performance of Various Virtual Screening Methods on the DUD-E Set
| Comparison to FINDSITEcomb2.0 and AutoDock Vinaa | ||||
|---|---|---|---|---|
| method | EF0.01 | AUPR | top 100 precisionb | top 100 recall |
| FRAGSITE | 30.20 | 0.397 | 0.475 0.557 | 0.305 |
| FRAGSITE_MACCSc | 28.28 | 0.367 | 0.459 | 0.282 |
| FRAGSITE_FP2d | 29.79 | 0.387 | 0.476 | 0.297 |
| FRAGSITE_no-mTC | 22.12 | 0.283 | 0.355 | 0.227 |
| FRAGSITE_no-DOT | 27.23 | 0.358 | 0.438 | 0.284 |
| FRAGSITE_no-HADA | 23.42 | 0.283 | 0.386 | 0.240 |
| FINDSITEcomb2.0 | 25.22 | 0.321 | 0.416 0.557 | 0.257 |
| AutoDock Vina:11 experimental target structure | 9.13 | 0.093 | 0.151 | 0.093 |
| AutoDock Vina: modeled target structure | 3.57 | 0.045 | 0.063 | 0.045 |
| Comparison to AtomNet and CNN Scoringe | ||
|---|---|---|
| method (no. of targets) | AUC | no. of targets having an AUC > 0.9 (%) |
| FRAGSITE (102) | 0.910 | 73 (71.6%) |
| FRAGSITE (102) (experimental target structure) | 0.924 | 77 (75.5%) |
| FINDSITEcomb2.0 (102) | 0.874 | 61 (59.8%) |
| FINDSITEcomb2.0 (102) (experimental target structure) | 0.892 | 65 (63.7%) |
| CNN scoring (102) | 0.868 | 49 (48.0%) |
| FRAGSITE (randomly selected 30)f | 0.915 | 20 (66.7%) |
| FRAGSITE (randomly selected 30, experimental target structure) | 0.916 | 21 (70.0%) |
| FINDSITEcomb2.0 (randomly selected 30)f | 0.881 | 16 (53.3%) |
| FINDSITEcomb2.0 (randomly selected 30, experimental target structure) | 0.888 | 18 (60.0%) |
| AtomNet (30) | 0.855 | 14 (46.7%) |
Since FRAGSITE and FINDSITEcomb2.0 perform similarly on experimental and modeled target structures, we present only results with modeled target structures. We have generated AutoDock Vina results locally using its default settings.
The second number is the precision of consensus prediction of FRAGSITE and FINDSITEcomb2.0.
FRAGSITE using the 256 bit MACCS fingerprint generated by Open Babel.60
FRAGSITE using the 1024 bit FP2 fingerprint generated by Open Babel.60
A sequence identity cutoff of 80% is used by both FINDSITEcomb2.0 and FRAGSITE for target structure modeling and template ligand selection and training in boosted tree regression.
Since AtomNet was only tested on 30 DUD-E targets and their identities are not known, we randomly selected 30 targets for comparison to AtomNet.