Introduction: A challenge of cancer therapy is to optimize therapeutical options to individual patients. Cancers with similar histology may show dramatically different responses to therapy, indicating that a refined approach needs to be developed to classify tumors by intrinsic characteristics that may predict response to chemotherapy. Global expression profile-based classification has the potential to identify such tumor-intrinsic subclasses. Pemetrexed effectiveness has been related to the expression of its target thymidylate synthase. The relatively frequent resistance of squamous cell carcinoma to Pemetrexed is correlated with high levels of thymidylate synthase expression. Methods: A global expression profile-based molecular classification of non-small cell lung cancer (NSCLC) was performed. Gene expression was used to predict Pemetrexed responsiveness. The distinct molecular attributes of NSCLCs predicted likely to be resistant to Pemetrexed were bioinformatically characterized. We tested if routine immunohistochemical markers can be used to distinguish putative Pemetrexed responders, predicted by gene signatures, from nonresponders. Results: Ninety NSCLCs were divided into six subclasses by gene expression signatures. The relevance of this novel phenotyping was linked to other tumor characteristics. Two of the subclasses correlated to putative Pemetrexed resistance. In addition, the identified signature genes characterizing putative Pemetrexed responsiveness predicted therapeutic benefit in a subset of squamous cell carcinoma. Conclusions: Gene expression signatures can be used to identify NSCLC subgroups and have potential to predict resistance to Pemetrexed therapy. We suggest that a combination of classical pathological markers can be used to identify molecular tumor subclasses associated with predicted Pemetrexed response.