Manipulations of PPIs (protein–protein interactions) are important for many biological applications such as synthetic biology and drug design. Combinatorial methods have been traditionally used for such manipulations, failing, however, to explain the effects achieved. We developed a computational method for prediction of changes in free energy of binding due to mutation that bring about deeper understanding of the molecular forces underlying binding interactions. Our method could be used for computational scanning of binding interfaces and subsequent analysis of the interfacial sequence optimality. The computational method was validated in two biological systems. Computational saturated mutagenesis of a high-affinity complex between an enzyme AChE (acetylcholinesterase) and a snake toxin Fas (fasciculin) revealed the optimal nature of this interface with only a few predicted affinity-enhancing mutations. Binding measurements confirmed high optimality of this interface and identified a few mutations that could further improve interaction fitness. Computational interface scanning of a medium-affinity complex between TIMP-2 (tissue inhibitor of metalloproteinases-2) and MMP (matrix metalloproteinase) 14 revealed a non-optimal nature of the binding interface with multiple mutations predicted to stabilize the complex. Experimental results corroborated our computational predictions, identifying a large number of mutations that improve the binding affinity for this interaction and some mutations that enhance binding specificity. Overall, our computational protocol greatly facilitates the discovery of affinity- and specificity-enhancing mutations and thus could be applied for design of potent and highly specific inhibitors of any PPI.
PPIs (protein–protein interactions) play a crucial role in virtually all physiological processes in the cell such as proliferation, apoptosis, assembly of macromolecular complexes and network regulation. Hence, modulating PPIs is of great interest for both basic sciences and applied research. PPIs could be modulated through combinatorial methods that involve construction of large combinatorial libraries of protein mutants, displaying them on phage or on cell surfaces and selecting the variants that bind with high affinity to a particular target [1–3]. This approach has proved to be a powerful means of enhancing binding affinity and specificity and for selecting novel binding domains [4–10]. Although very successful in obtaining the end-product, combinatorial methods do not reveal the basic chemical and structural principles that govern protein binding. Moreover, protein variants selected from combinatorial libraries usually contain multiple mutations whose individual effect on binding cannot be explained. However, for many biological applications, it is desirable to engineer single mutations that affect binding energetics in a predictable manner. Computational structure-based methods could provide a fast and convenient way of identifying such mutations in various biological systems.
To facilitate PPI engineering, we developed a computational saturated mutagenesis protocol  based on the framework of the ORBIT protein design package . The protocol utilizes an energy function that has been specifically optimized by our group for design of PPIs [13,14]. For a given PPI with an available high-resolution structure, the protocol first defines the binding interface and the residues in direct contact with the binding interface. Each position in the binding interface is then computationally mutated, the surrounding residues are repacked and the energy of the complex is calculated (Figure 1). The two chains in the complex are then taken apart and the energies for the unbound chains are computed and subtracted from that of the complex. The change in free energy of binding, ΔΔGbind, is determined by subtracting the intermolecular energy of the wild-type complex from that of the mutant complex.
Computing change in free energy of binding
To obtain a comprehensive picture of the binding interface, we probe each binding interface position with all 20 amino acids and compute ΔΔGbind for each substitution (Figure 2). The results could be conveniently visualized by colour coding values of ΔΔGbind from red to blue, where red represents mutations that destabilize the complex, blue represents mutations that stabilize the complex and green represents neutral mutations (Figure 2). Figure 2 shows the binding landscape of an exemplar PPI. Whereas mostly red landscapes indicate that the binding interface is near the maximum of fitness for binding, mostly green and blue landscapes indicate that the interface sequence lies far from the fitness maximum and hence could be substantially improved through mutations.
