Distinct allosteric networks in CDK4 and CDK6 in the cell cycle and in drug resistance
1Cancer Innovation Laboratory, National Cancer Institute, Frederick, MD 21702, U.S.A.
2Laboratory of Biochemistry and Molecular Biology, National Cancer Institute, Bethesda, MD 20892, U.S.A.
3Computational Structural Biology Section, Frederick National Laboratory for Cancer Research, Frederick, MD 21702, U.S.A.
4Department of Human Molecular Genetics and Biochemistry, Sackler School of Medicine, Tel Aviv University, Tel Aviv 69978, Israel
*Author for correspondence: Ruth Nussinov, Mardo Kõivomägi Email: NussinoR@mail.nih.gov mardo.koivomagi@nih.govAbstract
Cyclin-dependent kinases 4 and 6 (CDK4 and CDK6) are key regulators of the G1-S phase transition in the cell cycle. In cancer cells, CDK6 overexpression often outcompetes CDK4 in driving cell cycle progression, contributing to resistance against CDK4/6 inhibitors (CDK4/6i). This suggests distinct functional and conformational differences between these two kinases, despite their striking structural and sequence similarities. Understanding the mechanisms that differentiate CDK4 and CDK6 is crucial, as resistance to CDK4/6i—frequently linked to CDK6 overexpression—remains a significant therapeutic challenge. Notably, CDK6 is often upregulated in CDK4/6i-resistant cancers and rapidly proliferating hematopoietic stem cells, underscoring its unique regulatory roles. We hypothesize that their distinct conformational dynamics explain their differences in phosphorylation of retinoblastoma protein, Rb, inhibitor efficacy, and cell cycle control. This leads us to question how their dissimilar conformational dynamics encode their distinct actions. To elucidate their differential activities, molecular mechanisms, and inhibitor binding, we combine biochemical assays and molecular dynamics (MD) simulations. We discover that CDK4 and CDK6 have distinct allosteric networks connecting the β3-αC loop and the G-loop. CDK6 exhibits stronger coupling and shorter path lengths between these regions, resulting in higher kinase activity upon cyclin binding and impacting inhibitor specificity. We also discover an unrecognized role of the unstructured CDK6 C-terminus, which allosterically connects and stabilizes the R-spine, facilitating slightly higher activity. Our findings bridge the gap between the structural similarity and functional divergence of CDK4 and CDK6, advancing the understanding of kinase regulation in cancer biology.
Graphical abstract
Article notes
Competing Interest Statement
The authors have declared no competing interest.
Summary of Updates:
1Introduction
The almost synonymous and interchangeable cyclin-dependent kinases 4 and 6 (CDK4 and CDK6), with relatively minor sequence and structural differences, differentially regulate the G1-S phase transition of the cell cycle in CDK4/6 inhibitor (CDK4/6i)-resistant cancer cells, adult tissues, and hematopoietic stem cells (HSCs).[1–5] The crucial question of the molecular mechanisms by which overexpressed CDK6 in cancer resistant cells facilitate cell cycle progression— in lieu of CDK4— remains unanswered.
Cyclin dependent kinases (CDKs) and their regulatory subunits, cyclins, are essential in the ordered progression of cells through the cell cycle.[6] Dysregulation of cyclin-CDK complexes leads to uncontrolled cell growth, making them promising targets for anticancer therapies.[7] Closely evolutionarily related, CDK4 and CDK6 are activated by cyclin D binding and phosphorylation by CDK7 (Fig. S1).[8, 9] The two share overlapping functions in regulating progression through G1 by phosphorylating retinoblastoma protein (Rb).[10] Phosphorylation relieves Rb-mediated inhibition of E2F transcription factors, in turn stimulating expression of genes required for the transition into S-phase (such as cyclin E) and committing cells to replication (Fig. 1A).[11] Because of their critical role in initiating the cell cycle, dual CDK4/6 inhibitors (palbociclib, abemaciclib, ribociclib) are used as chemotherapeutics in HR positive and HER2 negative breast cancer. However, their use is associated with a range of side effects, most notably neutropenia.[7, 12, 13] Hematopoietic stem cells (HSCs) rely on CDK6 to drive G1 progression, whereas most cell types, including breast cancer cells, utilize CDK4. Despite their sequence and structural similarity, CDK4 and CDK6 differentially regulate G1 progression in a variety of cell types and exhibit remarkable diversity in regulation, activation, and substrate recognition. In the case of rapidly dividing HSCs, the greater intrinsic catalytic activity of CDK6 may facilitate faster entry into the cell cycle. A CDK4 specific inhibitor, atirmociclib, has successfully been shown to halt breast cancer growth with reduced effect on neutrophil levels. However, an in depth understanding of the mechanistic differences between CDK4 and CDK6, which could inform better drug design, has not yet been established. This is likely in part due to existing static structural analyses failing to capture allosteric networks which modulate activity and substrate specificity through dynamic communication.[14] CDK4 and CDK6 likely rely on distinct allosteric pathways to mediate their unique functions, which may explain their specialized roles in the cell cycle and underlie differences in drug sensitivity.[15–19] These subtle structural differences may also underlie the development of drug resistance specifically through CDK6 overexpression. When in complex with the canonically inhibitory protein, INK4, CDK6 adopts a drug resistant yet ATP-binding competent state through which it can still sustain G1 progression.[38] The differential response to inhibitors suggests distinct allosteric mechanisms of CDK4 and CDK6 which are not yet understood.[3, 7, 20–25] Understanding these allosteric networks could be key to uncovering the unique mechanisms of these kinases and improved inhibitors.
