Disease mutations and phosphorylation alter the allosteric pathways involved in autoinhibition of protein phosphatase 2A

Abstract

Mutations in protein phosphatase 2A (PP2A) are connected to intellectual disability and cancer. It has been hypothesized that these mutations might disrupt the autoinhibition and phosphorylation-induced activation of PP2A. Since they are located far from both the active and substrate binding sites, it is unclear how they exert their effect. We performed allosteric pathway analysis based on molecular dynamics simulations and combined it with biochemical experiments to investigate the autoinhibition of PP2A. In the wild type (WT), the C-arm of the regulatory subunit B56δ obstructs the active and substrate binding sites exerting a dual autoinhibition effect. We find that the disease mutant, E198K, severely weakens the allosteric pathways that stabilize the C-arm in the WT. Instead, the strongest allosteric pathways in E198K take a different route that promotes exposure of the substrate binding site. To facilitate the allosteric pathway analysis, we introduce a path clustering algorithm for lumping pathways into channels. We reveal remarkable similarities between the allosteric channels of E198K and those in phosphorylation-activated WT, suggesting that the autoinhibition can be alleviated through a conserved mechanism. In contrast, we find that another disease mutant, E200K, which is in spatial proximity of E198, does not repartition the allosteric pathways leading to the substrate binding site; however, it may still induce exposure of the active site. This finding agrees with our biochemical data, allowing us to predict the activity of PP2A with the phosphorylated B56δ and provide insight into how disease mutations in spatial proximity alter the enzymatic activity in surprisingly different mechanisms.

I. INTRODUCTION

Protein phosphatase 2A (PP2A) is a major serine/threonine phosphatase that is involved in cell cycle progression, growth, cell death and survival, DNA damage response, and several diverse signaling pathways.^1–3 The holoenzymes contain common scaffold A and catalytic C subunits as well as a variable regulatory B subunit. One type of regulatory B subunit, B′/B56 regulatory subunit, recognizes substrates by binding to short linear motifs (SLiMs) of the substrates. This regulatory subunit also contains a SLiM that binds to the SLiM binding groove, which can auto-inhibit the holoenzyme.^4–7 The deregulation of PP2A leads to human diseases, including multiple types of cancer and neurological disorders.^1,8–10 Mutations to the B56δ regulatory subunit (also known as PPP2R5D) have been identified in cancer and neurological disorders, including intellectual disability.^11–16 A recently solved cryo-electron microscope (EM) structure of the PP2A-B56δ holoenzyme reveals that the disordered N/C-arms of B56δ form an unstructured interface spanning over 190 Å, which connects to the catalytic and the scaffold subunits [Fig. 1(a)].¹⁷ In the closed latent form, the folded C-arm of B56δ blocks access to the active site on the catalytic subunit. The top inlay in Fig. 1(a) shows that a substrate bound in the active site would clash with the C-arm. The tail of the C-arm occupies the SLiM-binding groove and must detach to allow SLiM recognition [bottom inlay in Fig. 1(a)]. Together, the B56δ has a dual autoinhibition effect on the enzyme. Both N- and C-arms of B56δ can be phosphorylated at multiple sites, which activate the enzyme, likely by loosening the N/C-arm interface and revealing the active site and SLiM-binding grove. Multiple mutations associated with intellectual disability are scattered throughout this N/C-arm interface over the active site and SLiM-binding grove, but none of them are directly a part of the SLiM binding groove or the active site. Therefore, it is hypothesized that they have an allosteric effect on autoinhibition, but the mechanism by which these mutations disrupt normal autoinhibition is unknown.

FIG. 1.

C-terminal SLiM loosening in MD simulations. (a) Overview of the PP2A holoenzyme structure with the B56δ subunit (white). The catalytic subunit is colored green, and the scaffold subunit is colored blue. The N-arm and C-arms of the B56δ are colored yellow and magenta, respectively. The serine residues phosphorylated in WT4P are labeled. The top inlay shows that a bound substrate (from PDBID: 2NPP) is incompatible with the closed C-arm due to steric clashes. The magnitude of the C-arm SLiM fluctuation is depicted in the bottom inlay. (b) Root mean square fluctuation of the terminal 20 C-arm residues. The SLiM residues are underlined. The error bars correspond to the standard deviation of the mean calculated by bootstrapping the trajectories 30 times.

Allostery is a mechanism that couples two distal sites in a biological macromolecule; therefore, a perturbation at the far-away (allosteric) site can propagate to the functional (or orthosteric) site. Allosteric effects play an important role in signal transduction, transcriptional regulation, and metabolism.¹⁸ Drugs that target allosteric mechanisms hold great potential to overcome the limitations of traditional orthosteric compounds, greatly expanding the possible repertoire of druggable targets.^19,20 Both experimental and computational²¹ approaches proved valuable to uncover the allosteric mechanisms. From the experimental side, allostery has been successfully studied with site-directed mutagenesis,²² nuclear magnetic resonance (NMR),^23,24 and single-molecule spectroscopy,²⁵ among others. Computational models can further expand our understanding of these complex mechanisms with approaches such as perturbation response scanning,²⁶ topology analysis,²⁷ normal mode analysis,²⁸ and bioinformatic analyses.²⁹ Mutations in B56δ, including E198K, E200K, and, to a lesser degree, E420K and E197K, were previously suggested to be involved in allosteric signaling and were shown to alter the activity of PP2A.¹⁷ Our study will focus on E198K and E200K as they are the two most abundant variants, represent the most severe and a mild version of the disease, and produce starkly different effects on the SLiM binding.

