An apparatus, computer program product, and method are provided for the determination of one or more components of an EC number through the application of a level-by-level modeling approach capable of conducting feature reconstruction and classifier training simultaneously, based on encoded aspects of a sequence listing for a protein with an unknown function. The method includes receiving a sequence source data object associated with an enzyme; extracting a sequence data set from the sequence source data object; encoding the sequence data set into a first and second encoded sequence; generating a first predicted characteristic of the enzyme by applying the first and second encoded sequence to a first level of a model comprising a plurality of levels; and generating a second predicted characteristic of the enzyme by applying the first and the second encoded sequences to a second level of the model comprising a plurality of levels.