Gene expression profiles have been extensively discussed as anaid to guide the therapy by predicting disease outcome for thepatients suffering from complex diseases, such as cancer.However, prediction models built upon single-gene (SG) featuresshow poor stability and performance on independent datasets.Attempts to mitigate these drawbacks have led to the developmentof network-based approaches that integrate pathway informationto produce meta-gene (MG) features. Also, MG approaches haveonly dealt with the two-class problem of good versus pooroutcome prediction. Stratifying patients based on theirmolecular subtypes can provide a detailed view of the diseaseand lead to more personalized therapies. We propose and discussa novel MG approach based on de novo pathways, which for thefirst time have been used as features in a multi-class settingto predict cancer subtypes. Comprehensive evaluation in a largecohort of breast cancer samples from The Cancer Genome Atlas(TCGA) revealed that MGs are considerably more stable than SGmodels, while also providing valuable insight into the cancerhallmarks that drive them. In addition, when tested on anindependent benchmark non-TCGA dataset, MG features consistentlyoutperformed SG models. We provide an easy-to-use web service athttp://pathclass.compbio.sdu.dk where users can upload their owngene expression datasets from breast cancer studies and obtainthe subtype predictions from all the classifiers.