Functionally-Valid Unsupervised Compression of the Protein Space

ProtoNet is an agglomerative binary clustering procedure that provides a hierarchical tree of over 1,000,000 proteins. In this paper we show that the ProtoNet tree is composed of a large number of biologically valid clusters and that the tree expresses functional information. We confirm that several aspects of the tree highly correspond to various external sources, some of which are manually validated and are not based on sequence. Furthermore, an intrinsic parameter of the clustering process was used in order to yield an extensive compression of the tree to only 12.2% of its original size, with minimal loss of biological and functional quality. By applying an automatic method for function assignment, over 5,300 clusters that account for 78% of the remaining clusters following compression can be assigned biological names with a high level of confidence. The clustering and compression processes are robust and unsupervised, allowing a fully automatic and unbiased hierarchical arrangement of all proteins, based only on sequence.