Molecular dynamics (MD) simulations can provide valuable insights into the allostery of proteins with rich mechanistic and energetic details.^30–45 To study this phenomenon, an allosteric network can be constructed based on the correlated dynamics in the system.^30,33,34 Paths through such a network can elucidate the allosteric wiring in the protein and the residues that are most important for this process.^18,25–27 For example, Romero-Rivera et al. found that the residues most often crossed by the shortest paths in the allosteric network match with ones that were mutated in a retro-aldolase engineered through directed evolution, indicating that these residues were important for function.⁴⁶ When one is interested in the allosteric effect between specific sets of residues, one can focus only on the paths that connect them. These source and sink regions can be selected broadly, taking multiple residues as the start and end points of the paths.^34,41,47 However, residues in spatial proximity may play drastically different roles in allosteric signaling, for example, E198 and E200 in PP2A-B56δ;¹⁷ therefore, contributions of individual residues in the source and sink must be examined carefully, considering all pairwise combinations between source and sink regions containing multiple residues give rise to an ensemble of paths, which may be difficult to interpret. To address this problem, Bhattacharya and Vaidehi combined the pathways into a small number of “pipelines” based on the proximity of residues in the protein structure.³¹ While this approach facilitates interpretation, it does not consider the dynamic correlation; it can only cluster pathways that are close in the structure, neglecting the fact that they can account for functionally different correlated motions. Therefore, additional methodological development is required to attribute the functional role of such allosteric pathway clusters.

Here, we used MD simulations and allosteric network analysis to study the impact of the two most severe mutations in B56δ implicated in intellectual disability (E198K and E200K) as well as a variant harboring phosphoserines in positions 88, 89, and 90 on the N-arm and 573 on the C-arm, referred to as WT4P. We analyzed the allosteric pathways between the mutation sites as well as the pathways between the mutation sites and the C-arm SLiM. We found that these pathways utilize distinct channels sharing many residues within a channel in different mutations. To overcome the problem of identifying and analyzing allosteric channels, we devise a new approach to group individual paths into path channels based on their inter-correlation. We find that mutations and phosphorylation of B56δ, although utilizing neighboring residues, have different destabilizing effects on the C-arm SLiM, which likely free the SLiM binding groove for incoming substrates. Additionally, we show that these mutations and phosphorylation to wild type (WT) B56δ weaken the connection between the C-arm interface and the catalytic subunit, rendering the active site more accessible. Remarkably, WT4P displays similar effects on the active site and the C-arm SLiM as the E198K mutant, leading us to predict that phosphorylation of B56δ increases PP2A activity by allosterically enhancing substrate-SLiM binding and access to the active site.

II. METHODS

A. Structure modeling and MD simulations

The starting structure for MD simulation was based on the available cryo-EM structure of PP2A-B56δ in the closed form.¹⁷ Missing residues 513–568 of B56 δ of the structure were taken from a predicted structure (UniProt identifier: AF-Q14738-F1).⁴⁸ Mutations to residues E198 and E200 were introduced with PyMOL 2.4.1.⁴⁹ Phosphate groups in WT4P (S88, S89, S90, and S573 of B56δ) were added manually. The protonation state of residues was determined by PROPKA 3.4.0.^50,51 MD simulations were performed with the Amber ff99SB-ILDN forcefield⁵² using GROMACS 2022.2.⁵³ The structure was simulated in a dodecahedron box with the initial size chosen such that the closest protein residue would be 1 nm from the box boundaries. The complex was solvated in TIP3P water,⁵⁴ and water molecules were replaced with sodium ions until the system was neutral. The coulomb interactions were cut off at 1 nm, and long-range electrostatics was computed with particle mesh Ewald.⁵⁵ The Lennard-Jones interactions were cut off at 1 nm. The energy of the system was minimized with the steepest descent for 10 000 steps. Then, 1 ns MD simulations were performed at 300 K with the temperature being controlled by a V-rescale thermostat⁵⁶ and pressure set to 1 bar controlled by a C-rescale barostat.⁵⁷ The production MD simulations were carried out in 10 replicas for 120 ns with initial velocities assigned randomly to the 300 K distribution. Please refer to Table I for a summary of the size of the simulation box for various systems and simulation lengths. The first 20 ns of each MD trajectory was removed before analysis, and snapshots used for analysis were spaced by 100 ps.

TABLE I.

Summary of the MD simulations for various systems.

