Improving protein-ligand binding site prediction accuracy by classification of inner pocket points using local features

Table 2 Rescoring Fpocket and ConCavity predictions with PRANK: cross-validation results on CHEN11 dataset and the results of the final prediction model (trained on CHEN11-Fpocket) for all datasets

Dataset	Top-n [%]	Rescored [%]	All [%]	Δ	%possible*	P	R	MCC
Fpocket predictions
CHEN11 (CV)**	47.9	58.8	71	+10.6	47.1	0.60	0.32	0.41
CHEN11***	47.9	67.9	71	+20	86.4	0.87	1.0	0.98
ASTEX	58	63.6	81.1	+5.6	24.2	0.56	0.41	0.46
DT198	37.5	56.2	80.2	+18.8	43.9	0.31	0.38	0.33
MP210	56.6	67.7	78.8	+11.1	50	0.58	0.42	0.47
B48	74.1	81.5	92.6	+7.4	40	0.58	0.45	0.49
U48	53.7	77.8	88.9	+24.1	68.4	0.55	0.36	0.42
ConCavity predictions
CHEN11 (CV)**	47.9	50.7	52.3	+2.8	63.3	0.44	0.76	0.40
CHEN11***	47.9	52.3	52.3	+4.4	100	0.80	0.82	0.75
ASTEX	55.2	62.9	65.7	+7.7	73.3	0.60	0.55	0.46
DT198	45.8	61.5	65.6	+15.6	78.9	0.33	0.55	0.34
MP210	57.4	66.1	68.2	+8.7	80.6	0.63	0.53	0.49
B48	66.7	77.8	81.5	+11.1	75	0.61	0.53	0.47
U48	64.8	74.1	77.8	+9.3	71.4	0.58	0.46	0.43

Abbreviations: P precision, R recall, MCC Matthews correlation coefficient.
^*percentage of improvement that was theoretically possible to obtain by reordering pockets [ Δ / (All – Top-n)].
^**cross-validation results.
^***results where the test set was de facto the same as the training set for the Random Forest classifier (included here only for completeness).