Previously, our work resolved the question of how cyclin-D/CDK4 evolved to have slower catalytic phosphorylation in the long G1 phase, whereas cyclin-E/CDK2 evolved to accomplish fast phosphorylation of its downstream targets during the short G1-to-S transition timescale.[26, 27] We explained the mechanisms, on the conformational level, through which cyclin-D/CDK4 and cyclin-E/CDK2 complexes have adapted to their distinct functional requirements in the G1 phase, and G1/S phase transition, of the cell cycle. We clarified why CDK2 can transition from its inactive to the active state faster than CDK4,[28] and how the G1/S CDK2 complex can maintain faster phosphorylation rate than the G1 CDK4 complex.[29] We suggested the CDK2 structural features which are at play, including (i) stable G-loop, (ii) stable ATP binding site, (iii) stable hydrogen bonds between the hinge region and the adenosine ring of ATP, (iv) stronger protein-protein interactions between the kinase and its activating partner (cyclin) and (v), the stable hydrophobic R-spine that form the backbone of the kinase. These features are necessary but not sufficient for efficient catalysis. Building on these, here we explain how CDK4 and CDK6, despite sharing over 70% sequence identity, similar overall structures, and the same D-type cyclin partner, can exhibit certain distinct functional roles.
Here, we combine biochemical assays and molecular dynamics (MD) simulations analyses to dissect the distinct allosteric regulation of CDK4 and CDK6 and their drug resistance. We attribute the different activities of CDK4 and CDK6 to the local dynamics in the loops (β3-αC, β4-β5, G-loop) in the N-lobe. Our analysis shows that when engaged with cyclin-D, the β3-αC loop allosterically regulates the G-loop, a structure that caps the active site and is essential to kinase activity. In CDK6, the path length between the β3-αC loop and the G-loop is shorter, and the coupling strength between these regions is greater compared to CDK4, enhancing CDK6 activity. Of the few sequence differences between CDK4 and CDK6, the extra sequence in the C-terminus of CDK6 stands out, as it allosterically stabilizes the R-spine, enhancing its catalytic activity. This novel finding extends the long-standing understanding of the roles of the C-terminus in kinase activity,[30, 31] provides a deeper understanding of CDK4/6 function, and may guide the development of next-generation inhibitors.
2Materials and Methods
Biochemical assays for kinase activity
Purification of cyclin/CDK complexes
Human cyclin-D1/CDK4/6 fusion complexes were purified from budding yeast cells (Table S3) using a FLAG affinity purification method, modified from a previous protocol.[32] Briefly, N-terminally 3XFLAG-tagged cyclin/CDK fusions with a glycine-serine linker (3XGGGGS) were cloned into high-copy yeast vectors (pRS425) and overexpressed from the GAL1 promoter. The overexpressed 3XFLAG-tagged complexes were then purified by one-step immunoaffinity chromatography using ANTI-FLAG M2 affinity agarose beads (Sigma-Aldrich A2220) and eluted with 0.2 mg/mL 3XFLAG peptide (Sigma-Aldrich F4799).[33]
Purification of substrate protein
N-terminally 6His-tagged Rb C-terminal fragment (RbC) recombinant protein (aa: 772-928) was expressed in the E. coli strain BL21(DE3)RIL and the purification was performed using cobalt affinity chromatography. Proteins were eluted using buffer containing 25 mM HEPES pH 7.4, 300 mM NaCl, 10% glycerol, and 200 mM imidazole.
