Construction of training and testing datasets. We built the training and testing datasets step by step. First, we collected data from five different data sources. Then, squid, RNAfold and UNAfold were employed to further filter the data. Finally, we constructed one training set (697 positive samples and 5428 negative samples) and two testing sets: one contains 27 bran-new hsa and aga pre-miRNA, the other contains 5238 other hairpin sequences in miRBase13.0 besides hsa and aga.