B56δ variant	Number of atoms in the simulation box	MD simulations (ns)
WT	204 396	10 × 120
E198K	204 404	10 × 120
E200K	204 425	10 × 120
WT4P	212 906	10 × 120

B. Analysis of MD trajectories

The number of contacts and solvent-accessible surface area (SASA) were computed with the “MDTraj” Python library.⁵⁸ To compute the root mean square fluctuation (RMSF), snapshots were aligned to the average structure in each variant. Images of structures and pathways were produced with PyMOL.⁴⁹ Figures S2–S4 were produced with the “pyvis” Python library.⁵⁹

C. Allosteric network analysis

We followed Luthey-Schulten and colleagues to construct allosteric networks of each PP2A variant.³³ In particular, we chose two atoms to represent each protein residue: the Cα carbon and the most distant non-hydrogen atom on the sidechain (for Gly, Pro, and Ala, only the Cα was used). Then, the generalized linear correlation matrix was computed with “g_correlation.”⁶⁰ This matrix is based on the pairwise generalized linear mutual information (LMI) I_ij between the selected atoms,

$$I_{ij}=H_i+H_j-H_{ij}.$$ (1)

Here, H is the Shannon entropy defined as

$$H_i=-\int p\left(x_i\right)\log [p\left(x_i\right)]d{x}_j,$$ (2)

$$H_{ij}=-\int \int p\left(x_i,x_j\right)\log [p\left(x_i,x_j\right)]d{x}_{i}d{x}_j,$$ (3)

where x_i and x_j are the deviations of the atomic position from the mean and $p\left(x_i\right)$ and $p\left(x_j\right)$ are their marginal probability distributions and $p\left(x_i,x_j\right)$ is their joint probability distribution. Elements of the generalized linear correlation matrix are then defined as

$$C_{ij}={\left(1-e^{-\left(2/d\right)I_{ij}}\right)}^{-1/2},$$ (4)

where d is the dimensionality of x_i and x_j. The generalized linear correlations were multiplied with a contact map so that only the correlation between neighboring residues would be considered. The contact map K between each pair of selected atoms utilized a linear smoothing function as suggested by Botello-Smith and Luo³² and was computed in the following way:

$$K(r)=\left\{\begin{array}{c}1ifr\le r_{\mathit{full}},\hfill \\ \frac{e^{-\frac{r^2}{2{\mathrm{\sigma }}^2}}}{e^{-\frac{r_{\mathit{full}}^2}{2{\mathrm{\sigma }}^2}}}ifr>r_{\mathit{full}}.\hfill \end{array}\right.$$ (5)

Using two atoms per residue, the contact map was based on the distance between the selected atoms r in an MD snapshot rather than the conventional distance between the closest heavy atoms in a pair of residues. The cutoff distance for a full contact was selected as r_full = 0.8 nm and r_cut = 1.5 nm; therefore, $K\left(r_{\mathit{cut}}=1.5\right)=10^{-5}$ and $\mathrm{\sigma }=\sqrt{\frac{r_{\mathit{full}}^2-r_{\mathit{cut}}^2}{2\ln K\left(r_{\mathit{cut}}\right)}}\approx 0.264$. Then, $m_{ij}=\frac{1}{N_{\mathit{frames}}}{\sum }_{n=1}^{N_{\mathit{frames}}}K\left(r_{ij}\left(n\right)\right)$ are the elements of the contact map averaged over all frames in the simulation.

The generalized linear correlation matrices were converted to graphs, where each vertex represented a selected atom and the edges connected neighboring atoms. The edge weights in the network were computed as $w_{ij}=-\log \left|L_{ij}\right|$, where L_ij = m_ij C_ij. Paths in this network weighed by the edges explain how strongly the correlated motion of residues in the source affects residues in the sink. Although many paths exist in an allosteric network, the shortest path reflects the strongest allosteric connection between the source and the sink. To obtain the pathways between mutation sites (E/K198 and E/K200) and either the C-arm SLiM (L595, S598 and E600) or the active site (D57, D86) shortest paths in the network were computed with the Dijkstra algorithm⁶¹ implemented in the “NetworkX” Python library.⁶² With two atoms representing each residue, all the pairwise combinations between atoms in the source and sink residues make a total of 24 paths between the mutation sites and the C-arm SLiM and 16 paths between the mutation site and the active site.

Suboptimal paths can have a non-negligible contribution to the allosteric signal, which may not be captured in the shortest paths. They can be derived from a given shortest path by considering one vertex deviation from the original^63,64 or sampled via the Monte Carlo scheme.⁴¹ To obtain the suboptimal pathways, a recently developed algorithm “subset of adjacent nodes” (SOAN) was employed.⁶⁴ The SOAN algorithm first evaluates all the neighboring nodes of the optimal path. For each neighbor, it finds the shortest path to the source and to the sink using Dijkstra’s algorithm. Those paths are then combined at the neighbor junction, and the unique nodes from this initial set of paths are used to make a subgraph, which retains the connectivity of the original graph. This subgraph can be easily searched for k shortest paths using Yen’s algorithm,⁶⁵ which are the resulting suboptimal paths. Using this algorithm, 50 additional pathways were computed for each of the shortest paths, yielding 1200 paths and 800 paths to the C-arm SLiM and the active site, respectively. The weight of each path was computed as the inverse sum of all edge weights traversed by the path; therefore, a larger path weight corresponds to a stronger allosteric connection between the source and the sink. To compare path weights across mutants, the total weight of optimal and all suboptimal paths was normalized by the total weight of paths in the WT for a given pair of source and sink residues,

$$\sum _{k=1}^P\frac{1}{\sum _{ij}^{N_k}{w}_{ij}},$$ (6)

where P is the total number of optimal and suboptimal paths and N_k is the number of edges crossed by the path k.

