Interface prediction can be formulated as a problem in which a binary classsification (interface or not) of all residues of one protein with a given 3D structure is sought. Previous work includes the independent classification of each residue with machine learning approaches. Li et al. considered some of the interdependencies between labels of different residues by interpreting the task as a sequential labeling problem and using a linear-chain conditional random field (CRF) [PMID:17234636].

In contrast, our CRF makes the weaker assumption that the label of one residue is conditionally independent of the labels of residues further than a distance threshold, given the labels of the other residues within threshold distance (3Å - 12Å)

With this approach we achieve more accurate results than the SVM based PresCont server [PMID:22038731] where we consider the output of the PresCont server as the only feature class within our CRF models.

It is important to note that, this server only demonstrates our new method which we describe in our "CRF-based models of protein surface improve protein-protein interaction site predictions". If you are interested in using our method for your own data set, the tools need to be newly trained. We would be very glad to cooperate with you. Please contact with us.

Email: mehmet.gueltas[at]