Reviews in Computational Chemistry Volume 18

Application of Scoring Functions in Virtual Screening 71
FlexX score
top 5% = 382 cpds.
FlexX score
top 149
14
23
consensus
list, 149 cpds.
PLP score
top 5% = 382 cpds.
PLP score
top 149
Figure 9 Analysis of the consensus scoring concept with COX-2 as an example.
Numbers in the shaded areas are numbers of active compounds. The larger pie charts
at the top show the numbers of inhibitors in the top 5% of the database in terms of
scores. The smaller pie charts refer to fewer top ranking compounds for better
comparison with the smaller size of the consensus list.
scoring for the top 100 compounds by means of PLP and FlexX scores would
eliminate all but one of the singly hydrogen-bonded inhibitors.
Figure 9 shows a worked consensus scoring example for a virtual screening
experiment on COX-2. (Figure 5 shows the corresponding FlexX and PLP
enrichment curves.) There are 23 inhibitors in the top 5% of the FlexX score
rank list and roughly twice as many in the PLP rank list. Consensus scoring
retains 22 of the actives. Because many inactive compounds are filtered out,
the ratio of actives to false positives increases relative to either of the original
lists. A different picture is obtained when one regards only the top 149 compounds
from the individual FlexX and PLP rank lists—the same number of
compounds that are in the consensus list. It becomes clear that the PLP function
alone performs significantly better than does consensus scoring.
Thus, if one does not know in advance which scoring function will work
better, more robust results can be obtained with consensus scoring. If one has
a rough idea which function works better, one can decrease the number of
false positives more effectively by testing fewer compounds from the top of
a single rank list. Experience from seeding experiments with known inhibitors
or an analysis of the type of binding site of the target can help to identify a
suitable scoring function.
22
33
44