1. Cell Biology
  2. Human Biology and Medicine
Download icon

Ryanodine receptor dispersion disrupts Ca2+ release in failing cardiac myocytes

  1. Terje R Kolstad
  2. Jonas van den Brink
  3. Niall MacQuaide
  4. Per Kristian Lunde
  5. Michael Frisk
  6. Jan Magnus Aronsen
  7. Einar S Norden
  8. Alessandro Cataliotti
  9. Ivar Sjaastad
  10. Ole M Sejersted
  11. Andrew G Edwards
  12. Glenn Terje Lines
  13. William E Louch  Is a corresponding author
  1. Oslo University Hospital and University of Oslo, Norway
  2. University of Oslo, Norway
  3. Simula Research Laboratory, Norway
  4. University of Glasgow, Great Britain
  5. Bjørknes College, Norway
Research Article
  • Cited 0
  • Views 535
  • Annotations
Cite this article as: eLife 2018;7:e39427 doi: 10.7554/eLife.39427

Abstract

Reduced cardiac contractility during heart failure (HF) is linked to impaired Ca2+ release from Ryanodine Receptors (RyRs). We investigated whether this deficit can be traced to nanoscale RyR reorganization. Using super-resolution imaging, we observed dispersion of RyR clusters in cardiomyocytes from post-infarction HF rats, resulting in more numerous, smaller clusters. Functional groupings of RyR clusters which produce Ca2+ sparks (Ca2+ release units, CRUs) also became less solid. An increased fraction of small CRUs in HF was linked to augmented ‘silent’ Ca2+ leak, not visible as sparks. Larger multi-cluster CRUs common in HF also exhibited low fidelity spark generation. When successfully triggered, sparks in failing cells displayed slow kinetics as Ca2+ spread across dispersed CRUs. During the action potential, these slow sparks protracted and desynchronized the overall Ca2+ transient. Thus, nanoscale RyR reorganization during HF augments Ca2+ leak and slows Ca2+ release kinetics, leading to weakened contraction in this disease.

https://doi.org/10.7554/eLife.39427.001

eLife digest

The muscle cells of the heart coordinate how they contract and relax in order to produce the heartbeat. During heart failure, these cells become less able to contract. As a result the heart becomes inefficient, pumping less blood around the body.

For the cardiac muscle cells to contract, the levels of calcium ions in the cells needs to rapidly increase. In failing hearts, these increases in calcium ion levels are smaller, slower and less well coordinated. It was not known what causes these changes, making it difficult to treat heart failure.

Calcium ions are released in cardiac muscle cells through protein channels called ryanodine receptors. These receptors form clusters that allow them to synchronize when they open and close. Could the reorganization of ryanodine receptors account for the problems seen in failing hearts? To investigate, Kolstad et al. examined rat hearts using a technique called super-resolution microscopy. This showed that the clusters of ryanodine receptors break apart during heart failure to form smaller clusters. Further experiments showed that calcium ions ‘leak’ from these smaller clusters, reducing the amount of calcium that can be released into cardiac muscle cells during each heartbeat. Released calcium also spreads between the dispersed clusters, resulting in a slower rise of the calcium levels in the cells. Both changes contribute to weakened contractions of cells in failing hearts.

Therefore, heart failure can be traced back to very small rearrangements of the ryanodine receptors. This understanding will help researchers as they investigate new ways to treat heart failure.

https://doi.org/10.7554/eLife.39427.002

Introduction

The basic processes for cardiac excitation-contraction coupling are well described. Depolarization of the sarcolemma triggers the opening of voltage-gated L-Type Ca2+ channels (LTCCs), and the resulting Ca2+ influx elicits additional Ca2+ release via Ryanodine Receptors (RyRs) in the sarcoplasmic reticulum (SR). This process of Ca2+-induced Ca2+ release leads to a sharp increase in cytosolic Ca2+ concentration which initiates cardiomyocyte contraction. In ventricular myocytes, Ca2+ release is tightly controlled by the arrangement of LTCCs and RyRs in dyads, with LTCCs present in t-tubules juxtaposed from RyRs across a narrow 12–15 nm dyadic cleft (Bers, 2001). The RyRs themselves are organized into clusters; an arrangement that couples their gating, promoting synchronized opening and closing of neighbouring channels (Marx et al., 2001Sobie et al., 2006). Recent data have indicated that neighbouring clusters of RyRs can also act concertedly if the Ca2+ diffusion distance between them is sufficiently short (Macquaide et al., 2015). Referred to as ‘superclusters’ or Ca2+ Release Units (CRUs), these functional arrangements of RyR clusters generate Ca2+ sparks, the fundamental units of SR Ca2+ release in cardiomyocytes (Cheng et al., 1993). Ca2+ sparks are not only elicited by LTCC opening, but also occur spontaneously during diastole, where spark frequency and geometry can be measured to assess CRU function. While Ca2+ sparks are an important source of RyR-mediated Ca2+ leak from the SR, ‘silent’ or ‘non-spark’ events also occur, and involve the opening of a subset of RyRs within a CRU; so-called ‘quarky’ release (Brochet et al., 2011).

Impaired cardiomyocyte Ca2+ homeostasis is believed to importantly contribute to reduced cardiac contractility and arrhythmogenesis in heart failure (HF). SR Ca2+ release is reduced and slowed in this condition, and these changes have been linked to altered dyadic structure (Louch et al., 2010). We and others have observed marked remodeling of the t-tubular system in failing cardiomyocytes, while RyRs remain predominantly distributed along z-lines (Song et al., 2006; Louch et al., 2006; Heinzel et al., 2008). Thus, the coupling between LTCCs and RyRs is disrupted, with ‘orphaned’ CRUs exhibiting delayed Ca2+ release only after trigger Ca2+ diffuses from intact dyads. However, abnormal gaps occurring between t-tubules only account for a fraction of the overall de-synchronization of Ca2+ release in HF (Louch et al., 2006; Øyehaug et al., 2013). This suggests that other alterations might also occur, perhaps at the nanometer scale of CRU organization, which hinder efficient triggering of Ca2+ release. CRU reorganization could in principle contribute to increased Ca2+ leak, including silent leak, which is a hallmark of heart failure (Zima et al., 2010; Walker et al., 2014). Exaggerated Ca2+ leak in failing cells has been linked to reduced SR Ca2+ content and depressed contractile function, elevation of resting Ca2+ levels and impaired relaxation, pro-arrhythmic early and delayed afterdepolarizations, and energetic inefficiency as Ca2+ is redundantly cycled (Bers, 2014). Thus, a detailed understanding of CRU structure and function in failing cells is critical.

The advent of super-resolution microscopy techniques has markedly improved our ability to visualize and quantify CRU organization (Baddeley et al., 2009Macquaide et al., 2015; Jayasinghe et al., 2018). However, these techniques have not previously been employed to examine RyR configuration in HF. Using direct stochastic optical reconstruction microscopy (dSTORM), we presently report that CRUs become dispersed in failing myocytes, as RyR clusters are broken apart. With the aid of mathematical modeling, we directly link these changes in CRU structure to experimentally measured increases in RyR leak and slowed SR Ca2+ release, identifying a novel mechanism underlying pathological remodeling of Ca2+ homeostasis in this disease.