Computational saturated mutagenesis
To validate our computational saturated mutagenesis method experimentally, we picked two biological systems that represent high- and medium-affinity PPIs. The first PPI is a complex between an essential enzyme, AChE (acetylcholinesterase), and its potent inhibitor Fas (fasciculin) that comes from the venom of the Green Mamba snake. This PPI exhibits a naturally high binding affinity (Kd of 10−10–10−12) that has been attributed to high surface complementarity of the two binding partners and optimized electrostatic interactions with negative charges on the AChE side and positive charges on the Fas side [15,16]. Structural information is available for three homologous complexes where Fas interacts with AChE from three different species: Torpedo californica (tAChE), humans (hAChE) and mouse (mAChE) [15,17,18]. The three structures exhibit a very similar binding mode, where Fas binds at the surface of the enzyme and seals the tunnel leading to the deeply buried enzyme active site. In this study, we considered Fas interactions with two enzyme species, hAChE and tAChE, that exhibit approximately 60% sequence identity in the binding interface region. Although Fas binds in the picomolar range to both enzymes, its affinity for hAChE is ~4-fold tighter. Our goal was to test whether we can identify mutations in Fas that further enhance its affinity to either tAChE or hAChE. Several previous studies involved computational design of affinity-enhancing mutations; however, these studies were performed in low- or medium-affinity PPIs where the probability of finding affinity-enhancing mutations is relatively high [19–24]. In contrast, our study focused on a high-affinity PPI, where such mutations might be extremely rare and hence difficult to find. Indeed, the computational saturated mutagenesis protocol showed that Fas-binding interface is highly optimized for both tAChE and hAChE, with many highly destabilizing mutations and only a few mutations that improve free energy of binding. Interestingly, the Fas-binding interface is more optimized for interactions with hAChE compared with tAChE, in agreement with higher natural affinity of Fas for hAChE. Using computational saturated mutagenesis results as a guide, we chose a relatively large number of Fas single mutants for experimental testing. Among these mutations there were mutations with predicted enhanced affinity to either tAChE or/and hAChE as well as mutations predicted to destabilize the Fas–AChE complexes. We experimentally constructed 25 Fas single mutants and measured their binding affinity to both hAChE and tAChE. Binding measurements identified three mutations that significantly enhance Fas affinity to tAChE and two mutations that enhance its affinity to hAChE, proving that affinity-enhancing mutations could be found even in such a high-affinity complex. All of the affinity-enhancing mutations identified were polar in nature and were predicted to improve electrostatic and hydrogen-bond interactions with the enzyme. This is in contrast with most previous protein design studies that improved PPI affinity through introduction of hydrophobic mutations. In addition, we were able to identify several single mutations that result in more than 10-fold preference of Fas towards one enzyme species, attesting to the fact that species-specific mutations could be computationally designed. The summary of all binding affinity measurements showed that our computational predictions of ΔΔGbind exhibit high correlation with the experimental results. This is despite the simplicity of our computational protocol that includes only side chain, but not backbone, flexibility when calculating ΔΔGbind. Introducing backbone flexibility into our calculations only decreased the correlation between prediction and experiment. In conclusion, our computational saturated mutagenesis protocol was able to identify affinity-enhancing mutations in a high-affinity picomolar complex between Fas and AChE. A relatively small number of potential affinity-enhancing mutations might be a general characteristic of high-affinity complexes whose binding interface sequences might lie close to the fitness maximum for the binding function.
We next wanted to apply our computational saturated mutagenesis protocol on a PPI whose binding interface is not optimized for binding. For this purpose, we picked a complex between an MMP (matrix metalloproteinase) and its inhibitor TIMP (tissue inhibitor of metalloproteinases). MMPs are a family of 24 homologous enzymes that play a major role in degradation of the extracellular matrix and thus control tissue remodelling, cell differentiation and proliferation and other crucial processes . The MMP family of enzymes is regulated by four TIMPs that exhibit 40–50% sequence identity with each other and broad specificity against all MMP members. For our studies we picked MMP14 whose structure in complex with TIMP2 is available (PDB code 1BUV) and MMP9 whose structure is available in the unbound form only (PDB code 1L6J). Owing to high structural homology between the catalytic domains of MMP9 and MMP14, we were able to construct a model for the MMP9–TIMP2 complex. The MMP–TIMP2 complex structures show that TIMP2 contacts the enzyme through four interaction