We proposed a novel method for protein classification based on phylogenetic profiles. Each protein's profile was extended with extra bits encoding the phylogenetic tree structure and the likelihood, in the form of weights on profile indices, of the protein's functional family membership in each of the reference genomes. The extended profiles were then integrated as part of a kernel of a support vector machine, which was trained in a transductive learning scheme using the EM algorithm to update the weights. Classification accuracy was greatly increased when tested on the proteome of Saccharomyces cerevisiae using the MIPS classification as a benchmark.