Results

dSTORM imaging reveals dispersion of CRUs in failing myocytes

Imaging was performed on isolated, fixed cardiomyocytes with antibody labelling of RyR2. Using diffraction-limited confocal imaging (resolution ≈250 nm) and Structured Illumination Microscopy (SIM, resolution ≈120 nm), the localization of RyRs along z-Lines was clearly apparent, but organization of RyRs within CRUs was not discernable (Figure 1A). With dSTORM imaging, spatial resolution was markedly improved (mean localization precision = 21 ± 3 nm) enabling detailed CRU geometry to be assessed. For analysis of RyR cluster and CRU configuration, acquired raw images were fitted to a 30 × 30 nm grid, corresponding to the quatrefoil structure of the RyR protein (Baddeley et al., 2009). Thresholding was then performed to create binary images (Figure 1B), enabling quantification of RyR clusters, with an RyR counted as present if >half the area of a 30 nm square was above threshold. RyR clusters were defined by occupied, neighbouring grid positions, and CRUs were delineated by collecting neighbouring RyR clusters located within 150 nm (Macquaide et al., 2015) (red boundaries in Figure 1B) or 100 nm (Baddeley et al., 2009; Hou et al., 2015) (Figure 2—source data 1). RyR organization was compared in cardiomyocytes from rats with post-infarction HF and cells from Sham-operated controls. Overall RyR expression was similar in Sham and HF, as evidenced by Western blotting of ventricular homogenates (Figure 2— figure supplement 1), and equivalent RyR labeling density in cardiomyocytes (41.9 ± 1.4 RyR/µm, 40.4 ± 1.3 RyR/µm in Sham, HF respectively). In both groups, RyR staining showed a predominantly transverse, striated pattern (Figure 2A). However, despite rather similar organization of RyRs at the macroscale, nanoscale dSTORM imaging revealed fragmentation of RyR clusters in failing cardiomyocytes (see insets in Figure 2A). Cluster breakup resulted in a reduction in the number of RyRs per cluster, and a greater proportion of small clusters in HF (Figure 2B). The overall number of clusters increased accordingly in failing cells (Figure 2D), and inter-cluster distance was reduced (Figure 2E; see Figure 2—source data 1 for mean data across animals). Consistent with fragmentation of clusters into smaller adjacent groupings, the number of clusters contained in a CRU increased in HF (Figure 2F), although the number of RyRs per CRU decreased (Figure 2C) since RyR clusters were markedly reduced in size. Convex hull analysis (see methods) revealed a consequent decrease in CRU solidity in HF (Figure 2G). Thus, RyR reorganization in failing cells resulted in CRUs with a more sparse, dispersed configuration of smaller sub-clusters.

dSTORM imaging enables quantification of RyR localization within Ca2+ release units (CRUs).

RyR imaging was performed with antibody labelling of isolated and fixed rat ventricular cardiomyocytes. (A). Imaging of RyRs with confocal microscopy (left panel) or Structured Illumination Microscopy (SIM, centre panel) revealed a predominantly striated pattern of RyR localization across cells, but individual CRUs could not be discerned (magnified regions in lower panels). dSTORM imaging provided markedly improved spatial resolution enabling identification of RyR clusters (scale bars = 5 µm). (B). Quantification of RyR localization was performed by fitting raw images to a 30 × 30 nm grid (Baddeley et al., 2009), and performing thresholding to create binary images; an RyR was counted as present if > half the area of a 30 nm square was suprathreshold. CRUs were defined as collections of RyR clusters with an edge-to-edge distance < 150 nm (Macquaide et al., 2015) (red boundaries) or < 100 nm (Baddeley et al., 2009; Hou et al., 2015). (Scale bar = 2 µm).

https://doi.org/10.7554/eLife.39427.003
Figure 2 with 1 supplement see all
RyRs are dispersed in failing cardiomyocytes.

Alterations in nanoscale RyR organization were examined in cardiomyocytes from rats with post-infarction heart failure (HF). Representative images show that macroscale organization of RyRs was similar in HF and Sham-operated controls (A), upper panels). However, nanoscale examination revealed that RyR clusters were broken apart in HF. For the magnified regions in (A), conversion from raw dSTORM to binary images is shown in the middle and lower panels (saturation levels indicated by high-low look-up table). Mean measurements showed fewer RyRs per cluster in failing cells, with an increased fraction of small clusters (B). Dispersion of RyR clusters into smaller fragments resulted in an increased overall number of clusters (D), reduced inter-cluster distances (E) and inclusion of more clusters in each CRU (F). Overall CRU composition became less solid in failing cells ((G), assessed by convex-hull analysis), as the average CRU contained fewer RyRs (C). See Figure 2—source data 1 for analysis of 100 nm vs 150 nm CRU inclusion criterion (ncells = 46, 50 in Sham, HF; *=P < 0.05 vs Sham).

https://doi.org/10.7554/eLife.39427.004

RyR dispersion in HF augments ‘silent’ RyR Ca2+ leak

We examined the functional implications of altered nanoscale organization of RyRs, first hypothesizing that RyR dispersion would augment SR Ca2+ leak in failing cardiomyocytes. Total Ca2+ leak was assessed in SR microsomes obtained from the left ventricle of Sham and failing hearts.

Following initiation of microsomal Ca2+ uptake by addition of ATP, SERCA function was halted by thapsigargin treatment to reveal RyR-mediated Ca2+ leak (Figure 3A). The Ca2+ leak rate was markedly higher in HF compared to Sham (inset in Figure 3A, mean data in Figure 3C). In agreement with previous work (reviewed in (Bers, 2006)), we additionally observed significantly slowed SR Ca2+ uptake in HF (Figure 3B) and lower Ca2+ content (Figure 3D). Ca2+ spark-mediated RyR leak was assessed by confocal linescan imaging of freshly-isolated cardiomyocytes (Figure 3E). While the average Ca2+ release per spark (spark mass) was significantly increased in HF cells, this effect was offset by a tendency toward lower spark frequency (Figure 3F,G). Indeed, overall spark-mediated Ca2+ leak was similar in HF and Sham cells (Figure 3H). Since we observed an increase in total RyR leak in HF, these results are consistent with augmented'silent’, non-spark mediated leak in failing cells.

Failing cardiomyocytes exhibit increased ‘silent’ RyR leak.

Total RyR-mediated Ca2+ leak was assessed in SR microsomes using fura-2 fluorescence (A). Vesicular Ca2+ uptake was initiated by addition of ATP, and halted by addition of thapsigargin. SR Ca2+ leak was estimated as the thapsigargin-induced rate of rise of [Ca2+], normalized to releasable SR content (rise in [Ca2+] induced by the RyR opener 4-chloro-m-cresol, CMC). While the rate of SR Ca2+ uptake was reduced in HF relative to Sham (B), total RyR leak was increased (C) even with a slight reduction in the releasable Ca2+ store (D). (n = 8 from 3 Sham hearts, 7 from 3 HF hearts; *=P < 0.05 vs Sham). To assess whether elevated SR leak in failing cells could be attributed to Ca2+ sparks, line-scan confocal imaging of resting cardiomyocytes was employed (E). Ca2+ spark mass was increased in HF relative to Sham (F), due to augmented spark geometry (Figure 5B). However, since spark frequency tended to be reduced (G), overall Ca2+ spark-mediated leak was similar in Sham and HF (H). These results are consistent with increased ‘silent’ non-spark-mediated SR leak in HF cardiomyocytes. (FWHM = full width at half maximum, FDHM = full duration at half maximum; ncells = 43 in Sham, 50 in HF; *=P < 0.05 vs Sham).

https://doi.org/10.7554/eLife.39427.007

To investigate whether increased silent Ca2+ leak could be linked to RyR dispersion, we employed a mathematical model of the dyad (illustrated schematically in Figure 4—figure supplement 1A) that enabled simulation of Ca2+ sparks with varied placement of RyRs within the CRU. We first incorporated small idealized CRUs containing as few as 4 RyRs (Figure 4A), as our dSTORM imaging indicated that HF cells contain an increased fraction of small CRUs (Figure 2C). During repeated simulations, a single RyR was opened at a random position within the CRU, and subsequent triggered RyR openings were allowed to proceed stochastically. Simulated Ca2+ release events with amplitudes ΔF/F0 ≥0.4 were defined as sparks, based on the detection threshold determined experimentally (see methods). Ca2+ release from the smallest CRUs was never detected, but a progressively greater proportion of events yielded visible sparks as the number of RyRs in these idealized dyad geometries was increased (Figure 4A). These results support the assertion that an increased fraction of small CRUs in HF promotes undetectable, silent Ca2+ leak.