In vitro kinase assay
In vitro kinase assays were performed by incubating 1 µM RbC with cyclin-CDK fusion constructs (in the low nM range) in a reaction mixture containing 100 mM HEPES pH 7.4, 150 mM NaCl, 5 mM MgCl2, ∼0.02 mg/ml 3XFLAG peptide, 5% glycerol, 3 mM EGTA, 0.1 mg/ml BSA and 500 μM ATPγS. Reactions were quenched at 8 minutes by addition of EDTA to a final concentration of 25 mM. P-nitrobenzyl mesylate (pNBM) was added to a final concentration of 2 mM and incubated for >1 hour to alkylate thio-phosphorylated proteins. Reactions were separated on 12% SDS-PAGE gels and phosphorylation of substrate proteins was visualized by western blotting. Total protein levels were detected using Revert total protein stain (LiCOR; 926-11010). Western blots were performed using primary rabbit anti-thiophosphate ester antibody (1:2000; Abcam; ab92570) and mouse anti-FLAG antibody (1:5000; Sigma-Aldrich; F1804) coupled with donkey anti-rabbit 800CW and donkey anti-mouse 680CW antibody (1:12000; LiCOR). Fluorescent signal was detected using Odyssey CLX imager (LiCOR) and quantified with the ImageStudio Software (LiCOR). Data were analyzed using GraphPad Prism software (GraphPad). Kinase activity was determined by dividing phosphorylation signal (anti-thiophosphate ester) by both total RbC (detected via total protein staining) and total kinase (anti-FLAG) levels. The activity of all kinases was normalized relative to CDK6WT, which was set to 1. Kinase activity, as determined from six independent experiments per enzyme, were compared using one-way ANOVA.
Inhibition assays were conducted as described above, except for the addition DMSO or inhibitor dissolved in DMSO at indicated concentrations. The final DMSO concentration was kept at 10%. Kinase activity was determined by dividing phosphorylation signal (anti-thiophosphate ester) by total RbC (detected via total protein staining) only. Data were fit by non-linear regression using the model “[Inhibitor] vs. response (three parameters)” to determine IC50 values. IC50 values from three independent experiments per drug-enzyme pair were compared using one-way ANOVA.
Modeling of CDK complexes and their mutants
We modeled the conformations of the complexes of CDK4 and CDK6 in their cyclin-bound states in the presence of ATP. We constructed the initial coordinates of the active cyclin-D1/CDK4 complex using the crystal structures of cyclin-D3/CDK4 (PDB ID: 7SJ3), replacing the cyclin with cyclin-D1. Cyclin-D1 is modeled using PDB ID 2W9Z, with its C-terminus modeled using homology modeling based on that in cyclin-D3 from 7SJ3. The initial coordinates of cyclin-D1/CDK6 were modeled after the v-cyclin/CDK6 complex (PDB ID: 1XO2) with the cyclin replaced by the same cyclin-D1 as in the CDK4 complex. Both structures are phosphorylated at the active site of the A-loop (pT172 on CDK4, pT177 on CDK6). All missing loops were added using the CHARMM package. We also simulated the active conformations of both CDK4 and CDK6 monomers as well as the unphosphorylated A-loop, from their crystal structures listed above for reference. In all systems, ATP was loaded into both CDKs, replacing the inhibitors in the crystal structures of CDK4 in 7SJ3 and CDK6 in 1XO2 by aligning an ATP-bound crystal structure of CDK2 (PDB code: 1FIN) to both CDKs. The simulated CDK4 and CDK6 systems are listed in Tables S1, S2. The solvent was modeled using the TIP3P model with at least 15 Å shell of water in each direction of the periodic boundary box. The charge states of all titratable groups were set to reflect a neutral pH environment, with acidic side chains represented in their negatively charged forms and basic side chains in their positively charged forms. We added Na+ and Cl- to neutralize the solvated systems and maintain a physiological salt concentration of 150 mM.
MD simulation protocols
We performed MD simulations using methods similar to our previous works, in particular utilizing OpenMM.[34–40] The simulation process consisted of several steps up to the production run. We first performed energy minimization using OpenMM’s minimization integrator for 10,000 steps to optimize the systems and remove unfavorable atomic contacts. For each system, we then executed three 2 μs all-atom explicit solvent MD simulations under the NPT ensemble (constant number of atoms, pressure, and temperature) with 3D periodic boundary conditions. We employed OpenMM version 8.1 for the production runs, utilizing the CHARMM 36m all-atom force field. [41–45] Temperature was maintained at 310 K using the Langevin integrator with a friction coefficient of 1 ps-1. Pressure was kept constant at 1 atm using a Monte Carlo barostat with an update interval of 25 steps. We applied the SETTLE algorithm to constrain bonds involving hydrogen atoms, allowing for a 2-fs integration time step. Long-range electrostatic interactions were computed using the particle mesh Ewald (PME) method with a grid spacing of 0.1 nm. Short-range non-bonded interactions were calculated using a cutoff of 1.2 nm, with a switching function applied to van der Waals interactions starting at 1.0 nm. The simulations were run on CUDA-enabled GPUs to enhance computational efficiency. Trajectory frames were saved every 100 ps for subsequent analysis. We used MDAnalysis, MDTraj and OpenMM’s built-in analysis tools for post-processing and analyzing the simulation trajectories.[46, 47]
Principal component analysis of protein dynamics
To analyze the large-scale motions of the protein during the MD simulation, we performed principal component analysis (PCA) using the Bio3D package in R(v4.4.0).[48, 49] PCA was conducted on the Cartesian coordinates of the Cα atoms to capture dominant motions and reduce the dimensionality of the trajectory. A covariance matrix was constructed from the atomic fluctuations of the Cα atoms around their mean positions. The covariance matrix captures the correlated motions of the atoms during the simulation. Next, eigenvalue decomposition of the covariance matrix was performed using the pca.xyz function to obtain eigenvalues and eigenvectors. The eigenvalues represent the variance explained by each principal component (PC), while the eigenvectors describe the direction of the motion in the conformational space. Then, the first few principal components (PCs) were analyzed in detail. Typically, the first few PCs capture the majority of the protein’s collective motion. The contribution of each PC to the total motion was evaluated by the magnitude of the corresponding eigenvalue, and the cumulative variance was plotted to determine how many PCs are needed to describe the essential motions. To further determine the residues that contribute most to the dominant motions, the contribution of individual residues to each PC (loading plot) was calculated using Bio3D. The per-residue contribution was plotted to identify key regions of the protein involved in large-scale conformational changes.
