With the development of next-generation sequencing technologies, large number of transcripts has been accumulated in public databases. Long non-coding RNAs (lncRNAs), typically above 200 nucleotides in sequence length, have recently attracted increasing interests because of their important roles in various cellular processes. While it is straightforward to distinguishing lncRNAs from most small non-coding RNAs using sequence length as the major criteria, it is much more challenging to differentiate between lncRNAs and mRNAs because they share many similarities. In this study, we present a computational method which can effectively classify lncRNAs and mRNAs using sequence-derived features. The algorithm is organized into a two-layer structure. The first layer consists of multiple Support Vector Machine classifiers, each of which takes input a disjoint set of features such as k-mers and sequence-order correlation coefficient factors. The output of first-layer classifiers then serves as input features to the second-layer Naïve Bayes classifier. The final two-layer structured classifier achieved 99% accuracy and 0.98 MCC on a benchmark dataset of human RefSeq mRNAs and GENCODE lncRNAs. We also applied our method, which was trained on human transcripts, in predicting transcripts from 9 other vertebrates and achieved satisfactory cross-species performance.