Figure 4 with 1 supplement see all
CRU dispersion provides the structural basis for silent RyR leak in HF.

A mathematical model of the dyad was employed to examine the effects of CRU dispersion on Ca2+ sparks and non-spark mediated RyR leak. (A). As dSTORM imaging indicated an increased fraction of small CRUs in HF (Figure 2C), small idealized CRUs were initially modelled with as few as 4 RyRs. Simulated Ca2+ sparks (300 consecutive simulations) were never detected for the smallest CRUs, based on an experimentally determined spark detection threshold of ΔF/F0 = 0.4. Higher probability of visible spark generation (fidelity) was observed for larger CRUs. (B). Real CRU geometries obtained by dSTORM imaging were employed to simulate sparks from larger dyads. Four configurations were modelled with varying numbers of constituent RyR clusters, but similar total RyR number (≈55). While the single-cluster CRU exhibited high Ca2+ spark fidelity, lower probability of spark generation was observed in dispersed, multi-cluster CRUs (fidelity indicated by colour scale). These data support that CRU rearrangement during HF promotes silent RyR leak, due to an increased fraction of both small CRUs as well as larger CRUs with dispersed, irregular configurations.

https://doi.org/10.7554/eLife.39427.008

We next examined whether dispersion of clusters in larger more realistic CRUs could similarly contribute to increased silent Ca2+ leak in failing cells. To this end we incorporated real CRU geometries obtained by dSTORM imaging into the model (Figure 4B). Four CRUs were selected containing roughly the same number of RyRs, but with different numbers of RyR clusters (1, 3, 7 or 10 clusters). As in the simulations described above for idealized CRU geometries, a single, randomly chosen RyR was opened in each simulation, to determine the likelihood that such triggering would result in a detectable Ca2+ spark. While relatively high fidelity spark generation was observed for the single-cluster CRU, Ca2+ release was more rarely observed to propagate between clusters, and spark fidelity was significantly lower in multi-cluster CRUs (Figure 4B). This reduced efficiency of Ca2+ spark triggering in dispersed CRUs partly resulted from greater Ca2+ diffusion distance between neighbouring clusters, as demonstrated by progressively increasing the distance between RyR clusters in an idealized dyad (Figure 4—figure supplement 1B). Furthermore, released Ca2+ is less efficiently confined in the dyadic space when the junctional SR has a more distributed and irregular shape. This latter point was demonstrated in the model by altering the amount of junctional SR surrounding the CRU; increasing junctional SR ‘padding’ increased spark fidelity in both idealized dyads (Figure 4—figure supplement 1B) and dSTORM-based geometries (Figure 4—figure supplement 1C).

In summary, these results indicate that nanoscale reorganization of RyRs in HF promotes non-spark-mediated SR Ca2+ leak by two mechanisms: (1) by creating smaller CRUs which produce Ca2+ release events below the detection limit, and (2) by creating more distributed CRU configurations in which multiple RyR clusters are less likely to co-operatively generate sparks.

CRU dispersion in HF causes slowing of Ca2+ sparks

We next hypothesized that CRU dispersion would slow cardiomyocyte Ca2+ release; a hallmark of HF. Representative confocal recordings of Ca2+ sparks and their temporal profiles are shown in Figure 5A. Spark kinetics in Sham cells generally exhibited rapid rising and declining phases. While some sparks also showed fast kinetics in HF cells, others were markedly slow to rise and decay (Figure 5A). Indeed, measurements of spark rise time and duration exhibited broader distributions and were, on the average, prolonged in HF compared to Sham (Figure 5B). To investigate whether CRU dispersion in HF could underlie slowing of Ca2+ spark kinetics, we again employed our mathematical model with dSTORM-based CRU configurations. During the simulations, the time to opening of each RyR was registered, and the time course of the overall Ca2+ spark determined. Representative simulations show that RyR opening times were delayed in the dispersed, multi-cluster CRUs compared to the solid, single-cluster CRU (Figure 6A). Simulations of Ca2+ spark time courses further showed that the delayed RyR openings in multi-cluster CRUs resulted in more variable kinetics and overall slowing of spark rise time (Figure 6B, mean data Figure 6C), reproducing experimental observations. Of note, although CRUs were observed to contain fewer RyRs in HF than Sham (Figure 2C), simply reducing the RyR number to an equivalent degree in the mathematical model did not markedly alter Ca2+ spark kinetics (Figure 6—source data 1), further confirming a key role of CRU fragmentation in failing cells.

Ca2+ spark kinetics are slowed in HF.

(A) Representative line-scan images of Ca2+ sparks in Sham and HF, selected from the cell-wide scans presented in Figure 3E. Temporal profiles (right panels) show that spark kinetics were generally tightly constrained in Sham, with low values for both time to peak (TTP) and duration (full duration at half maximum, FDHM). Although many sparks were also brief in HF cells, a subset of sparks exhibited slowed kinetics. (B) Distributions of measurements for TTP and FDHM were right-shifted in HF, and mean values were significantly increased. Spark magnitudes tended to be larger in HF than Sham. (nsparks = 130, 100 from 75, 72 cells in Sham, HF; *=P < 0.05 vs Sham).

https://doi.org/10.7554/eLife.39427.011
Figure 6 with 1 supplement see all
RyR dispersion during HF results in slowing of Ca2+ sparks.

To examine whether altered CRU morphology could slow Ca2+ spark kinetics in HF, spark profiles were simulated for a variety of dSTORM-derived RyR configurations. (A) Sparks were triggered by opening a single RyR (circled) which was randomly placed in consecutive simulations (example RyR opening trajectories are shown in the upper panels, with a family of spark time-courses illustrated below). Time to opening was registered for each RyR in the CRU, and the resultant time course of the Ca2+ spark was plotted until the final RyR closure, at which point the simulation was stopped for computational efficiency. Opening times were similar for individual RyRs within a solid, single cluster CRU, and the overall temporal profile of elicited sparks showed rapid kinetics which were rather consistent between consecutive simulations. By contrast, delayed and variable opening times were observed for individual RyRs in multi-cluster CRUs. This resulted in variable and slowed Ca2+ spark kinetics with these CRU configurations, as indicated by temporal spark profiles (A), a right-shifted distribution of time-to-peak measurements (B) and mean data (C). (*=P < 0.05 vs single-cluster CRU).

https://doi.org/10.7554/eLife.39427.012

Slow Ca2+ sparks promote slowing and de-synchronization of the Ca2+ transient

Finally, we examined the consequences of increased variability in Ca2+ spark kinetics for the Ca2+ transient in failing cells. We observed that field-stimulated Ca2+ transients were significantly slower to rise in HF than Sham (Figure 7A–C). This slowing of Ca2+ release was associated with marked de-synchronization of the Ca2+ transient, which we quantified by measuring the variability in time to reach half-maximal fluorescence (TTF50) across the cell (see lower panels in Figure 7A). This ‘dyssynchrony index’ (Louch et al., 2006) was significantly increased in HF compared to Sham, with a strongly right-shifted distribution of values (Figure 7D). T-tubule disruption in failing cells (Figure 7—figure supplement 1) has been previously established in this model of HF (Frisk et al., 2016), and is a recognized cause of Ca2+ release dyssynchrony in this disease (Song et al., 2006; Louch et al., 2006; Heinzel et al., 2008). We examined whether alterations in Ca2+ spark kinetics also promote dyssynchrony, by examining local Ca2+ transients within narrow, 2 µm regions of the line scan. These regions were centered at the locations of spontaneous Ca2+ sparks observed when electrical pacing was halted. We specifically distinguished between locations with ‘slow’ sparks, defined by a rise time >13 ms (ie. 1 S.D. > mean rise time in Sham), and remaining ‘fast’ sparks. By this definition, 24% of sparks in HF cells were defined as slow, while only 13% of Sham sparks fit this definition. Representative examples of such sparks and their temporal profiles are shown in Figure 7E, with corresponding positions along the line scan indicated in Figure 7A. Local transients from slow spark locations in HF exhibited markedly slower rise times than those from fast spark locations in both HF and Sham (Figure 7F). The association between slow sparks and slow local transients was also apparent in ‘heat map’ plots (Figure 7G). These results show that by protracting Ca2+ sparks, CRU dispersion during HF slows and desynchronizes the overall Ca2+ transient.

