Motivation: Nuclear receptors (NRs) play a role in all developmental and physiological processes and are important drug targets in a wide variety of disease and healthy states. In the past years, to identify NRs and their subfamilies with high throughput and low-cost, many machine learning methods have been introduced. However, these predictors are all developed based on old dataset in the NucleaRDB, what’s more, no feature selection technique is employed, so that the performances are very limited. Result: In this study, a feature selection based two-level predictor, called NRPred-FS, is developed that can be used to identify a query protein as a nuclear receptor or not based on its sequence information alone, if it is, the prediction will be automatically continued to further identify it among the following eight subfamilies: (1) Thyroid hormone like (NR1), (2) HNF4-like (NR2), (3) Estrogen like, (4) Nerve growth factor IB-like (NR4), (5) Fushi tarazu-F1 like (NR5), (6) Germ cell nuclear factor like (NR6), (7) knirps like (NR0A), and (8) DAX like (NR0B). The nuclear receptor sequences are encoded as sequence-derived feature vectors formed by incorporating various physicochemical and statistical features. Furthermore, the features set are optimized by forward feature selection algorithm for reducing the feature dimensions and for getting higher classifying accuracy.