Table 2.

The ${\text{R}}^2$ between the actual solubility scores and those predicted by GraphSol based on individual feature groups, removing each feature group from the final GraphSol model, and recursively adding feature groups according to their importance, respectively

Feature groupsa	CVd	Ind. test	Feature groupsb	CVd	Ind. test	Features groupsc	CVd	Ind. test
–	–	–	GraphSol	0.476 $\pm$ 0.014	0.483	–	–	–
BLOSUM	0.329 $\pm$ 0.016	0.317	− BLOSUM	0.460 $\pm$ 0.011	0.465	BLOSUM	0.329 $\pm$ 0.016	0.317
AAPHY7	0.293 $\pm$ 0.014	0.289	− AAPHY7	0.465 $\pm$ 0.012	0.479	+ SPIDER3	0.413 $\pm$ 0.012	0.409
PSSM	0.333 $\pm$ 0.012	0.332	− PSSM	0.457 $\pm$ 0.017	0.467	+ PSSM	0.456 $\pm$ 0.011	0.453
HMM	0.337 $\pm$ 0.015	0.341	− HMM	0.455 $\pm$ 0.016	0.458	+ HMM	0.465 $\pm$ 0.012	0.479
SPIDER3	0.231 $\pm$ 0.019	0.227	− SPIDER3	0.428 $\pm$ 0.018	0.449	+ AAPHY7	0.476 $\pm$ 0.014	0.483

Italic values indicate the performance of using all feature groups in our model

^a Performances based on individual feature groups

^b by removing each feature group from all feature

^c by adding feature groups recursively

^d Performances by the fivefold cross-validation