Figure 7 with 1 supplement see all
Slow Ca2+ sparks promote slow, desynchronized Ca2+ transients in HF.

(A) Representative confocal linescan images of Ca2+ transients in field-stimulated cells (stimulus illustrated as a horizontal line). The overall Ca2+ transient was slowed in HF compared to Sham, as indicated by plots of spatially-averaged Ca2+ transients (A), right panel), and measurements of half rise time (TTF50, (B)) and time to peak (C). Slowed Ca2+ transient kinetics included de-synchronization of Ca2+ release across HF cells, as indicated by profiles of local TTF50 (lower panels in A). The standard deviation of these values, defined as the dyssynchrony index, (Louch et al., 2006) showed a right-shifted distribution in HF compared to Sham (D). To examine the relationship between slowed Ca2+ spark kinetics and de-synchronized Ca2+ transients in HF, local Ca2+ transients were examined within 2 µm regions of the linescan centered at the location of recorded sparks. Paired representative recordings of sparks and local Ca2+ transients are shown in (E and F), respectively, corresponding to indicated positions in A) (vertical arrows). Local Ca2+ release at ‘slow’ spark locations (rise time > 13 ms) was protracted during the action potential, in comparison with local transients with ‘fast’ sparks in both HF and Sham (F). This association is demonstrated by clustering of locations with slow Ca2+ spark and local transient kinetics in ‘heat maps’ (G), and links slowing of Ca2+ release kinetics at the level of the single CRU and whole cell. (Ca2+ transients: ncells = 43 in Sham, 57 in HF; nfast sparks= 18 in Sham, 19 in HF; nslow sparks= 18 in HF; *=P < 0.05 vs Sham).

https://doi.org/10.7554/eLife.39427.015

Discussion

In the present study, we have employed dSTORM imaging to reveal key changes in CRU morphology during heart failure. We specifically observed marked dispersion of RyRs, which resulted in a shift towards smaller RyR clusters and CRUs. Remaining larger CRUs became less solid, with more fragmented configurations. Experiments and mathematical modeling linked these changes in RyR arrangement to two central aspects of impaired Ca2+ homeostasis in failing cells: increased ‘silent’ RyR leak and slowing of Ca2+ release, which are believed causative for reduced contractility in this condition.

In contrast to lower resolution optical imaging techniques such as confocal and SIM microscopy, the dSTORM technique allows quantification of RyR organization within CRUs. We presently employed this technique with a grid-based quantification method for RyR counting, as previously developed by the Soeller group (Baddeley et al., 2009; Hou et al., 2015). Using images collected close to the cell surface (depth of 200 – 500 nm, see also Macquaide et al. [Macquaide et al., 2015]), we calculated that an average cluster contains approximately 14 RyRs in our healthy, Sham-operated rat ventricular cells. This estimation is in close agreement with previous estimates made at the cell surface (14 RyRs/cluster) (Baddeley et al., 2009), but considerably lower than estimates made deep within the cell interior (Hou et al., 2015), a discrepancy which may reflect regional differences in RyR organization across the cell. However, it should be noted that the grid-based method for RyR counting assumes that the channels lie parallel to the field of view, with a uniform, grid-like configuration; presumptions which are likely less valid when RyRs are visualized internally than at the cell surface. Due to superimposition of internal RyRs along the z-axis, it is therefore likely that present RyR cluster sizes are somewhat underestimated. Assuming that RyR clusters located within 150 nm cooperatively form a CRU (Macquaide et al., 2015), we calculated that average CRUs contain roughly 3 – 4 clusters, and a total of ≈30 RyRs. Using a narrower CRU definition, with inclusion of neighbouring clusters within 100 nm (Baddeley et al., 2009; Hou et al., 2015), reduced average CRU size to ≈25 RyRs (Figure 2—source data 1). Despite possible underestimation of RyR numbers/CRU due to methodological issues noted above, these estimates are in relatively close agreement with an electron microscopy tomography study of mouse ventricular cardiomyocytes (Hayashi et al., 2009). The authors reported that while there was great variability in CRU size and geometry, dyadic volume was an order of magnitude lower than previous estimates (Franzini-Armstrong et al., 1999; Scriven et al., 2013); an average-sized dyad could only hold up to 43 RyRs, with RyRs occupying ≈78% of the dyadic space (≈34 RyRs/CRU). It is estimated that typical sparks result from the activation of somewhat fewer RyRs, ≈20–30 (Shkryl et al., 2012). One possible explanation for this discrepancy is that not all RyRs within a CRU may contribute to every spark. Alternatively, previous work based on estimation of the dyadic size available for RyRs (Hayashi et al., 2009) may have overestimated RyR number if RyRs are not densely packed (Asghari et al., 2014). Our own present estimates have assumed that RyRs located within a fixed distance share the same junctional SR (jSR), which may also be an overestimation since true jSR geometry is unknown. Indeed, our modeling results showed that reducing the degree of jSR ‘padding’ around each RyR cluster markedly reduces the ability of neighbouring clusters to function as cooperative CRUs (Figure 4—figure supplement 1B,C). Future work may address this important issue by simultaneously assessing jSR and RyR arrangement by multi-colour super-resolution imaging, or direct visualization of RyRs and jSR using electron microscopy.

Previous work by Zima et al. has shown that silent, non-spark mediated leak is a significant contributor to overall RyR leak (Zima et al., 2010). We presently show that silent leak can be partly traced to small CRUs, which produce Ca2+ release events that are not detectable experimentally (see also (Walker et al., 2014)). Consistent with previous work (Baddeley et al., 2009; Hayashi et al., 2009; Hou et al., 2015), we observed that many CRUs have very small geometry even in healthy cells; 42% of CRUs contained five or fewer RyRs in Sham (Figure 2C). However, larger CRUs can also contribute to non-spark mediated RyR leak, when Ca2+ release from CRU sub-clusters does not propagate to remaining clusters; so-called ‘quarky’ Ca2+ release (Brochet et al., 2011). Our results show that the decreased fidelity of spark triggering in these dispersed CRUs results from the spacing between neighbouring clusters, which inhibits cooperative, diffusion-based triggering. Furthermore, released Ca2+ more easily escapes from less densely packed CRUs, making it less likely to trigger additional RyRs. This finding is in agreement with previous modeling studies showing that spark fidelity declined when clusters deviated from circular and compact shapes (Walker et al., 2014), and that RyR activation during triggered release is less likely when CRUs are broken into sub-clusters (Cannell and Soeller, 1997). We presently link augmented silent leak in HF to an increased fraction of both small CRUs and CRUs with larger, distributed configurations. These results have important implications. Increased RyR leak during heart failure is widely believed to promote arrhythmia via generation of both early and delayed afterdepolarizations. Furthermore, greater leak also promotes depression of contraction, via loss of SR Ca2+ content, and poor relaxation, due to elevation of resting Ca2+ (reviewed in (Bers, 2014)). The present results provide a structural basis for these maladaptive functional alterations, and suggest that the nanometer scale of these changes prevented their previous detection by lower resolution imaging techniques.