Community analysis
A network-based community analysis was conducted to identify clusters of residues that move collectively. The dynamic cross-correlation matrix (DCCM) was calculated using Bio3D to quantify correlated and anti-correlated motions between residue pairs. The Cα atom fluctuations from the MD trajectory were used to generate the DCCM. DCCM was then transformed into a distance matrix, and a protein structure network (PSN) was generated where nodes represent residues (Cα atoms), and edges represent significant correlations between them. The edges were weighted based on correlation strength, with a threshold set at 0.5 to filter noise. Communities were detected using the Girvan-Newman algorithm implemented in Bio3D, identifying groups of residues that share similar dynamic properties.
Binding free energy calculations
To quantitatively evaluate the binding affinities and specificities between cyclin-D/CDK4 and cyclin-D/CDK6 complexes, we determined the binding free energies using the molecular mechanics/generalized Born surface area (MM/GBSA) approach, implemented in AmberTools and Amber.[50] This method allows us to assess the interactions between cyclin-D and CDK4 or CDK6 by calculating the binding free energies from MD simulations of the complexes. The total binding free energy (ΔGb) is calculated as the sum of the molecular mechanics gas-phase energy (ΔEMM), solvation free energy (ΔGsol), and entropy contributions (-TΔS), where T is the temperature (in Kelvin): The ensemble averages, denoted by the angle brackets, were calculated over the final 1 μs of equilibrated configurations from the MD simulations. The components of the gas-phase molecular mechanics energy (ΔEMM) include the internal energy (ΔEinter), electrostatic interactions (ΔEelec), and van der Waals interactions (ΔEvdW): The solvation free energy (ΔGsol) consists of the polar and nonpolar contributions, computed using the generalized Born (GB) model in AmberTools with the igb=5 parameter (GBneck2 model) for the polar solvation energy and a surface area-based model for the nonpolar component. The nonpolar contribution is calculated using the linear combination of pairwise overlap (LCPO) model: The entropy contributions (-TΔS) to the binding free energy were estimated using normal mode analysis (NMODE) performed within AmberTools. These include translational (TΔStrans), rotational (TΔSrot), and vibrational (TΔSvib) components. The vibrational contributions were computed from quasi-harmonic approximations based on the trajectory snapshots: The net binding free energy (ΔGb) was determined by calculating the difference between the binding free energy of the bound complex and the sum of the free energies of the unbound CDK and cyclin proteins:
Allosteric pathway analysis
To investigate allosteric pathways in our protein systems, we conducted a dynamical network analysis using the Bio3D package,[48] implementing the weighted implementation of suboptimal paths (WISP).[51] Each amino acid residue was treated as a node, with edges connecting residues whose Cα atoms were within a 5 Å threshold. Edge weights, -ln(cij), where cij is the correlation between residues forming the edge, are inversely proportional to the dynamic correlations calculated from MD simulations, indicating residue coupling strength. We selected source and sink residues based on dynamic cross correlation and their functional relevance, typically involving key binding regions undergoing conformational changes during allosteric modulation. The path length is defined as the sum of edge weights along a path, with longer paths indicating weaker coupling between source and sink residues.
Using WISP, we calculated the top 100 suboptimal paths between source-sink pairs, capturing multiple communication routes and potential alternative pathways for signal propagation. Path lengths and their distributions were analyzed to assess coupling strength across the network. These analyses inform the protein’s allosteric landscape, showing key residues mediating connectivity between functionally important sites.
Quantification and statistical analysis
For statistical robustness, each complex was subjected to three independent MD simulations, with the final binding free energy values representing the average over these runs. The convergence of the simulations was assessed by monitoring the potential energy, electrostatic energy, and the root-mean-square deviation (RMSD) of the Cα atoms, ensuring that equilibrium was reached. To improve statistical accuracy, only the last 1 μs of each trajectory was used for energy calculations. Result analyses were carried out using OpenMM,[44, 45] VMD,[52] PyMOL,[53] MDAnalysis,[46] and custom Python scripts. All free energy results were reported as mean ± standard deviation across the independent simulation replicates.
