Skip to main content

Table 11 Overview of training features and performance of naïve Bayesian predictive models created from the String0.7 and Reactome networks, utilizing network-wide and target class-specific training

From: Centrality of drug targets in protein networks

  1. The latter outperformed network-wide training in both databases. Centrality features were chosen for each model with the aid of the Knime forward feature selection workflow (blue boxes on the left hand of the table). The ‘radiality’ feature (marked with an asterisk* in the table), equivalent to average shortest path or closeness centrality, performed well in several models likely due to its narrow range relative to the model training parameters