A new approach for HMM based protein sequence family modeling and its application to remote homology classification

Currently probabilistic models of protein families, namely HMMs, are the methodology of choice for remote homology analysis. Unfortunately, the topology of such so-called Profile HMMs is rather complex which, despite sophisticated regularization techniques, is problematic for robust model estimation when only little training data is available. We propose a new HMM based protein family modeling method using building blocks which capture the essentials of particular targets only. They are estimated in a fully data-driven and unsupervised procedure. Contrary to current motif detection procedures we use a feature based protein sequence representation we developed earlier. Such small building blocks are automatically combined to global protein family HMMs which can be applied to remote homology analysis tasks. The results of an experimental evaluation on a challenging task of remote homology classification prove that robust models containing substantially smaller amounts of parameters can be estimated using the new modeling approach. The smaller the number of parameters to be trained, the smaller the number of training samples required which is of major importance for e.g. drug discovery tasks