3Results
CDK6’s longer C-terminus and CDK4’s elongated β3-αC loop define their structural contrasts
A high degree of functional and structural similarity has resulted in CDK4 and CDK6 often being referred to as “CDK4/6”. However, both possess unique evolutionarily conserved structural features (Fig. S2). CDK6 has extended N- and C-termini, each containing an additional α-helix (αB and αJ) (Fig. 1B-D). While the function of these extensions has yet to be investigated, in transcriptional CDKs, C-terminal extensions help to shape the active site.[54] CDK4 has a longer loop connecting its β3-strand to its αC-helix (β3-αC loop), 40PNGGGGGG47 compared to 48QTGEE52 in CDK6. Salt bridge formation between these two structural elements is required to form the ATP binding pocket, thus their relative positions are of critical importance in kinase activation.[28, 55] For example, decreasing the length of the β3-αC loop has been shown to lock BRAF in an active conformation and reduce inhibitor efficacy.[56] By contributing to distinct allosteric networks, these subtle structural variations may impart the distinct regulatory and functional properties of CDK4 and CDK6 in cell cycle control. This work explores the significance of these allosteric structural differences in relation to their activity and functional implications.
Enhanced flexibility in CDK4 loops and C-lobe correlates with reduced activity
Despite only minor differences in sequence and structure between CDK4 and CDK6, biochemical assays reveal that wild-type CDK6 exhibits more than 10-fold greater kinase activity toward the Rb C-terminus (RbC) when compared to wild-type CDK4 (Fig. 2A and S3). To understand the structural origins of this difference, we performed principal component analysis (PCA) on the MD trajectories of cyclin-D/CDK4 and cyclin-D/CDK6. We denote these two systems CDK4WT and CDK6WT. Analysis shows vastly different conformational clustering patterns for each kinase, with conformational ensembles depicted by PCA revealing distinct local fluctuation patterns for CDK4WT and CDK6WT in the first principal component in particular (Fig. 2B). The first two principal components explain a larger portion of the variance in CDK4WT (39%) than in CDK6WT (34%), suggesting greater conformational flexibility in CDK4WT (Fig. 2C). Residues in the β3-αC and β4-β5 loops display high degrees of conformational variability, as indicated by their contribution to PC1 (Fig. 2D). Conformational ensembles play a crucial role in determining function. [57, 58] Despite their sequence homology, CDK4 and CDK6 occupy different conformational spaces, likely reflecting their non-redundant functional roles in cell cycle progression.[16, 54, 59] This is consistent with findings that CDK6 has a more stable αC-helix conformation compared to CDK4, potentially explaining its higher basal activity.[60] The dynamic cross-correlation map for all residue pairs shows distinct patterns of intramolecular communication within CDK4WT and CDK6WT (Fig. 2E). CDK4WT shows more extensive anti-correlations between the N- and C-lobes. Excessive anti-correlation indicates less stable structure leading to lower efficiency compared to other cell cycle CDKs.[29]
Longer β3-αC loop in CDK4 disrupts coordinated kinase dynamics and activity
CDK4WT has a longer and more flexible loop connecting the β3-strand to the αC-helix relative to CDK6WT (Fig. 3A and 3B and Fig. S4). During the transition from the inactive to the active state, the longer loops in CDK4 block the inward movement of the αC-helix upon cyclin binding, contributing to its slower activation compared to CDK2.[28] However, once CDK4 crosses the activation barrier and becomes the active complex (as in the state studied here), the β3-αC loop also plays a role in modulating the dynamics of the complex by decoupling the correlated motion of the αC-helix and the G-loop, hence decreasing the activity. The binding free energy indicates a stronger affinity between cyclin-D and CDK6WT compared to CDK4WT (Fig. 3C) due to more extensive contacts with cyclin-D (Fig. S5), resulting in the stronger coupling. This stronger association may contribute to the higher catalytic efficiency of CDK6, analogous to highly efficient cyclin-E and CDK2 having a greater binding affinity than cyclin-D and CDK4 in our previous study.[29] The flexible loop decreases binding affinity since flexibility increases the entropic contribution to the binding free energy, which decreases the binding affinity (Fig. S6).[61]
We next used community analysis to understand how the longer β3-αC loop in CDK4 affects the coordinated rearrangement of structural components essential for kinase activation. Community analysis, which segments proteins into smaller, interconnected groups of residues that move together as semi-rigid bodies, offers valuable insights into the cooperative dynamics of protein regions and the propagation of conformational changes throughout the structure. [14, 62–64] Structural analysis of the cyclin-D/CDK6WT complex places the G-loop and αC-helix in the same structural community, indicating coordinated movement between both structures (Fig. 3E). These structural elements play critical roles in facilitating catalysis, with the G-loop capping the active site and αC-helix ensuring that the kinase remains in an active state through its interactions with the β3-strand. Coordinated motion between the G-loop and αC-helix is associated with enhanced kinase activity, as seen in eukaryotic protein kinases.[65] Cyclin-D/CDK4WT lacks this coordination, as community analysis splits the G-loop and the αC-helix into different structural communities (Fig. 3D). This is at least partially due to the longer β3-αC loop in CDK4, as swapping the β3-αC loop in CDK6 for that of CDK4 (CDK6β3-αC(CDK4)) reduces catalytic activity by 50% and results in communities like those observed for CDK4WT (Fig. 3F and S7). Thus, this difference in dynamic organization may explain the higher basal activity observed in CDK6.
