JAK2 plays essential roles in transmitting signals from multiple cytokine receptors, and has emerged as a prominent drug target in hematological malignancies. JAK2 kinase activity is negatively regulated by its JH2 domain, in which a gain-of-function mutation is found in the majority of patients with myeloproliferative neoplasms. Understanding of how the JH2 domain regulates JAK2 kinase activity (JH1) thus is urgently needed. In the absence of full-length JAK2 structures, we developed a model of the JAK2 JH1–JH2 complex using computational modeling. We assessed this model by mutating critical residues in the predicted complex interface in JH2 and showed that they indeed hyperactivated JAK2 kinase activity. Our model requires further experimental validation. Nevertheless, it represents a verifiable working hypothesis that facilitates structure-function interrogation of mechanisms underlying JAK2 signaling. Importantly, our model was built by a novel strategy based on allosteric sites on interacting partners. This step-wise computational strategy we devised may be easily adopted for studying novel protein-protein interactions in a general manner.
Several important lessons were learned from our study. First, our study provides “proof-of-principle” evidence that information on allosteric sites of each interacting partner can be used to guide the generation of protein complex structures. In addition to four models constructed based on available kinase dimeric interfaces, we applied the MutInf algorithm to identify sites exhibited correlated torsional motions in JH1 and JH2 to guide protein-protein docking. The MutInf method applies equilibrium molecular dynamics simulations to identify correlated motions between spatially unrelated residues, so that novel allosteric sites might be identified in an unbiased, statistically robust manner 
. Detailed analysis of MD simulation trajectories clearly indicates that Model 2, derived from MutInf, was the most energetically favorable and structurally stable model among the six models built. This work represents the first application of using MutInf in a prospective prediction of sites on two protein domains involved in a protein-protein interface.
Second, a priori knowledge of the starting conformation for JH1 and JH2 is important for model generation. In contrast to Kroemer's model in which JH2 was built in the inactive conformation, our initial JH1–JH2 model was built such that JH1 is in the inactive conformation and JH2 in the active conformation. This is based on the latest findings that JH2 possesses kinase activity and auto-phosphorylates two JAK2 residues to maintain basal auto-inhibition. In our model, the active conformation of JH2 traps JH1 in an inactive conformation via direct interfacial contacts. Our experimental data that mutating the important interfacial residues V706 and L707 in the JH2 activation loop hyperactivated JAK2 strongly support our model. The importance of V706 and L707 would not have been noticed if JH2 were built in an inactive conformation. It should be noted that although it is hard to argue that JH2 is not in an active conformation as it phosphorylates negative regulatory JAK2 sites in the basal state, the exact conformation of JH2 and JH1 would have to await a crystal structure of full-length JAK2.
Third, the presence of linker loops between JH1 and JH2, and between SH2 and JH2 play a critical role in model construction. We modeled the loops after we determined the best packing mode between JH1 and JH2 – a necessary step in refining the final complex structure. The JH1–JH2 loop reduced the spatial sampling needed in protein-protein docking and restrained the inter-domain arrangement although it remains challenging of loop prediction for those longer than 12 residues 
. In addition, we also found that the SH2-JH2 linker loop can stabilize the π stacking interaction between residues F617 and F595 in JAK2-V617F. Consistent with our results, mutations in the SH2-JH2 linker loop (the exon 12 mutations), similar to V617F, hyperactivate JAK2 and are found in patients with myeloproliferative neoplasms 
Mutating critical interfacial residues in JH2 hyperactivated JAK2 kinase activity, lending strong support to our model. Among these critical residues, R588 was previously identified in our random mutagenesis screen of residues essential for JAK2 auto-inhibition. E592 is adjacent to residue S591, where a S591L mutation was identified in the same random screen 
. Differ from Kroemer's model, our model predicts that R588 in JH2 forms strong salt bridge interaction with E1028 in JH1 instead of E890 (Figure S5
). Importantly, our model predicts the critical role of residues V706 and L707 in JH2 in stabilizing the inactive conformation of JH1, which could not have been identified in Kroemer's model (Figure S5
). Surprisingly, mutating the corresponding interfacial residues in JH1, instead of hyperactivating JAK2, resulted in reduced basal JAK2 kinase activity. Among these residues, none are conserved within the kinase family except for I973 (Table S2
). These residues thus are not likely to disrupt the kinase fold or directly reduce enzymatic activity. We envision that they serve dual roles in regulating JAK2 kinase activity. First, they interact with JH2 to trap JH1 in an inactive conformation in the basal state. Second, they regulate JH1 activity upon release from JH2. These JH1 residues may control its conformational transition from an inactive to an active state. Alternatively, they may interact with other JAK2 domains such as the FERM domain to activate JH1 activity. Therefore, mutating these residues, although relieves the inhibitory JH1–JH2 interaction, also hinders JH1 kinase activity. Another hypothesis put forth was that dimerization and autophosphorylation of JH2 might be the predominant mechanism to inhibit JH1 
. However, how JH2 phosphorylation of negative regulatory sites results in JH1 inhibition remains elusive. Our results and the contribution of the different mechanisms in JH2-mediated JAK2 regulation await confirmation by experimental structures and further experiments.
During the revision of this manuscript, the Hubbard group reported the X-ray structure of the JAK2 JH2 domain 
. Superimposition of the crystal structure with our modeled complex structure showed that the two structures are well aligned except in the predicted interfacial regions including the αC helix and the activation loop (Figure S6
). Both regions are highly flexible in our simulations and in simulation results reported by Hubbard's group. Importantly, our predicted JH1–JH2 interfaces can still be identified, especially for residues E592 and V706, which lends further support to our model. In addition, the JH2 activation loop is less open compared to our modeled structure. Future work will utilize this new crystal structure of JH2 to further refine our JH1–JH2 complex model.
In summary, we hypothesized that the JH1–JH2 interface involves sites on each partner with a high degree of correlated motions with other sites (i.e. potential allosteric sites), and tested this working hypothesis using a hierarchical protein-protein docking and refinement protocol. Our JH1–JH2 model generated from this approach is more energetically favorable compared to those generated in parallel based on available kinase dimeric forms. We then tested our model with prospective mutational analyses. We note that our approach – predicting potential allosteric sites on each partner using MutInf and subsequently adding restraints between these sites to guide protein-protein docking – is particularly novel, and should be useful in predicting interface regions involved in other protein-protein complexes. We expect that our JAK2 JH1–JH2 structure model may facilitate the further exploration of the atomic events of regulatory mechanisms in JAK protein family (structure models available at http://www.huanglab.org.cn/JAK2_MODEL
). We also believe that the computational approach we used here will be applicable in predicting novel protein-protein interactions in other systems in general.