Extraction of Protein Domains and Signatures through Unsupervised Statistical Sequence Segmentation

We present a novel information theoretic method for protein domain and statistical signature extraction. We apply a new algorithm [20] for unsupervised segmentation of sequences into alternating Variable Memory Markov sources, to families of protein sequences. The algorithm is based on competitive learning between Markov models, implemented as Prediction Suffix Trees [18], which we have shown in earlier works to model well protein families [6, 7]. By clustering the statistical models themselves, based on rate distortion theory combined with deterministic annealing, we obtain a hierarchical segmentation of sequences between alternating Markov sources that seems to automatically avoid over segmentation. This paper demonstrates the potential of this method in protein sequences analysis. We analyze in detail the resulting segmentation achieved for several diverse protein families, and demonstrate the automatic differentiation between different known domains within the sequences. The method also has the potential to pinpoint exact domain boundaries or sub-divide a family into biological sub-families where other computational tools do not.