Finally, we assessed how the structural differences between CDK4 and CDK6 affects β3-strand αC-helix contacts. In active cyclin-CDK complexes, the β3-strand lysine (K35 in CDK4, K43 in CDK6) forms a salt bridge with the αC-helix aspartate (E56 in CDK4, E61 in CDK6) (Fig. 3G), creating a link to the G-loop, thereby stabilizing both elements. Our simulations reveal that CDK6WT has a far higher probability of salt bridge formation than CDK4WT, suggesting a more efficient catalysis-ready conformation. Mutations affecting this residue disrupt the relative positioning of the G-loop and αC-helix, altering the active site and reducing catalytic efficiency.[65, 66] Given the importance of active site precision for efficient catalysis[66], coordinated motion between the G-loop and the αC-helix likely optimizes this conformation (Fig. 3H), leading to enhanced kinase activity.
Stronger allosteric coupling in CDK6 stabilizes active site for efficient catalysis
We next used network analysis to define the mechanism of allosteric regulation between the β3-αC loop and the G-loop. This network-based approach allows us to understand how signals travel through the kinase and how different structural elements work together dynamically. The strength of the connection between two residues is based on how correlated their motions are: if they tend to move together, they form a path and are strongly connected. We focused on the allosteric signaling pathway linking the β3-αC loop (source) to the G-loop (sink) with source and sink residues selected based on the dynamic cross-correlation map (Fig. 2E). The selected source residues are G46 of CDK4 and E51 of CDK6 in the β3-αC loop and the sink residues are A16 of CDK4 and A23 of CDK6 in the G-loop.
The allosteric signaling path in CDK4WT is long and meandering, starting from its extended β3-αC loop and progressing through the β3-strand, β2-strand, and β1-strand before reaching to the G-loop (Fig. 4A and C). In contrast, the CDK6WT allosteric pathway is more direct, with the allosteric signal only passing through a small portion of the β3- and β2-strands without influencing the β1-strand (Fig. 4B and E). To investigate how strongly the residues along these paths are connected, we calculated the length of the allosteric path. The path length is the sum of edge weights in a network path connecting two sites, where each edge weight is typically calculated as -ln(|cij|) with cij being the correlation between connected residues i and j, rather than the physical distance between the source and sink residues, or the number of residues between the source and sink along a path. The path length distribution in CDK4WT (peak at ∼3.44) indicates slightly longer paths with weaker coupling between the source and sink residues compared to CDK6WT (peak at ∼3.1) (Fig. 4D and F). The shorter allosteric path length in CDK6WT indicates stronger correlated motions between the β3-αC loop and the G-loop, potentially positively modulating the kinase activity. The different allosteric pathways can be attributed to the different length of the β3-αC loop. CDK4WT has a longer β3-αC loop than that of CDK6WT (Fig. 1D). Loop swapping of CDK6β3-αC(CDK4) by adopting the β3-αC loop of CDK4 results in longer allosteric paths, resembling those observed for CDK4WT (Fig. S7).
The G-loop connecting the β1- and β2-strands contains a conserved GxGxxG motif and is located on top of ATP, which is important for ATP positioning. Superimposition of the conformations from the simulated trajectories shows that the G-loop of CDK4WT has greater flexibility (Fig. 4G) compared to that of CDK6WT (Fig. 4H), as indicated by the residual RMSFs (Fig. S9). These observations could provide an explanation for the different responses of CDK4 and CDK6 to certain inhibitors, such as the CDK4-selective inhibitor atirmociclib (PF-07220060), which exhibits a remarkably high selectivity for CDK4 over CDK6. The enhanced flexibility of the CDK4 G-loop in part allows for more efficient binding of the bulky isopropyl alcohol [(CH3)2COH] group in this inhibitor (Fig. S10).[67] Indeed, replacing the β3-αC loop in CDK6 with that of CDK4 (CDK6β3-αC(CDK4)) significantly increases its sensitivity to this CDK4-selective inhibitor (Fig. 4I and S11).
