Drug-Drug Interactions Prediction Using Fingerprint Only

Table 2

Performance of RF classifiers with different combinations of feature types under the composition tenfold cross-validation^#.

Test set	Model	Accuracy	Precision	Recall	F1-measure	MCC^a
ODIT^b test dataset	Addition	79.62%	74.27%	83.18%	78.46%	59.61%
	Subtraction	78.60%	72.65%	82.46%	77.23%	57.63%
	Hadamard	69.19%	65.84%	70.56%	68.10%	38.49%
	Addition + subtraction	79.07%	72.71%	83.32%	77.63%	58.65%
	Addition + Hadamard	79.86%	75.20%	82.99%	78.89%	60.06%
	Subtraction + Hadamard	79.63%	74.74%	82.84%	78.57%	59.56%
	Addition + subtraction + Hadamard	79.39%	73.49%	83.33%	78.08%	59.21%
NDIT^c test dataset	Addition	63.58%	37.60%	78.53%	50.56%	31.88%
	Subtraction	63.56%	39.66%	76.27%	52.01%	30.99%
	Hadamard	62.40%	54.91%	64.56%	59.21%	25.15%
	Addition + subtraction	63.41%	37.72%	77.78%	50.50%	31.34%
	Addition + Hadamard	64.49%	41.56%	76.82%	53.73%	32.64%
	Subtraction + Hadamard	64.39%	42.04%	76.18%	53.99%	32.26%
	Addition + subtraction + Hadamard	63.92%	39.72%	77.16%	52.21%	31.88%

^#Numbers in italics indicate the highest values in the corresponding column. ^aMCC: Mathews correlation coefficient. ^bODIT: One Drug In Train set. ^cNDIT: No Drug In Train set.