D. Allosteric path lumping

To aid the interpretation and quantify the effect of multiple similar allosteric pathways, we develop a new approach to lump the allosteric paths into a small number of path channels in accordance with their inter-correlation. Using the LMI matrices with contact map cutoff, the inter-correlation between two paths M and N can be written as

$$S_{MN}=\frac{1}{2}\frac{1}{mn}\sum _{i\in M}\sum _{j\in N}{L}_{ij}{L}_{ji},$$ (7)

where m and n are the number of sites that paths M and N pass through, respectively, and L_ij is the masked LMI between atoms i and j. To compare the paths between PP2A variants (WT, E198K, E200K, or WT4P), the average of the LMI matrices between the variants masked with their respective contact maps was used to compute S_MN. Using S_MN, the pairwise inter-correlation matrix of all allosteric paths was computed. Then, the spectral clustering algorithm⁶⁶ was adopted to identify the path channels based on the eigenstructure of the similarity matrix. The number of path channels is chosen according to the largest gap in the eigenvalue spectrum of the Laplacian matrix of the inter-correlation matrix. Thus, the paths that were highly correlated with each other were grouped into the same channel. The weights of allosteric channels were computed by summing all paths belonging to a channel and normalizing them by the total weight in the WT variant, thereby reflecting the overall strength of the allosteric signal through a channel.

E. Protein expression and purification

Baculovirus of PP2A subunits were prepared according to manufacturer’s instructions in Spodoptera frugiperda 9 (Sf9) cells using a Bac-to-Bac Baculovirus Expression System (Invitrogen). PP2A-B56δ holoenzymes were expressed in Hi-5 suspension cells (Thermo Fisher Scientific). Hi-5 cells grown to a density of 1.5 × 10⁶ cell/ml were co-infected with PP2A His-Aα, His-Cα, and GST-B56δ baculovirus for 48 h at 27 °C. Cells were harvested by centrifugation and lysed by Dounce homogenization in the lysis buffer containing 25 mM Tris-HCl (pH 8.0), 150 mM NaCl, 50 µM MnCl₂, 2 mM DTT, and protease inhibitors [10 µM leupeptin, 0.5 µM aprotinin, and 1 mM phenylmethanesulfonyl fluoride (PMSF)]. The insoluble proteins were removed by centrifugation, and the soluble fraction of cell lysates was gravity-loaded to GS4B (Glutathione Sepharose 4B) resin (Cytiva) column for three times followed by two washes using five column volumes (CV) of lysis buffer. The proteins left on the resins were digested by the Tobacco Etch Virus (TEV) protease on the column at 4 °C until the His- and GST-tags were completely removed, and the flow-through was further fractionated by anion exchange chromatography (Source 15Q column, Cytiva) and gel filtration chromatography (Superdex 200 column, Cytiva).

F. GST-mediated pulldown assay

To test the interaction between WT and mutant PP2A-B56δ holoenzyme with substrate B56 SLiMs, 12 µg of GST-tagged SYT16 (132-147) or GST-tagged cAMP response element-binding protein (CREB) (99-161) was immobilized on 5 µl of GS4B resin. The unbound protein was washed out by assay buffer containing 25 mM Tris (pH 8.0), 150 mM NaCl, 3 mM DTT, and 1x protease inhibitor cocktail (P8340, Sigma). Then, 10 µM of PP2A-B’δ holoenzymes was then added to the immobilized GST-tagged protein in a final volume of 50 µl assay buffer supplemented with 1 mg/ml of bovine serum albumin (BSA) or 5 mg/ml lysosome. After 5 min of incubation, the unbound proteins were removed, and the resins were washed three times by the assay buffer supplemented with 0.1% of Triton X-100 three times. The proteins that remained on the resin were examined by sodium dodecyl sulfate (SDS)-PAGE and visualized by Coomassie blue staining.

G. In vitro phosphatase assay

The enzyme kinetics of the purified PP2A-B56δ holoenzymes were determined using the PiColorLock phosphatase assay kit (Abcam, Ab270004). A 10 µl sample of phosphorylated hexapeptide (KRpTIRR) peptide prepared at six times the target concentrations was added in a 96-well clear plate. Then, 50 μl of 30 nM indicated PP2A-B56δ holoenzymes in the assay buffer, containing 25 mM Tris pH8.0, 150 mM NaCl, and 50 μM MnCl₂ supplemented with 0.05 mg/ml BSA, was added to the wells to start the reactions. After 3 min, the reactions were quenched by adding 15 µl of the quench buffer provided in the kit, and the color was allowed to develop for 15 min.

The absorbance at 635 nm of the reactions was then measured using SpectraMax Plus 384 Microplate Reader (Molecular Devices) in end-point mode. Initial velocities (V₀) were determined at varying concentrations of the substrate and fit to the Michaelis–Menten equation [Eq. (8)] to determine the steady-state kinetics of the PP2A-B’δ holoenzymes,

$$v=\frac{k_{\mathit{cat}}\left[E\right]\left[S\right]}{K_m+\left[S\right]}.$$ (8)

In Eq. (8), k_cat is the rate constant, $\left[E\right]$ and $\left[S\right]$ are the enzyme and substrate concentrations, respectively, and K_m is the Michaelis–Menten constant reflecting the binding affinity between the peptide substrate and the enzyme.