CDK6 C-terminus stabilizes the R-spine, enhancing catalytic activity
Relative to CDK4, CDK6 has an extended C-terminal tail, a structural feature known to regulate the kinase activity. For example, the C-terminal tail of PKA supports its active conformation.[30, 31] Some CDKs are known to have flexible C-terminal tail that helps to stabilize the N-lobe by folding onto it, enclosing the ATP in its active site.[54] However, it is not known how the C-terminal tail of CDK6 affects its conformational dynamics. To elucidate the role of the C-terminus, we simulated CDK6 with a truncated C-terminal tail (CDK6ΔCterm) and calculated the cross-correlation of atomic fluctuations. The dynamic cross-correlation map reveals highly correlated residue motions between the N- and C-lobes of CDK6WT (Fig. 5A and C). In contrast, these residue motions are uncorrelated in CDK6ΔCterm (Fig. 5B). Instead, correlation is primarily observed within the C-lobe (Fig. 5D). We hypothesize that in CDK6ΔCterm, the absence of the correlated residue pairs between the N- and C-lobes renders the inter-lobe motions rigid, hindering the opening and closing of the active site and resulting in lower activity. Consistent with this, biochemical assays with CDK6ΔCterm show reduced kinase activity compared to CDK6WT, resulting in a lower rate of RbC phosphorylation (Fig. 5E). One explanation for the reduced activity of CDK6ΔCterm is that the C-terminal deletion disrupts the coordinated movements of the regulatory-spine (R-spine) of CDK6 (Fig. 5F). Once the R-spine is disrupted, the kinase is less likely to maintain its stability and resulting in reduced activity.[68] In CDK6WT, we observed multiple strong and short allosteric signaling paths connecting the C-terminal edge of the kinase domain to an R-spine residue (Fig. 5G), but this is not the case in CDK6ΔCterm (Fig. 5H). CDK6ΔCterm disrupts the intrinsic dynamics of the kinase, resulting in longer allosteric paths from the C-terminus through both αH- and αF-helices to the R-spine. This is in part due to the high mobility of the truncated C-terminal tail (Fig. S12), which increases the dynamics of the C-lobe, breaking the intrinsic coordinated motions within the kinase that separate R-spine into three distinct communities (Fig. 5F). Our finding represents a novel mechanism of kinase activity in which the C-terminal tail plays a direct role in modulating the opening and closing of the C- and N-lobes through R-spine dynamics.
4Discussion
Though largely identical, CDK4 and CDK6 evolved distinct activities in regulating the G1 progression. This distinction is particularly relevant to CDK4/6 inhibitor-resistant cancer cells, as well as in adult tissues and HSCs. The key question is how the overexpressed CDK6 in cancer cells enables cell cycle progression in lieu of CDK4, and what molecular mechanisms differentiate the higher activity of CDK6 from that of CDK4. Explaining these differences at the conformational level could clarify the shifts towards the CDK6 complex for cell cycle progression in resistant cancer cells, thereby bypassing CDK4/6 inhibition.
Our experimental results indicate that CDK6WT is more active than CDK4WT and CDK6ΔCterm. The latter, which lacks the C-terminal part of CDK6 is more similar to CDK4 in terms of activity. We use MD simulations to explain the molecular origin of their catalytic differences. The higher activity of CDK6WT can be attributed to several factors, including the allosteric role of the β3-αC loop upon cyclin binding. A longer β3-αC loop in CDK4 results in a disjointed kinase, while a shorter loop in CDK6 is better at facilitating coordinated motion for efficient catalysis. Furthermore, the weaker coupling between the G-loop and the β3-αC loop in CDK4WT allows decoupling of the β-stands (β1-β5) and the αC-helix. This results in insufficient stabilization of ATP and, consequently, lower activity due to the longer β3-αC loop of CDK4WT. In contrast, CDK6WT has a shorter β3-αC loop, resulting in stronger coupling between the G-loop and the β3-αC loop. This enhanced coupling leads to better stabilization of ATP and higher catalytic activity. Secondly, the unstructured C-terminus of CDK6WT acts on the kinase allosterically, stabilizing the R-spine for more efficient catalysis. In contrast, CDK4WT lacks this C-terminus, leading to a less stable R-spine, as shown in our previous works.[29] Thirdly, the dynamics of CDK4WT reveal a significant degree of opening and closing of the N- and C-lobes, leading to an unstable structure as evidenced by the unstable R-spine.[29] This instability likely contributes to the slower catalytic activity observed for CDK4, as the excessive conformational changes typically hinder efficient substrate binding and phosphoryl transfer (Fig. 6). In contrast, as discussed in our previous work,[29] CDK2WT displayed a relatively stable structure with less dynamic opening and closing motions of the N- and C-lobes for ATP binding, resulting in higher catalytic activity. We suspect that CDK6WT may exhibit the right balance of N- and C-lobe movements between CDK2 and CDK4, and a coordinated and efficient catalytic process. The ability of CDK6WT to maintain a dynamic yet stable conformation is likely responsible for its higher activity compared to both CDK4WT and CDK6ΔCterm.
