Feature ranking determined using multinomial logistic regression with LASSO regularization. (A) Feature weights in each class with respect to log lambda, which is a penalty factor to shrink feature weights. Weights of features less discriminative of the three categories shrink to 0 as lambda is increased. Top ranking features are those with non-zero weights at high lambda values: p300, H3K4me1 and MED12 for Enh group; CpG islands, H3K4me3, G + C percent and RNAPII-ser5 for PrL groups. (B) Signatures used in the model chosen from cross validation (log lambda = −4) are shown in each category with their degree of contribution.