H. In vitro holoenzyme phosphorylation by protein kinase A (PKA)

A 100 µl sample of PP2A-B56δ holoenzymes was prepared to a final concentration at 0.1 mg/ml in assay buffer containing 50 mM Tris-HCl (pH7.5), 10 mM Mg₂Cl, 200 μM ATP, 0.1 mM EDTA, 2 mM DDT, and 0.01% Brij35. Phosphorylation was initiated by the addition of 1 µl protein kinase A (PKA) at 0.1 mg/ml to the PP2A-B56δ holoenzymes [PKA: PP2A-B56δ = 1:100 (w/w)] at 37 °C. A 10 µl reaction solution was taken out at the indicated time from the reaction, mixed with 3 µl of 4× SDS dye (40% glycerol, 8% SDS, 0.25M Tris‐HCL, pH 6.8, 0.4M DTT, and 0.04% bromophenol blue) and immediately heated at 95 °C for 3 min to quench the reaction. Then, 0.05 µg of PP2A-B56δ holoenzyme from each time point was examined by a western blot using an antibody that specifically recognizes phosphorylated S573 of the PP2A-B56δ holoenzyme.¹⁷

III. RESULTS AND DISCUSSIONS

A. Phosphorylation and mutations far away from the C-arm SLiM alter its stability

Since the mutations that are most strongly associated with intellectual disability (E198K and E200K) are located at a distance away from both the active site and the SLiM binding site, it is hypothesized that they have an allosteric effect on these sites. To study this allosteric mechanism, we conducted all-atom MD simulations of the E198K and E200K variants of PP2A-B56δ as well as the WT and a WT with four phosphorylated serine residues (N-arm: S88, S89, S90; C-arm: S573) labeled as WT4P. The phosphorylated residues were selected according to the previously published frequency of B56δ phosphorylation from the PhosphositePlus database (https://www.phosphosite.org).⁶⁷ Our MD simulations are based on the cryo-EM structure of PP2A-B56δ in the closed conformation in which the C-arm SLiM (L595, S598, and E600) is bound to the groove on the core of B56δ and the N-arm overlaps on top [Fig. 1(a)]. The SLiM binding groove serves as an anchoring point for substrate SLiMs; however, it is blocked by the tip of the C-arm in the closed form.

In our MD simulations, the C-arm SLiM of the E198K and WT4P variants fluctuates significantly more compared to WT while the E200K variant shows a slight decrease in fluctuation [see the RMSF curves in Fig. 1(b)]. This result is unexpected since the residues E198 and E200 are spatially adjacent, and their respective mutation to lysine contributes the same change in the magnitude of charge and size of their sidechains. It is also surprising that WT4P destabilizes the C-arm SLiM in a similar magnitude as E198K despite the fact that they differ in charge by four in this region due to pS573. This observation suggests that the structural properties of these mutations alone cannot explain their effect on protein activity. Since the mutation sites and pS573 lie more than 30 Å away from the closest C-arm SLiM residue, we hypothesize that a perturbation of an underlying dynamic allosteric network may explain our observation. The increased fluctuation of the C-arm SLiM suggests that the binding groove is more accessible to a substrate. Alternatively, this could also mean that substrate SLiMs would have a lower affinity. To address this alternative possibility, we computed the fluctuation of residues lining the SLiM binding groove (Fig. S1). We found no significant differences, leading us to rule out the possibility that the substrate-SLiM binding would be destabilized.

B. The E198K mutation and phosphorylation of B56δ loosen the C-arm SLiM by repartitioning allosteric channels

In our MD simulations, the B56δ does not undergo drastic conformational changes (with RMSD below four, see Fig. S2); therefore, we hypothesized that mutations and phosphorylation of the N/C-arm interface disrupt the coordinated dynamics in B56δ, and this disruption spreads to the C-arm SLiM and its binding groove. To learn how far-away mutations can alter the stability of the C-arm SLiM, we analyzed the allosteric networks in variants of PP2A-B56δ. The networks were computed based on the correlated motion of atoms in the whole protein complex. Furthermore, the shortest pathways through those networks were computed between the mutation sites (residues E/K198 and E/K200) and the conserved C-arm SLiM residues (L595, S598, and E600) (see Sec. II for details). In addition to the 24 optimal pathways, we consider 1200 suboptimal pathways. Based on this ensemble of pathways, we investigate how the fluctuations at the source propagate through the structure and how strongly they impact the sink residues. The pathways in the WT engage residues exclusively in the core of B56δ before directly connecting to the C-arm SLiM [Fig. 2(a)]. Pathways of E200K are surprisingly similar with the WT, mostly using the same residues (Fig. 3). Remarkably, pathways in E198K and WT4P partition into a second spatially distinct group of pathways that branch off close to the E200 and follow the C-arm.

FIG. 2.

