Machine learning methods for metabolic pathway prediction

BMC Bioinformatics

Table 3 Best-performing numeric features, ordered by AUC

Feature	AUC	max. ACC	SN (max. ACC)	SP (max. ACC)	max. FM	PR (max. FM)	RC (max. FM)
fraction-reactions-with-Enzymes	0.902	0.878	0.662	0.935	0.715	0.641	0.807
fraction-reactions-present	0.899	0.879	0.618	0.948	0.699	0.612	0.815
fraction-reactions-present-or-Orphaned	0.899	0.879	0.619	0.948	0.7	0.69	0.709
best-fraction-reactions-present-in-linear-path	0.898	0.879	0.662	0.936	0.703	0.682	0.726
evidence-info-content-norm-all	0.894	0.866	0.638	0.927	0.689	0.617	0.781
enzyme-info-content-norm	0.89	0.855	0.69	0.899	0.69	0.584	0.844
enzyme-info-content-unnorm	0.88	0.847	0.665	0.895	0.683	0.556	0.887
evidence-info-content-unnorm	0.875	0.841	0.526	0.925	0.657	0.511	0.918
num-reactions-with-enzymes	0.873	0.838	0.635	0.892	0.681	0.543	0.914
enzymes-per-reaction	0.871	0.842	0.686	0.883	0.688	0.567	0.875

See section "Feature Extraction and Processing" and Section 1 of Additional file 2 for description of features.
Columns 2 through 8 correspond to various performance measures: AUC = area under the ROC curve; max. ACC = maximum thresholded accuracy; SN (max. ACC) = sensitivity at maximum-accuracy threshold; SP = specificity at maximum-accuracy threshold; max. FM = maximum thresholded F-measure; PR (max. FM) = precision at maximum-F-measure threshold; RC (max. FM) = recall at maximum-F-measure threshold.

ISSN: 1471-2105