These findings suggest that achieving optimal kinase function requires the right balance of conformational plasticity and show that the evolution of CDK6, as exemplified by its role in fast-proliferating short-term HSCs, has favored allosteric networks that enhance its activity relative to CDK4. In CDK4/6i-resistant cancer cells, CDK6 is often overexpressed. When complexed with CIP/KIP family inhibitors, such as p18 or p27, CDK6 becomes insensitive to the inhibitors, in part due to the allosteric mechanisms it has acquired through evolution.[22, 60] These mechanisms include the longer C-terminus, which stabilizes the R-spine, and stronger coupling between the β3-αC loop and the αC-helix upon cyclin binding.
The divergence in sequence, function, and activity of CDK4 and CDK6 is not unique in evolution, and our results show that such differences provide opportunities for designing inhibitors that selectively target one kinase but not the other. Notably, FGFR1 and FGFR2 follow a similar pattern.[69, 70] Despite their high sequence similarity, these kinase pairs exhibit different catalytic efficiencies and regulatory mechanisms. The different dynamics of FGFR1 and FGFR2 have been leveraged in the design of the inhibitor lirafugratinib to selectively target FGFR2.[70] The G-loop (or P-loop) in FGFR1 is highly flexible, rapidly transitioning between an extended conformation and a more contracted state. In contrast, the G-loop in FGFR2 maintains a relatively extended conformation, which is far less dynamic than that of FGFR1, allowing efficient covalent binding of lirafugratinib to C491 in FGFR2. Although lirafugratinib stabilizes the G-loop conformation in FGFR1, the covalent bond is too far away from C488 for efficient covalent binding. These examples show the importance of considering protein dynamics in drug design.
Closely related kinases, such as CDK4 and CDK6, often have different allosteric mechanisms that can potentially be exploited to selectively target one over the other. For example, AKT1 is more biochemically active and responsible for the tumor initiation, while AKT2 is mainly responsible for tumor progression and metastasis.[71] Their divergence is in part allosterically regulated by their differences in the PH domain.[72]
Here, we discover two novel allosteric mechanisms that CDK4 and CDK6 utilize to differentially regulate their activity: allosteric signaling between the β3-αC loop and the G-loop, “sensing” cyclin binding, and the C-terminus of CDK6, modulating its activity through coordinated R-spine movement. The weak allosteric coupling of β3-αC loop and G-loop in the CDK4 offers a more flexible G-loop compared to CDK6. These allosteric mechanisms may also in part explain why overexpression of CDK6, rather than CDK4, leads to drug resistance. This is because CDK6 may be more efficient at promoting cell cycle progression even when partially inhibited, and the relatively inflexible G-loop of CDK6 is less likely to allow binding of bigger ligands such as inhibitors, while simultaneously allowing binding and unbinding of ATP, facilitating ATP hydrolysis.[22, 73] To overcome CDK6 overexpression in drug resistance to CDK4/6i, proteolysis-targeting chimeras (PROTACs) and molecular glue degraders have been proposed.[59, 73] This strategy is highly promising as it potentially eliminates the overexpressed CDK6 through its targeted degradation, rather than merely inhibiting its activity.
5Conclusions
Our combined approach, leveraging molecular simulations and biochemical assays, uncovers new molecular mechanisms underlying the allosteric differences between the closely related CDK4 and CDK6.Two allosteric networks are identified: one in the N-lobe (connecting the β3-αC loop and G-loop) and another in the C-lobe (linking the C-terminus and R-spine). In CDK6 these networks have shorter paths or stronger coupling than those in CDK4, providing favorable conditions for higher kinase activity. These findings reveal how these kinases evolved to fulfill distinct functional roles in different cell types and offer new mechanistic details into their allosteric regulation and inhibitor specificity.
Supporting information
Acknowledgments
We thank members of the Kõivomägi and Nussinov lab for the discussions and comments on this work. This project was supported in whole or in part by federal funds from the National Cancer Institute, National Institutes of Health, under contract HHSN261201500003I and ZIABC012133. The contents of this publication do not necessarily reflect the views or policies of the Department of Health and Human Services, nor does mention of trade names, commercial products, or organizations imply endorsement by the U.S. government. This research was supported [in part] by the Intramural Research Program of the NIH, National Cancer Institute, Center for Cancer Research. All simulations had been performed using the high-performance computational facilities of the Biowulf cluster at the National Institutes of Health, Bethesda, MD (https://hpc.nih.gov/).
Conflict of interest
The author declares that there is no conflict of interest that could be perceived as prejudicing the impartiality of the research reported.