Molecular dynamics (MD) simulations can provide detailed analysis of protein interfaces at the atomic level for more  accurate prediction of hot spots (Gonzalez-Ruiz and Gohlke, 2006). Rajamani

∗To whom correspondence should be addressed.

et al. (2004) studied 11 protein complexes and found that anchoring residues in protein interfaces show  restricted  mobility  and may act as hot spots. Kollman and co-workers (Huo et al., 2002) applied MD to find computational alanine scanning of 1:1 human growth hormone–receptor complex and reported a good agreement with the experimental data. Although these energy- and MD-based methods are successful to identify hot spots of inpidual protein complexes, they are not applicable, in practice, for large-scale hot spot predictions due to their computational cost.

The importance of conservation in protein interfaces is well studied (Caffrey et al., 2004; Grishin and Phillips, 1994; Valdar and Thornton, 2001). Residues at protein interfaces (Fraser et al., 2002) and functional sites (Panchenko et al., 2004) were observed to be mutating at a slower pace compared with the rest of the protein surface. There are several studies focusing on the detection of hot spots based on conservation: correlation between hot spot residues and structurally conserved residues were found to be remarkable (Ma et al., 2003). These hot spots are also found to be buried and tightly packed with other residues (Keskin et al., 2005) resulting in densely packed clusters of networked hot spots, called ‘hot regions’. It was shown that central residues are highly conserved in sequence alignments and non-exposed to the solvent in the protein complex and concluded that these residues either correspond to experimental hot spots or are in contact with an experimentally annotated hot spot (del Sol and O’Meara, 2005).

Hot  spots  in  binding  regions  are  located  around  clefts   (Li et al., 2004). Predicted clefts using physicochemical properties and conservation of protein surfaces may correspond to binding hot spot regions (Burgoyne and Jackson, 2006). Another study has illustrated that there is a correlation between energy change and decrease in the accessible surface area (ASA) of inpidual residues as a consequence of complexation (Guharoy and Chakrabarti, 2005). In a recent work, solvent accessibility is combined with conservation in an empirical formula to identify hot spots computationally (Guney et al., 2008). Moreira et al. (2007) have supported that hot spots are protected from solvent by a rim region; however, they concluded that more computational analysis should be applied to elucidate this theory. Another approach to predict hot spots is graph analysis of the proteins. Brinda et al. (2002) have used graph representation of homodimeric protein complexes (residue network). Spectral analysis of the residue network identified some important residues involved in dimer stability that might correspond to hot spots. Recently, a neural network-based approach using various features of interfaces such as sequence profiles, solvent accessibility and evolutionary conservation is employed in computational hot spot prediction (Ofran and Rost,  2007).  The  method  has  advantage of using only sequence; thus, it is applicable when the structure is not available and also when the binding partner is unknown. A hybrid computational model combining decision tree (using atomic contacts, physicochemical properties and shape specificity contributions) with computational alanine scanning method is proposed to predict hot spots (Darnell et al., 2007). In a recent work, Grosdidier and Fernandez-Recio (2008) predict hot spots  by

efficient method to determine computational hot spots in protein– protein interfaces from structure. The method is based on a few simple rules involving solvent  accessibility  and  pair  potentials of residues. Computational effectiveness of this model makes it favorable for hot spot prediction at large scale. As a  result, by using only two features (ASA in complex and pair potential) we reached noteworthy accuracies both in training set and test set. Particularly, use of knowledge-based potentials between  residues is found to be critical in identifying hot spots. We further performed an exhaustive comparison of our empirical method with various ML-based methods by using independent training and testing data. The empirical model, containing solvent accessibility and pair potentials, outperforms other empirical and ML-based methods with its performance values both on ASEdb and BID, 70% and 70% accuracy, respectively.