Integrating water exclusion theory into βcontacts to predict binding free energy changes and binding hot spots

Background Binding free energy and binding hot spots at protein-protein interfaces are two important research areas for understanding protein interactions. Computational methods have been developed previously for accurate prediction of binding free energy change upon mutation for interfacial residues. However, a large number of interrupted and unimportant atomic contacts are used in the training phase which caused accuracy loss. Results This work proposes a new method, βACV ASA , to predict the change of binding free energy after alanine mutations. βACV ASA integrates accessible surface area (ASA) and our newly defined β contacts together into an atomic contact vector (ACV). A β contact between two atoms is a direct contact without being interrupted by any other atom between them. A β contact’s potential contribution to protein binding is also supposed to be inversely proportional to its ASA to follow the water exclusion hypothesis of binding hot spots. Tested on a dataset of 396 alanine mutations, our method is found to be superior in classification performance to many other methods, including Robetta, FoldX, HotPOINT, an ACV method of β contacts without ASA integration, and ACV ASA methods (similar to βACV ASA but based on distance-cutoff contacts). Based on our data analysis and results, we can draw conclusions that: (i) our method is powerful in the prediction of binding free energy change after alanine mutation; (ii) β contacts are better than distance-cutoff contacts for modeling the well-organized protein-binding interfaces; (iii) β contacts usually are only a small fraction number of the distance-based contacts; and (iv) water exclusion is a necessary condition for a residue to become a binding hot spot. Conclusions βACV ASA is designed using the advantages of both β contacts and water exclusion. It is an excellent tool to predict binding free energy changes and binding hot spots after alanine mutation.


Evaluation under leave-one-complex-out cross-validation
The performance of βACV ASA and ACV ASA are also evaluated under the leave-one-complex-out crossvalidation where those mutations in each of 22 protein-protein complexes are used for testing and other mutations for training. Under this evaluation strategy, βACV ASA has a F1 value 0.588 with a precision 0.595 and a recall 0.581, and ACV ASA has a F1 value 0.454 with a precision 0.481 and a recall 0.430. This performance slightly decreases compared with the leave-one-out cross-validation. But the small performance decrease is insignificant, and all the conclusions above are still true using the performance of leave-onecomplex-out cross-validation: βACV ASA is still superior to other methods to predict ∆∆G. Meanwhile, the small performance decrease is reasonable because the number of mutations are not big and leave-onecomplex-out cross-validation has less mutations in training process.

Discussion of using the 396 mutations
The benchmark dataset used by this work consists of 396 residues. However, under the definition of binding interfaces by FoldX, only 378 of them are interfacial residues. Therefore, FoldX could make only 378 predictions for alanine mutations. Similar situation happened to Robetta which made only 338 predictions for the same dataset of 396 residues (details shown in Supplementary Table 3). In fact, some of those residues not detected by FoldX or Robetta are actually in the rim of binding interfaces (rim mutations for short), and the other have close contacts with their partner proteins, such as the two mutations of Pro in position 306 of Chain B in 1A22 and Glu in position 80 of Chain M in 1DX5. Although these two mutation residues are not defined to be in binding interfaces by FoldX and/or Robetta, their smallest spatial distance are less than 6Å to the atoms in their partner proteins. These two mutations can also affect protein binding significantly: the alanine mutation of Pro has ∆∆G 3.31 kcal/mol, and the alanine mutation of Glu has ∆∆G 3.4 kcal/mol. If these two residue mutations were considered to have a small predicted ∆∆G, Robetta would have a lower regression performance. Nevertheless, only the detected mutations by FoldX and Robetta are used in their regression performance assessment. In the classification performance comparison, all 396 mutations are used, and all non-predicted mutations by every classifier are considered to have a small predicted ∆∆G. This is fair to all of the classifiers.

Dataset from BID and the evaluation
The mutations from BID [1] are also used for independent data testing. Those mutations have no explicit ∆∆G but with the class labels 'Strong', 'Intermediate', 'Weak','Insignificant' and so on. Often, the mutations with 'Strong' are considered as binding hot spots. In BID, we found 22 complexes. However, many of them are protein binding to a short sequence of peptides, such as with less than 20 residues in Chain E in 1CDL, B in 1DDM, X in 1dVA, C and D in 1EBP and D in 1JPP, and with less than 40 residues in A in 1DX5, and P in 1K4U. Protein-peptide binding complexes are not considered in this work, because they have different physicochemical properties. Meanwhile, some other complexes have very small interfaces (1IHB and 1KTZ) or have mutations in linear peptide structures (1G3I and 1UB4). They are also not considered because small interfaces have great effect on interfacial residues' ASA. 1GL4 is also not used due to the big difference of the definition of binding hot spots between BID data and the work in [2]. Finally, only protein-protein complexes are used for testing, including 1ES7, 1FAK, 1FE8, 1FOE, 1JAT, 1MQ8, 1NFI, 1NUN and 2HHB.
This dataset only has 37 mutations with 7 binding hot spots, called BID-propro.
Finally, the prediction results on BID are shown in Table 1. It can be seen from Table 1 that on the small BID dataset BID-propro, βACV ASA has F1 value 0.571, higher than FoldX's F1 0.480 and Robetta's F1 0.556. βACV ASA achieves better performance than the existing methods again, although the number of the binding hot spots is too small to have a significant evaluation. Table 1 also shows the prediction results on both protein-protein complexes and protein-peptide complexes in BID with two different binding hot spot definitions. If only the mutations with the label 'Strong' are considered as binding hot spots, FoldX has highest F1 value, while Robetta and βACV ASA have higher F1 values if the mutations with the label 'Strong' or 'Intermediate' are considered as binding hot spots. According the difference between the performances on BID-propro and BID, it seems that βACV ASA is better to be used to predict binding hot spots for proteinprotein complexes than for protein-peptide complexes. This is partially due to that ASA used in βACV ASA affects the prediction performance, but ASA of interfacial residues in protein-protein complexes is different from that from protein-peptide complexes (interfacial residues of protein-peptide complexes might be more exposed than interfacial residues in protein-protein complexes). The reason is supported by the improved performance of βACV on BID where ASA is not used.

3.074
Continued on next page

8.300
Continued on next page      PHE CE2 PHE CZ TRP CD1 TRP CE2 TRP CH2 TRP CE3 TRP CZ3  TRP CG TRP CD2 TRP CZ2 TYR CZ TYR CG TYR CD1 TYR CD2  TYR CE1 TYR CE2 'H+': more than one hydrogen 'H0': without any hydrogens 'N/O ': nitrogen/oxygen atoms without charged 'N+': nitrogen atoms with positively charged 'O-': oxygen atoms with negatively charged 'C NO(N)': carbon atoms with(out) covalent-bond nitrogen or oxygen 'ARC': carbon atoms in an aromatic ring For 'XXXYYZZ' or 'XXX YZZ' or 'XXX YZ ' in the second column, XXX represents a residue type, and Y or YY denotes an atomic name, while Z or Z indicates a specific position of Y or YY in XXX.