Slowing of Ca2+ release is another key component of the failing phenotype. Prolonged rise time of the Ca2+ transient during HF (Figure 7A–C) has been previously linked to slowed isolated cardiomyocyte contraction in both animals (Bokenes et al., 2008; Mørk et al., 2009) and patients (Davies et al., 1995). In vivo systolic tissue velocity is also reduced in HF patients (Vinereanu et al., 2005) which decreases contractile power. Thus, understanding the mechanisms underlying slowed Ca2+ release in failing cells is critical. Mathematical modeling studies showed that while out-of-focus release events can theoretically reduce spark kinetics artefactually, these events rapidly become undetectable as they are shifted further from the focal plane (Figure 6—figure supplement 1). Furthermore, there is no clear basis for expecting that dispersion of RyR clusters in HF should change the proportion of out-of-focus sparks relative to control, and for this reason we do not expect that variation in the focal plane has systematically impacted our measured rise times. Our modeling results also suggest that an altered number of RyRs/CRU, due either to methodological under-estimation or loss of RyRs from CRUs in HF, will not in and of itself promote slowing of sparks kinetics (Figure 6—source data 1). Indeed, we observed that spark kinetics were relatively insensitive to changes in RyR number for medium-sized CRUs, in agreement with previous work (Cannell et al., 2013). Rather, our results point to an important role of CRU dispersion in slowing Ca2+ sparks, as multi-cluster dyads exhibited progressive triggering of individual clusters by diffusion of released Ca2+ (Figure 6). Importantly, sites with slow spontaneous Ca2+ sparks were observed to also have slow local triggered Ca2+ release during the action potential (Figure 7). Thus, dispersed CRUs promote overall slowing and de-synchronization of the Ca2+ transient. While these results link visible spark-mediated leak to slowing of overall Ca2+ release, small, undetectable RyR openings (silent leak) might also play an important role. Previous work has shown that RyR clustering allows the channels to be functionally coupled, whereby they exhibit coordinated opening and closing (Marx et al., 2001; Wang et al., 2004Sobie et al., 2006). Wang and colleagues reported that this thermodynamic stability is lost when RyRs are present alone or with a small number of neighbouring channels, and slow Ca2+ release kinetics result (Wang et al., 2004). An increased fraction of small CRUs in failing cells may therefore augment slow, but undetectable Ca2+ release events which nevertheless contribute to an overall Ca2+ transient which is slow and de-synchronized.

Dyssynchronous Ca2+ release during heart failure has been previously linked to t-tubule disruption in a large number of studies (reviewed in (Louch et al., 2010)), and we have similarly observed t-tubule disorganization in this post-infarction rat model (Figure 7—figure supplements 1 and Frisk et al., 2016). An important question is therefore whether CRU morphology and slow Ca2+ sparks occur independently from t-tubule reorganization, or whether these two aspects of structural remodeling are related. In previous work, we (Louch et al., 2013) and others (Meethal et al., 2007) employed simultaneous imaging of t-tubules and Ca2+ and observed that sparks occurred almost exclusively at t-tubule sites in both healthy and failing cells. Similarly, de-tubulation experiments have been shown to dramatically reduce the occurrence of Ca2+ sparks in the cell interior, suggesting that SR-t-tubule junctions are important for spark initiation (Brette et al., 2005). It may be postulated, therefore, that CRU dispersion resulting in slow sparks occurs at intact dyads in HF (ie. not at sites with ‘orphaned RyRs’), and that t-tubule and CRU remodeling may occur independently. Verification of this point will likely require simultaneous nanoscale imaging of t-tubules and RyRs, as small degrees of t-tubule drift out of dyads may be critical, and not detected by lower resolution imaging techniques. Regardless, we believe that t-tubule and CRU disruption have additive, detrimental effects, resulting in markedly de-synchronized and slowed SR Ca2+ release.

What signals RyR dispersion in HF? Emerging data indicate that RyRs are not firmly anchored within the CRU, but exhibit a highly malleable organization dependent on factors such as phosphorylation status and cytosolic Mg2+ levels (Asghari et al., 2014). However, while changes in such conditions were shown to influence whether RyRs are positioned in a ‘checkerboard’ or side-by-side arrangement, it is unclear whether they can lead to reorganization of clusters and CRUs on the scale of changes presently observed in HF. Another important dyadic regulator is Junctophilin-2 (JPH-2), which anchors the t-tubule to the SR (Takeshima et al., 2000; Minamisawa et al., 2004), and interacts with the RyR (Beavers et al., 2013Munro et al., 2016). Munro et al recently reported that JPH-2 levels regulate RyR clustering; however, while they observed that JPH-2 overexpression triggered the formation of larger RyR clusters, JPH-2 knockdown did not reduce cluster size (Munro et al., 2016). Although others have observed JPH-2 loss during HF (Minamisawa et al., 2004; Wei et al., 2010), we did not presently observe reduced JP-2 protein levels in our rat HF model (Figure 2—figure supplement 1), suggesting that JP-2 downregulation is not a prerequisite for reorganization of CRUs (Figure 2) or t-tubules (Figure 7—figure supplement 1). Another dyadic regulator, BIN1, is a well-established regulator of t-tubule growth and structure (Hong et al., 2014), and recent data have suggested that this protein may also attract phosphorylated RyRs to the dyad (Fu et al., 2016). Although BIN1 loss has been reported in other HF models (Lyon et al., 2012; Caldwell et al., 2014), our data indicate that BIN1 expression is unaltered in our rat model (Figure 2—figure supplement 1). Thus, we do not believe that BIN1 changes are related to the CRU reorganization presently observed in failing myocytes, and the precise trigger for such changes remains unclear.

In conclusion, our results contribute to an emerging understanding that cardiomyocyte dyads are highly plastic structures. While previous work has shown that t-tubule structure is impressively malleable, and degraded during heart failure, our present findings show that there is also detrimental reorganization of RyRs in this disease. Dispersion of RyRs within the CRU was linked to increased silent RyR leak, slowing of Ca2+ sparks, and de-synchronization of the overall Ca2+ transient, indicating a novel mechanism underlying impaired contractility in HF.

Materials and methods

Key resources table
Reagent type
(species) or resource
DesignationSource or referenceIdentifiersAdditional
information
Strain, strain background
(Rattus norvegicus, M)
Male Wistar-
Hannover rats
Janvier-labsRjHan: WI;
RGD: 13792727
AntibodyMouse-anti-RyR2
primary antibody
ThermoFischer
Scientific
Cat# MA3-916;
RRID:AB_2183054
IHC, 1:100;
WB, 1:1000
AntibodyAlexa Fluo 647
conjugated
goat-anti-mouse secondary Ab
Molecular
Probes/Invitrogen
Cat# A-21237;
RRID:AB_2535806
IHC, 1:200
AntibodyAnti-goat IgG-HRP
linked whole antibody
R and D SystemsCat# HAF109;
RRID:AB_357236
WB, 1:3000
AntibodyMouse IgG HRP
linked Whole Ab
GE HealthcareCat# NA931V;
RRID:AB_772210
WB, 1:3000
AntibodyRabbit IgG HRP
linked Whole Ab
GE HealthcareCat# NA934V;
RRID: AB_772206
WB, 1:3000
 aAntibodyGoat Anti-GAPDH
Polyclonal antibody
Santa Cruz
Biotechnology
Cat# sc-20357;
RRID:AB_641107
WB, 1:500
AntibodyGoat Anti-
Junctophilin-2
Polyclonal Antibody
Santa Cruz
Biotechnology
Cat# sc-51313;
RRID:AB_2296391
WB, 1:1000
AntibodyBIN1
(Amphiphysin
II (2F11)
Antibody)
Santa Cruz
Biotechnology
Cat# sc-23918;
RRID:AB_667901
WB, 1:500
Software,
algorithm
Dense Stochastic
Sampling Imaging
(DSSI) algorithm
SoftWoRx,
GE Healthcare
Software,
algorithm
dSTORM image
post-processing
algorithm
PMID: 26490742Described in the
Github repository
(Kolstad, 2018; copy archived at https://github.com/elifesciences-publications/Ryanodine_Receptor_Dispersion_during_Heart_Failure)
Software,
algorithm
Mathematical
Model
PMID: 22495592;
PMID: 23708355;
this paper
Described in the
Github repository
(Kolstad, 2018; copy archived at https://github.com/elifesciences-publications/Ryanodine_Receptor_Dispersion_during_Heart_Failure)
Software,
algorithm
SigmaPlotSigmaPlotRRID:SCR_003210

