Table 1 Three categories of pathway-based feature selection algorithms. The filter and embedded methods are two typical types for the gene-based feature selection algorithms. As defined by [32], filter methods access the relevance of features by calculating some functional score while embedded methods search for the optimal subset simultaneously with the classifier construction

Penalty: add an extra penalty term which accounts for the pathway structure to the objective function, then optimize the resulting function to get the final gene subset

Embedded feature selection methods, carry out feature selection and coefficient estimation simultaneously, moderate to heavy computing burden

Need the pathway topology information for all genes, e.g., are they connected and the distance between them

Net-Cox [Zhang et al. 2013] netSVM [Chen et al. 2011]

Stepwise forward: order genes based on one specific statistic, and then add gene one by one until there is no gain on the pre-defined score.

Usually filter methods, the beneath concepts and theory are simple. However, they also inherits the filter methods’ drawbacks of inferior model parsimony and thus high false positive rate.

Usually ignore the pathway topology information, the decision hinges mainly on the genes’ expression values

SAM-GSR [Dinu et al, 2009] SurvNet [Li et al. 2012]

Weighting: create some kind weights according to the pathway knowledge and then combine with other feature selection methods to identify the relevant genes

With different weights, the chance of those “driving” genes with subtle change being selected increases. However, if the estimated weights subject to big biases, the resulting model might even be inferior to those without weights.

Account for the pathway topology information.

RRFE [Johannes et al. 2010] DRW [Liu et al. 2013]