Large cohorts of samples from multiple batches are usually required for global metabolomic studies to characterize the metabolic state of human disease. As such, it is critical to eliminate systematic variation and truly reveal the biologically associated alterations. In this study, we proposed a reference material-based approach (Ref-M) for data correction by liquid chromatography-mass spectrometry and represented by an analysis of multibatch human serum samples. The reference material was generated by mixing serum from healthy donors and distributed to each extraction batch of subject samples. Pooled quality control samples and isotopic internal standards were then applied in each acquisition batch for data quality control. Finally, each metabolite in subject samples was normalized by its counterpart in the reference serum. We demonstrated that Ref-M significantly enhanced the numbers of efficient features and effectively eliminated the batch variation of 522 serum samples of healthy individuals, benign pulmonary nodules, and lung cancer patients. Twenty differential metabolites were identified to distinguish lung cancer from healthy controls in the training set. The discriminant model was validated in an independent data set with an area under the receiver operating characteristics (ROC) curve (AUC) of 0.853. Another 40 serum samples further tested with Ref-M were achieved an AUC of 0.843 by the established model. Our results showed that the reference material-based approach presents the potential to improve the data comparability and precision for biomarker discovery in large-scale metabolomic studies.