The computational prediction for prokaryotic genes have been carried out either with very simple assumption of long open reading frame (ORF) or with Markov models with many parameters determined through machine learning algorithms. The latter usually gives highly reliable prediction of genes, but it shed little light to the structure of genes under study because of an enormous amount of parameters. We attempt to develop a mathematical model that explicitly accounts for universal features related to genes and its translation mechanisms. Our model contains relatively fewer parameters that have clear biological meaning. We design also an algorithm MED 2.0 that enables an unsupervised learning process for genome-specific parameters before the prediction of genes. The MED 2.0 not only predicts a set of genes for any newly sequenced prokaryotic genome, but also yields related parameters characterizing gene starts. We report the performance of the algorithm and discuss how mathematical models may help to gain biological insight.