Current approaches for the functional annotation of proteins rely on training a classifier based on a fixed reference database. As more genes are sequenced, the size of the reference database grows and classifiers are retrained with the old and some new data. Considering the ever-increasing number of (meta-)genomic data, repeating this process is computationally expensive. An alternative is to update the classifier continuously based on a stream of data. Thus, in this study, we propose an incremental and semi-supervised learning approach to train a classifier for the functional analysis of protein sequences. Our method proves to have a low computational cost while maintaining high accuracy in nredicting protein functions.