A comparison of feature selection methodologies and learning algorithms in the development of a DNA methylation-based telomere length estimator

Table 7 Performance scores for models constructed on Dunedin data and tested on EXTEND and TWIN data

Data set	Estimator	Regression algorithm	MAE	MAPE	Correlation
EXTEND	MI-EN TL	Elastic Net Regression	0.640	43.51	0.203
	MI-LARS TL	Least Angle Regression	0.627	42.64	− 0.012
	MI-PLS TL	Partial Least Squares Regression	0.447	34.12	0.076
	MI-SVR TL	Support Vector Regression	0.675	46.06	0.181
	MI-MLP TL	Multi-layer Perceptron	0.637	43.21	0.111
TWIN	F-test-0.01-EN TL	Elastic Net Regression	0.728	106.52	0.119
	MI-LARS TL	Least Angle Regression	0.746	107.55	0.055
	MI-PLS TL	Partial Least Squares Regression	0.678	135.28	− 0.141
	MI-SVR TL	Support Vector Regression	0.780	106.85	− 0.110
	MI-MLP TL	Multi-layer Perceptron	0.742	108.84	0.106

Metrics include MAE, MAPE and Pearson correlation between predicted and actual TL. These estimators utilise mutual information as the initial feature-selection stage for the EXTEND and TWIN data—followed by an array of varying regression algorithms. For comparison the performances of MI-EN TL and F-test-0.01-EN TL are included for the EXTEND and TWIN cases respectively

ISSN: 1471-2105