Molecular subtypes have been recently derived for various types of cancer, in an attempt to characterize the intertumoral heterogeneity. In this work we explore the possibility of constructing predictors for molecular subtypes based on histopathology images. For this, we introduce a novel 2-level bag-of-features method and we apply it to a collection of colorectal cancer samples. The resulting image features capture some relevant tumor morphology patterns and led to a classifier performing similarly to one constructed from features annotated by an expert pathologist. The significance of our results extends beyond subtype prediction since they demonstrate a possible approach to multimodal (histopathology and molecular) data mining and biomarker identification.