Rat model of post-myocardial infarction congestive HF

All experiments were approved by the Norwegian National Animal Research Authority (project license no. FOTS 5982, 7786), and were performed in accordance with the National Institute of Health guidelines (NIH publication No. 85 – 23, revised 2011) and European Directive 2010/63/EU. Large anterolateral myocardial infarctions were induced in ~300 g male Wistar-Hannover rats, by ligation of the left coronary artery as previously described (Lunde et al., 2012). Development of HF was verified six weeks later using a Vevo 2100 echocardiography imaging system (VisualSonics, Toronto, Canada). Inclusion of failing animals was based on established criteria (Sjaastad et al., 2000), including dilation of the left atrium (diameter >5 mm) and ventricle, and increased lung weight (>2.5 g). Sham-operated rats served as controls. Experiments were performed over a two year period, using animals from 10 rounds of animal surgery. Sample sizes were determined by power analysis, assuming that only 50% of post-infarction animals would be included in the final data set, and based on a pilot project of variability in CRU morphology in healthy controls.

Cell isolation

Cardiac myocytes from failing and Sham-operated rats were isolated using a standard enzymatic dispersion technique (Louch et al., 2011). Excised hearts were mounted on a Langendorff setup, and retrogradely perfused through the aorta with Ca2+-free solution containing (in mmol/L): 130 NaCl, 25 Hepes, 5.4 KCl, 0.5 MgCl2, 0.4 NaH2PO4, 5.5 D-glucose, pH 7.4. Once cleared of blood, hearts were then perfused with the above solution including collagenase (2 mg/mL, Worthington Biochemical Corp., Lakewood, NJ, USA) and low [Ca2+] (0.05 mmol/L). After 10 min of digestion, hearts were cut down, minced, and filtered, and isolated cardiomyocytes were allowed to sediment.

Immunostaining

Isolated cardiomyocytes were transferred to cell culture medium (DMEM 1X, Life Technologies with 10% FBS, Biowest Nuaillé, France and 1% Penicillin-Streptomycin, Sigma), and plated on laminin-coated, glass bottom culture dishes (MatTek corporation, Ashland MA). Staining was performed according to a described protocol (Swift et al., 2007), with consecutive steps for chemical fixation (4% Formaldehyde in 1 mol/L HEPES buffer, 10 min), quenching (PBS + 100 mmol/L Glycine, 10 min), permeabilization (PBS + 0,03% Triton X-100, 10 min), and blocking (NaCl 150 mmol/L, Na3 citrate 17.5 mmol/L, 5% goat serum, 3% BSA, 0.02% NaN3, 2 hr). PBS washing was performed in between each step. The cells were then incubated overnight with 1/100 diluted mouse-anti-RyR2 primary antibody (ThermoFischer Scientific, MA3–916) in low blocking buffer, containing 150 mmol/L NaCl, 17.5 mmol/L Na3 citrate, 2% goat serum, 1% BSA, and 0.02% NaN3 at 4°C. This protocol has previously been reported to result in the binding of multiple primary antibodies to each RyR tetramer (Baddeley et al., 2009). The following day, cells were washed with PBS and incubated with 1/200 diluted secondary antibody (Alexa Fluo 647 conjugated goat-anti-mouse secondary Ab, Molecular Probes/Invitrogen) in low blocking buffer for 2 hr. Cells were then washed and stored in PBS until image acquisition. Of note, the fab-fragment secondary antibody employed places the fluorescent label far closer to the epitope than traditional antibodies. Thus, under our experimental conditions, the steric error is generally <10 nm, and dwarfed by the localization of the dSTORM technique (≈20 nm, see below). dSTORM imaging and analysis dSTORM imaging was performed using an OMX V4 system (Applied Precision, GE Healthcare) with a 60 × 1.49 NA TIRF objective (Olympus), a pco.edge sCMOS camera (PCO), a 100 mW 642 nm laser, and a 683/40 emission filter. Focusing was performed with a 30V300nanoX CL focusing unit (Piezosystem, Jena). Cells were placed in ‘switching buffer’ (0.5 mg/mL glucose oxidase, 40 μg/mL catalase, 10% wt/vol glucose, 50 mmol/L β-mercaptoethylamine in Tris-buffer, pH 8.0, all Sigma–Aldrich), and fluorophores were pushed into the dark state by illumination with the 642 nm laser at a highly inclined, but sub-TIRF angle (Highly Inclined and Laminated Optical sheet, HILO (Tokunaga et al., 2008)). Spontaneous blinking occurred without the use of an activation laser, and was recorded at a depth of 200–500 nm during ten-thousand frames per field of view (20.48 × 20.48 µm), with a maximum of 350,000 blinks recorded.

Data were processed with built-in software (softWoRx, GE Healthcare) using a Dense Stochastic Sampling Imaging (DSSI) algorithm and multiple Gaussian fits to localize events. Drift correction was performed with a model-based algorithm. Average localization precision was 21.6 nm for the events included in the final reconstructions.

Images were further processed using a custom analysis program written in Python, which was similar to one previously employing scikits-image, scipy.spatial and Opencv (http://opencv.willowgarage.com/) modules (Macquaide et al., 2015). These algorithms are publicly available in an online repository, (https://github.com/TerjePrivate/Ryanodine_Receptor_Dispersion_during_Heart_Failure) (Kolstad, 2018; copy archived at https://github.com/elifesciences-publications/Ryanodine_Receptor_Dispersion_during_Heart_Failure). Images with 10 × 10 nm pixels were convolved with a 2D Gaussian function equal to the calculated resolution of the image (~20 nm), and downscaled by a factor of 3 to produce a final pixel size of 30 × 30 nm. RyR locations were defined using a modified automated thresholding algorithm (Kolstad, 2018) (Otsu method), excluding the brightest 0.3% of the signal. This prevented skewing of the threshold by regions which were constantly in the active state. To minimize inclusion of autofluorescence artefacts, blinks appearing in 10 or more consecutive frames were excluded from the final reconstruction. Acquired images were fitted to a 30 × 30 nm grid corresponding to the quatrefoil structure of the RyR protein (Baddeley et al., 2009). An RyR was counted as present if >half the area of a 30 nm square was above threshold, and RyR clusters were defined by occupied, neighbouring grid positions. Based on previous (Macquaide et al., 2015) and present (Figure 4—figure supplement 1) calculations clusters with edge-to-edge distance <150 nm were assumed to cooperatively generate Ca2+ sparks, and grouped into CRUs accordingly. A stricter CRU definition, with edge-to-edge distances < 100 nm, (Baddeley et al., 2009; Hou et al., 2015) was also examined. Inter-cluster distances were calculated from the centroid of each cluster. CRU solidity was calculated as the proportion of the bounding polygon (convex hull method) which was filled with RyRs; clusters containing less than 5 RyRs were excluded from this calculation. The solidity ratio was 1 if totally filled and 0 if completely empty; therefore, lower values indicate greater CRU fragmentation. Of note, all analyses of RyR localization were performed by automated protocols in a blinded manner.

To address whether unspecific secondary antibody binding affected measurements of RyR configurations, dSTORM imaging of cardiomyocytes was performed in the absence of primary antibody. The obtained signal was then added to RyR-labeled images obtained by the standard protocol (primary plus secondary antibodies), and RyR configuration was analyzed. Non-specific labeling was observed to only negligibly increase the number of RyRs/cluster and RyRs/CRU by 3.7% and 3.2%, respectively. Similarly, RyR density was increased by 0.5%, and inter-cluster distance was reduced by 0.2%, supporting that unspecific labelling had a minute influence on the dataset.