The E198K mutation and phosphorylation of B56δ loosen the allosteric connection to SLiM. (a) Allosteric pathways from mutation to the SLiM in the WT, E198K, E200K, and WT4P. The C-arm is colored magenta, and the SLiM residues are labeled in WT. The percentage labels reflect the number of paths in channel 1 or channel 2 (labeled blue and orange, respectively). (b) The mean weight of allosteric channels beginning at the mutation sites and ending on the SLiM normalized by the WT. The confidence interval corresponds to the standard deviation of the mean based on networks derived from ten samples of bootstrapped trajectories. (c) The number of contacts between the C-arm SLiM residues and the core of B56δ. (d) Effects of mutations on substrate-SLiM binding assessed by the pulldown of WT and mutant holoenzymes by GST-tagged B56 SLiM in the substrate. Each data point is normalized by the mean value for WT. See Sec. II for details of the GST-mediated biophysical pulldown assay. The P values were obtained with Welch’s t-test.

FIG. 3.

Optimal pathways starting from the mutation sites and ending on the C-arm SLiM residues. The source residues (E198 and E200) are colored yellow, and sink residues (C-arm SLiM residues, L595, S598, and E600) are colored red. Residues belonging exclusively to channel 1 are colored blue. Residues belonging exclusively to channel 2 are colored orange. Residues shared among channels are green. The thickness of the arrows corresponds to the number of pathways connecting two residues.

To quantify the strength of allosteric communication through each channel, we developed a new approach to group the ensemble of paths into two channels (see Sec. II for details) based on the gap in the eigenspectrum of the pathway inter-correlation matrix. We applied this method to cluster all paths in between the mutation sites and the C-arm SLiM residues across all studied PP2A variants and assigned each path into one of the two channels (Fig. S3). We then computed the weight of each pathway in the network so that a larger channel weight indicates a stronger correlation of atomic motions. The percentage of all paths in a given channel is labeled in Fig. 2(a), and the total weight of paths in a channel normalized by the WT is shown in Fig. 2(b).

The weight of allosteric channel 1 is the highest in the WT followed by the E200K, while the weights of E198K and WT4P are significantly lower, which is mirrored by the percentage of paths belonging to each channel [100% for WT and E200K, 73% for WT4P and only 50% for E198K, see Fig. 2(a)]. Since the channel 1 connects directly to the SLiM, the weight reflects its stability in the groove, agreeing with the observed differences in the fluctuations [Fig. 1(b)]. Channel 2, on the other hand, connects to the C-arm above the SLiM, implying that the C-arm dynamics are correlated with the mutation sites rather than with the B56δ core; hence, channel 2 destabilizes the C-arm SLiM. The weight of Channel 2 is higher in E198K (0.23 ± 0.11) where it connects to L595 and S598, while in WT4P (0.37 ± 0.07), it only connects to L595. This suggests that the E198K mutation and phosphorylation can selectively reduce C-arm SLiM binding without altering the affinity of B56δ to SLiM-bearing substrates. However, this effect is more pronounced in E198K. While E200K reduces the weight of Channel 1 (0.71 ± 0.08) relative to WT (1.00 ± 0.15), it does not engage Channel 2 at all [as shown in Fig. 2(a), the percentage of channel 2 paths in E200K is 0%]; hence, the C-arm SLiM stability is unlikely to be disrupted in this variant. Loosening of the C-arm SLiM in E198K and WT4P is also manifested in the straightforward measurement of the number of contacts between the core of B56δ and the C-arm SLiM [Fig. 2(c)].

These results are further validated by our experimental measurements of substrate-SLiM binding to the B56δ [Fig. 2(d)]. This measurement was achieved with a pulldown assay by adding substrate SLiMs to immobilized holoenzymes (see Sec. II for details). The substrate SLiMs that bind remain with the immobilized layer, and the ligands that do not bind are washed away allowing for binding affinity to be measured. In these results, higher substrate-SLiM binding indicates lower C-arm SLiM binding, implying that the C-arm SLiM is destabilized the most in E198K, followed by E200K, and the WT is the most stable. Currently, it is hard to obtain experimental substrate-SLiM binding affinity to the phosphorylated PP2A-B56δ since the exact phosphorylation pattern is affected by diverse kinases in human proteome and likely by auto-dephosphorylation once the holoenzyme is activated. Nevertheless, our results suggest that phosphorylation of S573 and S88-90 will destabilize the C-arm SLiM by diminishing the strength of allosteric channel 1 and engaging channel 2, similar to E198K.

C. Mutations and phosphorylation of B56δ expose the active site

The enzymatic activity of PP2A is realized in the active site of the catalytic subunit, and the closed C-arm of B56δ blocks access to the active site [Fig. 1(a)]. Our biochemical experiments with SLiM-less PP2A substrates determined that E198K and E200K mutations raise the catalytic rate of the PP2A-B56δ [Fig. 4(c)]. Since the active site lies ∼20 Å away from the mutation sites, we hypothesized that these mutations could alter the accessibility of the active site by loosening the C-arm through a network of interactions, similar to C-arm SLiM.

FIG. 4.

Modifications of B56δ loosen the allosteric pathway to the active site. (a) Allosteric pathways from mutation to the active site residues in the WT, E198K, E200K, and WT4P. The paths belonging to each variant are colored blue, orange, green, and red, respectively. (b) Mean and standard deviation of path weights from the mutation sites to the active site residues D57 and D58 obtained from networks of ten bootstrapped trajectories normalized by the WT. (c) Enzyme kinetics of the WT and mutant PP2A-B56δ holoenzyme activity toward a SLiM-independent substrate, KRpTIRR.