regions. The first interaction region consists of N-terminal residues that insert into the active site of MMP and directly co-ordinate the catalytic Zn2+ ion. Several previous mutational studies directed at enhancing TIMP2-binding specificity towards a particular MMP type were focused on this N-terminal region [26,27]. We, in contrast, did not consider mutations in the direct vicinity of the catalytic Zn2+ since our computational methods cannot predict the effects of mutations on protein catalysis. Since TIMP2 in Nature binds to multiple MMPs, we hypothesized that it is not optimal for binding each particular MMP type. Indeed, computational saturated mutagenesis revealed the non-optimized nature of the TIMP2-binding interface when interacting with MMP14 and MMP9 and predicted many affinity-enhancing mutations at multiple positions. A total of 13 single TIMP2 mutations were experimentally constructed and tested for binding to MMP14 and MMP9. Out of 13 experimentally explored mutations, ten showed enhanced binding affinity to MMP14 and nine to MMP9. Two of the mutations showed very significant (10- and 20-fold) enhancement in affinity to MMP14, testifying to the non-optimal nature of the MMP14–TIMP2 interface and high potential for its improvement. Among the experimentally verified affinity-enhancing mutations, we found both hydrophobic and polar mutations that improved hydrophobic burial and hydrogen-bond interactions respectively. Most of the mutations that enhanced TIMP2 affinity for MMP14 also enhanced its affinity for MMP9, whereas a few mutations had an opposite effect on binding affinity for the two enzymes, demonstrating enhancement in binding specificity. Combination of such specificity-enhancing mutations will be explored in the future in an attempt to create a TIMP2-based specific inhibitor for a particular MMP type, which would have high therapeutic potential. In conclusion, our computational and experimental results in the MMP–TIMP2 system confirm that the TIMP2-binding interface sequence lies far from the fitness maximum and could be improved through multiple mutations for binding to a specific MMP type.
Although the goal of the studies discussed was to find point mutations that enhance binding affinity, single mutations in most cases will not be sufficient to achieve several orders-of-magnitude enhancement in binding specificity. To increase binding specificity significantly, we proposed to introduce a large number of mutations into the protein-binding interface that collectively improve or do not change binding affinity towards one target, but destroy binding to a number of alternative targets . This computational protocol was applied to redesign the entire binding interface of CaM (calmodulin), a universal Ca2+ sensor that binds to hundreds of various proteins with similar affinities and hence possesses low-binding specificity. Using slightly different energy function parameters, we computationally designed six CaM mutants with enhanced binding specificity towards CaMKII (Ca2+/CaM-dependent protein kinase II) and against the phosphatase CaN (calcineurin) . CaM mutants, containing five to nine mutations in the binding interface region were experimentally constructed and tested for binding to the desired target CaMKII and the alternative target CaN. Experimental results showed that five out of six designed CaM mutants exhibited enhancement in binding affinity to CaMKII; however, the enhancement was relatively modest (up to 1 kcal/mol; 1 kcal=4.184 kJ). All six CaM mutants demonstrated decreased binding to the alternative target CaN with a substantially larger effect than that observed for CaMKII (up to 4.5 kcal/mol). All designed CaM mutants exhibited enhancement in binding specificity, although the degree of enhancement differed. The best CaM mutant with nine mutations showed almost three orders of magnitude enhancement in binding specificity towards CaMKII, demonstrating the largest binding specificity improvement obtained by computational design.
Conclusions and future perspectives
Our work in several biological systems demonstrated that computational saturated mutagenesis is a powerful technique that allows us to map binding landscapes of various PPIs, to look at interface optimality, and to identify affinity- and specificity-enhancing mutations. In the high-affinity complex, we were able to identify a few affinity-enhancing mutations, whereas in the medium-affinity multispecific PPIs, we found many mutations that stabilize the complex, thus showing correlation between the functional role of the PPI and the optimality of its binding interface sequence. In the future, we plan to perform a high-throughput study and to test whether this relationship between PPI function and its interface optimality holds for other protein–protein complexes and is a general property of all PPIs.
Protein Engineering: New Approaches and Applications: A joint Biochemical Society/Protein Society Focused Meeting held at the University of Chester, U.K., 10–12 April 2013. Organized and Edited by Ross Anderson (Bristol, U.K.) and Dafydd Jones (Cardiff, U.K.).
This work was supported by the Deutsche Forschungsgemeinschaft [grant number EI 423/2-1], the Abisch Frenkel Foundation and the Israel Science Foundation [grant number 1372/10]