Ca2+ spark and transient imaging and analysis

Using an LSM 7Live confocal microscope (Zeiss), Ca2+ sparks were recorded from quiescent cardiomyocytes loaded with fluo-4 AM (20 µmol/L, Molecular Probes, Eugene, OR) and superfused with a HEPES Tyrode solution containing (in mmol/L): 140 NaCl, 1.0 CaCl2, 0.5 MgCl2, 5.0 HEPES, 5.5 glucose, 0.4 NaH2PO4, 5.4 KCl, pH 7.4, 37°C. Scans were performed with a 1024 pixel line drawn along the longitudinal axis of the cell with a temporal resolution of 1.5 ms. Ca2+ sparks were analysed with a custom program (CaSparks 1.01, D. Ursu, 2003), as previously (Louch et al., 2013). Sparks were defined as local increases in fluorescence with a minimum amplitude (ΔF/F0) of 0.4, to minimise the inclusion of false positives. Linescan images of cells obtained during inhibition of Ca2+ sparks (prolonged exposure to 10 mM caffeine) confirmed the appropriateness of this detection threshold. Ca2+ spark frequency was normalized to cell length and recording time, and spark geometry was assessed by measurements of time to peak (TTP), full width at half maximum (FWHM), and full duration at half maximum (FDHM). Spark-mediated Ca2+ leak was calculated as the product of spark mass (amplitude X FWHM X FDHM) and frequency.

Ca2+ transients were elicited by field-stimulation through a pair of platinum wires (3 ms supra-threshold current pulses at 1 Hz), and recorded as confocal linescans under the same experimental conditions as Ca2+ spark measurements. Global Ca2+ transient characteristics were analyzed by averaging the Ca2+ signal along the linescan, with measurements of transient magnitude (normalized to resting fluorescence, F/F0), time to half maximal fluorescence (TTF50), and TTP. Local Ca2+ transients were averaged across narrow 2 µm bands of the linescan. Synchrony of Ca2+ release was assessed as previously described, (Louch et al., 2010) by plotting the profile of TTF50 measurements across the cell and measuring the standard deviation of these values (the ‘dyssynchrony index’).

SR Ca2+ content was assessed by rapidly applying 10 mM caffeine and measuring the amplitude of the elicited Ca2+ transient.

Microsomal Ca2+ uptake, leak and release

Ca2+ handling was additionally examined using crude homogenates from rat left ventricle, based on methods described by O’Brien and modified by Li et al. (O'Brien, 1990; Li et al., 2002). Fresh ventricular tissue was weighed and homogenized in ice cold buffer (1:10 wet weight/vol, pH 7.9) containing (in mmol/L): 300 sucrose, 5 NaN3, 1 EDTA, 40 L-histidine, 40 Tris HCl and protease inhibitors. Homogenization was performed with a Polytron 1200 (Kinematica AG, Luzern, Switzerland) at 25000 rpm for 3 × 20 s, with a 20 s break between bursts. Homogenates were then aliquoted, frozen in liquid N2, and stored at −80°C until use.

Ca2+ uptake and release were measured in 2.2 ml of assay buffer, containing (in mmol/L): 165 KCl, 22 Hepes, 7.5 oxalate, 11 NaN3, 0.0055 TPEN, 4.5 MgCl2, 9 Tris HCl and 0.002 fura-2 salt (pH = 7.0, 37°C). Ca2+ fluxes were monitored with an LS50B luminescence spectrometer (Perkin Elmer Ltd, Beaconsfield, Buckinghamshire, United Kingdom) after addition of 100 µl of freshly-thawed and vortexed homogenate. Ca2+ uptake by the vesicles was initiated by addition of Na4ATP (2.2 mmol/L), and then blocked by application of thapsigargin (1.5 µmol/L) to assess RyR leak. Releasable SR Ca2+ content was estimated by measuring Ca2+ release induced by application of the RyR opener 4-chloro-m-cresol (4-CMC) (5.5 mmol/L). The fluorescence ratio was calibrated to [Ca2+] using the following equation: [Ca2+]=Kd *((R - Rmin)/(Rmax - R))*(Sf2/Sb2), where R is the 340 nm/380 nm fluorescence ratio, Kd is the dissociation constant of fura-2 and Sf2/Sb2 is the ratio of measured fluorescence intensity at 380 nm when fura-2 is Ca2+ free or saturated, respectively. Rmin is the ratio at very low [Ca2+]i and Rmax is the ratio at saturating [Ca2+]i, obtained by adding 3.3 mmol/L EGTA and 4.8 mmol/L CaCl2 respectively to the cuvette at the end of each.

Western blotting

Frozen tissue from rat left ventricles was homogenized in cold buffer (210 mM sucrose, 2 mM EGTA, 40 mM NaCl, 30 mM HEPES, 5 mM EDTA) with the addition of a Complete EDTA free protease inhibitor cocktail tablet (Roche Diagnostics, Oslo, Norway) and a PhosSTOP tablet (Roche). SDS was then added to the homogenates to a final concentration of 1%, and protein concentrations were quantified using a micro BCA protein assay kit (Thermo Fischer Scientific Inc., Rockford, IL). Bovine serum albumin (BSA) was used as standard protein.

The following primary antibodies were employed for immunoblotting: RyR (1:1000; MA3-916, Thermo Scientific), BIN1 (1:500; sc23918, Santa Cruz Biotechnology), junctophilin-2 (1:1000; sc-51313, Santa Cruz Biotechnology) and GAPDH (1:500; sc-20357, Santa Cruz Biotechnology). Secondary antibodies were anti-rabbit (NA934V, GE Healthcare), anti-mouse (NA931V, GE Healthcare) or anti-goat (HAF109, R and D Systems) IgG-HRP linked whole antibody. Data were normalized to GAPDH and then to Sham values (Figure 2—figure supplement 1).

Protein homogenates (5 or 15 µg/lane) were size fractionated on 4–15% or 15% Criterion TGX gels (Biorad Laboratories, Oslo, Norway) and transferred to 0.45 μM PVDF-membranes (GE Healthcare). The membranes were blocked in 5% non-fat milk or 5% Casein (Roche Diagnostics) in Tris-buffered saline with 0.1% Tween (TBS-T) for 1 hr at room temperature, and then incubated with primary antibody overnight at 4°C. Secondary antibodies were incubated for 1 hr at room temperature and blots were developed using Enhanced Chemiluminescence (ECL prime, GE healthcare). Chemiluminiscense signals were detected by a LAS 4000 (GE healthcare) and protein levels were quantified using ImageQuant software (GE Healthcare).

Mathematical model

A mathematical model was created to simulate the effects of varied RyR localization and CRU geometry on Ca2+ spark characteristics. We have made all simulation results, geometries, and code specific to this study available in an online repository (Kolstad, 2018), along with code for the full reaction-diffusion simulator. The model extended from the work of Hake et al. (2012), with included RyR stochasticity developed from previous work by Cannell and colleagues (Cannell et al., 2013). We have chosen this model for the relative simplicity of its gating (no direct inter-RyR coupling, or explicit luminal Ca2+ regulation), and because it was built for a similarly constructed (spatially discretized) geometry, (Cannell et al., 2013) unlike most other recent RyR2 gating (Williams et al., 2011; Wescott et al., 2016). A set of coupled partial differential equations was employed to describe the temporal evolution of the free and bound [Ca2+] in the SR and cytosol:

ct=Dc2ci=14Ri(c,bi), xΩC
bit=Di2bi+Ri(c,bi), i=1,2,3,4 xΩC
st=Ds2sR5(s,b5), xΩS
b5t=R5(s,b5), xΩS

Here ΩC is this cytosolic domain, including the cleft space, and ΩS is the SR, including both junctional and network SR components. Four buffers were included in ΩC: ATP, calmodulin, troponin and Fluo-4, and one buffer, calsequestrin, was included in ΩS. These buffers are numbered from 1 to 5 and their corresponding concentrations are denoted bi. Troponin and calsequestrin were regarded to be stationary, and the corresponding diffusion coefficients (σ) were therefore set to zero (see Supplementary file 1). The calcium concentrations in ΩC and ΩS are denoted c and s respectively.

