PPI-hotspot^ID: A Method for Detecting Protein-Protein Interaction Hot Spots from the Free Protein Structure

Reviewer #1 (Public Review):

Summary:

The paper describes a program developed to identify PPI-hot spots using the free protein structure and compares it to FTMap and SPOTONE, two webservers that they consider as competitive approaches to the problem. On the positive side, I appreciate the effort in providing a new webserver that can be tested by the community but have two major concerns as follows.

(1) The comparison to the FTMap program is wrong. The authors misinterpret the article they refer to, i.e., Zerbe et al. "Relationship between hot spot residues and ligand binding hot spots in protein-protein interfaces" J. Chem. Inf. Model. 52, 2236-2244, (2012). FTMap identifies hot spots that bind small molecular ligands. The Zerbe et al. article shows that such hot spots tend to interact with hot spot residues on the partner protein in a protein-protein complex (emphasis on "partner"). Thus, the hot spots identified by FTMap are not the hot spots defined by the authors. In fact, because the Zerbe paper considers the partner protein in a complex, the results cannot be compared to the results of Chen et al. This difference is missed by the authors, and hence the comparison of the FTMap is invalid. I did not investigate the comparison to SPOTONE, and hence have no opinion.

(2) Chen et al. use a number of usual features in a variety of simple machine-learning methods to identify hot spot residues. This approach has been used in the literature for more than a decade. Although the authors say that they were able to find only FTMap and SPOTONE as servers, there are dozens of papers that describe such a methodology. Some examples are given here: (Higa and Tozzi, 2009; Keskin, et al., 2005; Lise, et al., 2011; Tuncbag, et al., 2009; Xia, et al., 2010). There are certainly more papers. Thus, while I consider the web server as a potentially useful contribution, the paper does not provide a fundamentally novel approach.

Higa, R.H. and Tozzi, C.L. Prediction of binding hot spot residues by using structural and evolutionary parameters. Genet Mol Biol 2009;32(3):626-633.

Keskin, O., Ma, B.Y. and Nussinov, R. Hot regions in protein-protein interactions: The organization and contribution of structurally conserved hot spot residues. J Mol Biol 2005;345(5):1281-1294.

Lise, S., et al. Predictions of Hot Spot Residues at Protein-Protein Interfaces Using Support Vector Machines. PLoS One 2011;6(2).

Tuncbag, N., Gursoy, A. and Keskin, O. Identification of computational hot spots in protein interfaces: combining solvent accessibility and inter-residue potentials improves the accuracy. Bioinformatics 2009;25(12):1513-1520.

Xia, J.F., et al. APIS: accurate prediction of hot spots in protein interfaces by combining protrusion index with solvent accessibility. BMC Bioinformatics 2010;11:174.

Strengths:
A new web server was developed for detecting protein-protein interaction hot spots.

Weaknesses:
The comparison to FTMap results is wrong. The method is not novel.

Reviewer #2 (Public Review):

Summary:

The paper presents PPI-hotspot a method to predict PPI-hotspots. Overall, it could be useful but serious concerns about the validation and benchmarking of the methodology make it difficult to predict its reliability.

Strengths:

Develops an extended benchmark of hot-spots.

Weaknesses:

(1) Novelty seems to be just in the extended training set. Features and approaches have been used before.

(2) As far as I can tell the training and testing sets are the same. If I am correct, it is a fatal flaw.

(3) Comparisons should state that: SPOTONE is a sequence (only) based ML method that uses similar features but is trained on a smaller dataset. FTmap I think predicts binding sites, I don't understand how it can be compared with hot spots. Suggesting superiority by comparing with these methods is an overreach.

(4) Training in the same dataset as SPOTONE, and then comparing results in targets without structure could be valuable.

(5) The paper presents as validation of the prediction and experimental validation of hotspots in human eEF2. Several predictions were made but only one was confirmed, what was the overall success rate of this exercise?