We examined the effect of mutations on the catalytic site by computing the optimal pathways in the allosteric network between mutation sites and two aspartate residues (D57 and D85) coordinating the catalytic metal ions. In addition to the 16 optimal pathways between all pairwise combination of source and sink atoms, we consider 800 suboptimal pathways. Here, we do not observe a clear separation of pathways into distinct path channels; therefore, we analyze all the allosteric paths together [Fig. 4(a)]. The total weight of paths shows that the allosteric connection is the strongest in the WT (1.0 ± 0.15), followed by E200K (0.74 ± 0.07) while the WT4P and E198K show a significantly lower path weight (0.56 ± 0.03 and 0.59 ± 0.05) [Fig. 4(b)]. A smaller path weight indicates that the C-arm interface between the B56δ and the catalytic subunit can move independently, likely increasing access to the active site; therefore, a lower path weight should correspond to a higher catalytic activity. Indeed, these results reflect the experimentally measured catalytic rate with the pThr hexapeptide, KRpTIRR, which does not have a SLiM [Fig. 4(c)]. Our results suggest that both mutations and phosphorylation expose the active site; however, the E198K mutation is the most efficient in this respect. Pathways from mutation sites to the active site engage more residues compared with pathways to the C-arm SLiM (compare Figs. 3 and S4); however, many residues are also shared between all variants (Fig. S5), suggesting that the dynamics are communicated through conserved pathways.

Previous experimental measurements of PP2A-B56δ phosphorylation suggest that S573 is the most frequently phosphorylated site.⁶⁷ This residue lies adjacent to the active site and can directly hinder access to the active site. We computed the solvent-accessible surface area (SASA) of this residue and found that in contrast to WT and E200K, both E198K and the WT4P (which carries a phosphate group on the S573) display a large increase in SASA (from 0.12 ± 0.03 nm² in WT and 0.12 ± 0.03 nm² in E200K to 0.41 ± 0.07 nm² in E198K and 0.45 ± 0.06 nm² in WT4P) [Fig. 5(a)], indicating that this part of the C-arm is more exposed to the solvent and hence the active site is more accessible. Consistently, the experimentally measured phosphorylation rate of the PP2A-B56δ holoenzyme in vitro by PKA was significantly enhanced by E198K and barely affected by E200K [Fig. 5(b)]. These results indicate that this part of the C-arm is more exposed to the solvent in case of E198K mutation or upon activation phosphorylation, and hence, the active site is more accessible. Phosphorylation of B56δ has been previously associated with the activation of PP2A.^68–70 Thus, mutations and phosphorylation of the PP2A-B56δ N/C-arm interface can increase its activity by loosening the closed C-arm and increasing access to its active site. A more exposed S573 in the E198K mutant is likely associated with higher holoenzyme activity under both basal and activated conditions. Based on the weakened allosteric connection between B56δ and the active site as well as greater exposure of S573 to the solvent, WT4P and E198K should display enhanced catalytic rates.

FIG. 5.

Exposition of S573 to the solvent and its phosphorylation rates. (a) SASA of S573 in variants of PP2A. (b) Effects of B56δ disease mutations on S573 phosphorylation of the holoenzyme by PKA in vitro. The data are normalized to the maximum phosphorylation of WT in a 40-min time course. The number of repeats, their scatter plots, averages, and SEM are shown. The P values for the comparison of WT and disease variants, E198K and E200K, are 0.002 and 0.3, respectively (calculated using two-sided Wilcoxon rank sum test).

For E200K, our computed SASA for S573 and our experimental measurement of phosphorylation indicate a reduced difference in exposure of S573 relative to WT. Similarly, our results indicate that the SLiM binding is not increased in E200K [Figs. 2(b) and 2(c)]. However, the allosteric connection to the active site [Fig. 4(b)] is reduced, and correspondingly, the experimentally measured catalytic rate is enhanced relative to the WT [Fig. 4(c)]. Therefore, E200K can still alter the enzymatic activity, however, less than E198K. Since the human proteome has ∼1500 B56 SLiMs, even the mild effects of E200K on basal SLiM-binding and the phosphatase active site might reduce the robustness of, if not perturb significantly, many signaling processes required for optimum neurological functions.

We have demonstrated how mutations and phosphorylation can alter the allosteric wiring of PP2A-B56δ. The E198K mutation as well as phosphorylation of B56δ can allosterically weaken the correlated dynamics that link the C-arm SLiM to the core of B56δ. Simultaneously, this mutation engages alternative allosteric pathways, thus explaining the mechanism of experimentally observed increased substrate-SLiM binding in the E198K mutant compared to WT [Fig. 2(d)]. This insight was made possible by examining all pairwise combinations of atoms in the source and sink residues rather than the single optimal allosteric pathway between the source and sink.

Our newly adopted method for lumping allosteric paths is useful in two ways: it helps interpretation by combining numerous paths into a few channels, and it allows for the quantification of the relative importance of each channel. It also provides an objective metric for selecting the optimal number of channels by utilizing the gap between the eigenvalues of the inter-correlation matrix. Compared to the previous approach of clustering allosteric pathways based on the geometric proximity of residues,³¹ our method is based on correlation similarity, meaning that pathways in the same channel will account for similar correlated dynamics. This is required to distinguish the difference between paths where residues are close in the structure but have a different role in the correlated motion. This is exemplified by the two channels between residues E198, E200, and the C-arm SLiM in WT4P, which share many residues between channels 1 and 2 (Fig. 3) and thus would be clustered into a single channel based on spatial proximity. Our allosteric pathway lumping performs well in comparing pathways across different variants of PP2A-B56δ; however, it relies on an average correlation matrix to compute the inter-correlation between different systems; thus, it works best when clustering across similar systems, such as point mutants of a specific protein. Despite this, functional interpretation of the allosteric channels can still be challenging. Our results indicate that mutations to E198 and E200 can loosen the correlated dynamics between B56δ and the catalytic subunit, which leads to greater exposure of the active site. However, here, the lumping of allosteric pathways does not reveal a clear role of individual channels; we thus leave this analysis out. Additional insight into substrate binding to the active site and catalysis in PP2A-B56δ may allow one to attribute a specific role to the channels in this case.

Our allosteric pathway lumping approach inherits the methodology from the transition path theory,^71,72 which is based on the reactive flux of trajectories. This theory has been widely applied to elucidate major pathways for conformational changes.^73–86 In the original application, the importance of individual pathways can be evaluated by iteratively subtracting the flux of the strongest pathway from the overall flux. This approach could be useful to analyze allosteric pathways; however, the path weights based on an LMI matrix are not directly related to flux. Instead, the flux through allosteric paths may be computed based on the current flow between the source and sink residues.³² With an allosteric network based on current flow, it may be possible to extract the contribution of specific paths to the overall flux and evaluate its impact on the allosteric signaling. In the future, current flow combined with transition path theory may provide an alternative approach to obtain an ensemble of suboptimal allosteric pathways, which can then be passed to our allosteric path lumping algorithm. This approach could also be applied in conjunction with Markov state models to reveal the role of allosteric networks and pathways through them in the context of thermodynamic and kinetic properties of the system.^43,44

IV. CONCLUSION

In this study, we have combined computational and experimental approaches to show how mutations and phosphorylation of PP2A-B56δ disrupt its normal autoinhibition mechanism by perturbing the allosteric pathways. We find that the E198K mutation exerts its effect by repartitioning the allosteric pathways leading to the C-arm SLiM, weakening its connection to the core of B56δ and, thereby, facilitating substrate-SLiM binding. We also showed that our new approach to lumping allosteric paths into channels is instrumental in interpreting the role of numerous pathways. We expect that it can be applied broadly to other complex dynamic systems. Additionally, we demonstrate that E198K and E200K mutations as well as phosphorylation of B56δ may raise the catalytic activity of PP2A by weakening the correlated dynamics between the mutation sites and the active site, rendering the latter more accessible. We find that modifications on the interface of the N/C-arms operate through common pathways; however, identical mutations at neighboring residues can produce markedly different effects and perturb different interfaces independently. Our computational results agree well with our experimental measurements of substrate-SLiM binding and catalytic activity with SLiM-independent substrates with E198K as well as phosphorylation at S573. Supported by these results, we suggest that phosphorylation of B56δ activates PP2A through a similar mechanism: loosening of the N/C-arms leading to destabilization of the C-arm SLiM and exposure of the active site. Overall, our findings provide a better understanding of the regulation of PP2A-B56δ, revealing a non-straightforward relationship between identical point mutations to neighboring residues and the dual-autoinhibitory mechanism. In this work, we focused on the closed, inactive conformation of PP2A-B56δ. Since a functional conformational change can alter the residue–residue contacts and may redefine allosteric signaling pathways,⁸⁷ future studies should investigate how mutations impact the opening of the N/C-arms and binding of substrate proteins.

SUPPLEMENTARY MATERIAL

The supplementary material consists of five supplementary figures: Figs. S1–S5.

AUTHOR DECLARATIONS

Conflict of Interest

The authors have no conflicts to disclose.

Author Contributions

Kirill A. Konovalov: Conceptualization (supporting); Data curation (lead); Formal analysis (lead); Methodology (equal); Writing – original draft (lead); Writing – review & editing (equal). Cheng-Guo Wu: Data curation (equal); Formal analysis (equal); Writing – original draft (supporting). Yunrui Qiu: Data curation (supporting); Formal analysis (supporting); Methodology (equal); Writing – original draft (supporting). Vijaya Kumar Balakrishnan: Data curation (supporting); Formal analysis (supporting); Writing – original draft (supporting). Pankaj Singh Parihar: Data curation (supporting); Formal analysis (supporting). Michael S. O’Connor: Visualization (equal); Writing – original draft (equal); Writing – review & editing (equal). Yongna Xing: Conceptualization (equal); Data curation (supporting); Formal analysis (supporting); Funding acquisition (lead); Supervision (equal); Writing – original draft (supporting); Writing – review & editing (supporting). Xuhui Huang: Conceptualization (equal); Funding acquisition (equal); Methodology (supporting); Project administration (equal); Resources (equal); Supervision (lead); Writing – original draft (supporting); Writing – review & editing (supporting).

DATA AVAILABILITY

The data that support the findings of this study are available from the corresponding authors upon reasonable request.