The buffering reactions are of the form

Rc,bi= koncBtot-bi-koffbi

where Btot is the total buffer concentration, and kon and koff are the on and off rates for the buffer, respectively.

The two domains are coupled through a flux condition over the SR membrane:

Dccn=Dssn=J(c,s)

where

J(c,s)={JRyRxΓRyRJSercaxΓSerca0elsewhere

The RyR flux is computed by:

JRyR(c,s)=gRyRγ(cs)

where γ0,1 is a stochastic variable that switches between the conductive (O) and non-conductive (C) states according to:

k+CO.k-

While we have chosen a different form for the equations expressing the default transition rates, they are equivalent to those in the original model of Cannell et al. (2013) with the exception that we have set limits to both k+ and k- at low dyadic calcium:

k+c=fcK+n+, kmin+, kmax+
k-c=fcK-n-, kmin-, kmax-

where for a<b:

f(y,a,b)={a, if y<ay, if aybb, if y>b

The parameters are given in Supplementary file 1.

The SERCA formulation is taken from Tran et al. (2009) and is of the form:

JSerca(c,s)=a1c2a2s2a3c2+a4s2+a5.

Geometry, numerics, and implementation

The computational domain in our model (ΩSΩc) was a (2 µm) (Marx et al., 2001) cube containing a single CRU (Illustrated in Figure 4—figure supplement 1A). Unlike the original work of Cannell and colleagues, for which RyR locations were fixed for all simulations, our simulations involve algorithm-defined changes in the jSR geometry and location of RyRs to reflect the structural differences captured by the dSTORM recordings. In all geometries the domain consisted of a 12 nm wide cleft space sandwiched between the junctional SR (jSR) surface and the t-tubular surface. The latter was represented as a non-conductive slab inside Ωc, serving as a barrier to diffusion. RyRs were located on the opposing jSR surface, each occupying a space 36 × 36 nm, with neighbouring RyRs placed 36 nm apart (centre-to-centre distance). The jSR was modeled as a physical extension (‘padding’) around each RyR by a defined distance equivalent to 1 RyR diameter (36 nm). As mentioned above, RyRs were arranged according to dSTORM-derived locations, and the jSR shape was adjusted according to the RyR locations. In some simulations (Figure 4A, Figure 6—source data 1) the geometries of the jSR were fixed while idealized RyR lattice geometries were varied to explore the independent effect of modifying RyR number with fixed jSR volume and the locally releasable calcium pool (Figure 4A). To specifically investigate the effects of RyR dispersal, example dSTORM-identified RyR patterns were selected with similar total RyR number but with arrangement into a varied number of sub-clusters (1, 3, 7, or 10). The ratio of RyR number to jSR volume for these example CRUs is presented in Figure 4—source data 1. To limit the effect variation in jSR volume has on the releaseable Ca2+ store, we fixed the total SR volume (and thus initial SR Ca2+ content) across all geometries by modifying the non-junctional SR (nSR) volume as required. Of note, while nSR concentration was modeled as a continuum, global SR calcium concentration was effectively clamped at initial values as expected for the time-scale and spatial-scale simulated.

Dyadic Ca2+ release was initiated by opening a single RyR in the CRU; this ‘trigger’ RyR was selected randomly and varied between consecutive simulations. Simulated triggered Ca2+ release from the CRU was then allowed to proceed, with the above equations discretized in space with a finite volume approach (12 nm edge length throughout the domain), and solved in time using explicit Euler time stepping. Specifically, we used operator splitting and solved each of the reaction and diffusion sub-problems with a fixed ∆t = 0.1µs, except when calculating the RyR release current. Due to the small element volumes and high fluxes this calculation is very stiff, so instead we solve it analytically. If we use x to denote the SR calcium concentration and y to denote the cleft calcium concentration, we can write this sub-problem as:

x˙=K(yx)
y˙=K(xy)

where K is the channel conductance per element volume. The solution to this subsystem is given by:

x(t)=SDe2Kt
yt=S+De-2Kt

Where:

S=(y0+x(0))2
D=(y0-x(0))2

Using this scheme it is possible to take arbitrarily long time steps without introducing instabilities.

For the RyR gating model half-maximal activation was achieved at 80 µM Ca2+, which allowed cooperative opening of adjacent RyR clusters located up to ≈150 nm apart (ie. 4 RyR lengths) if the clusters shared jSR (Figure 4—figure supplement 1B,C). Thus, we defined CRUs as groupings of RyR clusters with edge-to-edge distances < 150 nm, in agreement with recent work (Macquaide et al., 2015), but also compared data with a stricter CRU definition (cluster distances < 100 nm) employed in other publications (Baddeley et al., 2009; Hou et al., 2015). RyR rates are shown in Supplementary file 2.

Statistical analyses

All results are expressed as mean values ± standard error of the mean. All statistical significance was calculated in SigmaPlot by Student’s t-test or ANOVA with Bonferroni post-hoc comparison for normally distributed data, as appropriate. Skewed distributions of experimental and modelled Ca2+ spark parameters were respectively assessed by the nonparametric Mann-Whitney Rank Sum Test and Kruskal-Wallis ANOVA with Dunn’s test for post-hoc comparisons. dSTORM-based measurements of RyR geometries were compared with averages taken both across cells and animals, with respective statistical testing by t-tests and linear mixed effects models (Lindstrom and Bates, 1988). Statistical significance was defined as p<0.05.

Access to raw experimental data and analysis source code

All raw data acquired and analyzed in this study are publicly available at https://github.com/TerjePrivate/Ryanodine_Receptor_Dispersion_during_Heart_Failure (Kolstad, 2018; copy archived at https://github.com/elifesciences-publications/Ryanodine_Receptor_Dispersion_during_Heart_Failure).

References

  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
  7. 7
  8. 8
  9. 9
  10. 10
  11. 11
  12. 12
  13. 13
  14. 14
  15. 15
  16. 16
  17. 17
  18. 18
  19. 19
  20. 20
  21. 21
  22. 22
  23. 23
  24. 24
  25. 25
  26. 26
  27. 27
    Newton—Raphson and EM Algorithms for Linear Mixed-Effects Models for Repeated-Measures Data
    1. MJ Lindstrom
    2. DM Bates
    (1988)
    Journal of the American Statistical Association 83:1014–1022.
  28. 28
  29. 29
  30. 30
  31. 31
  32. 32
  33. 33
  34. 34
  35. 35
  36. 36
  37. 37
  38. 38
  39. 39
  40. 40
  41. 41
  42. 42
  43. 43
  44. 44
  45. 45
  46. 46
  47. 47
  48. 48
    Junctophilins: a novel family of junctional membrane complex proteins
    1. H Takeshima
    2. S Komazaki
    3. M Nishi
    4. M Iino
    5. K Kangawa
    (2000)
    Molecular Cell 6:11–22.
  49. 49
  50. 50
  51. 51
  52. 52
  53. 53
  54. 54
  55. 55
  56. 56
  57. 57

Decision letter

  1. Harry C Dietz
    Senior Editor; Howard Hughes Medical Institute and Institute of Genetic Medicine, Johns Hopkins University School of Medicine, United States
  2. Gordana Vunjak-Novakovic
    Reviewing Editor; Columbia University, United States

In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.

Thank you for submitting your article "Ryanodine Receptor Dispersion Disrupts Ca2+ Release in Failing Cardiac Myocytes" for consideration by eLife. Your article has been reviewed by Harry Dietz as the Senior Editor, a Reviewing Editor, and three reviewers. The following individual involved in review of your submission has agreed to reveal his identity: David Baddeley (Reviewer #3).

The reviewers have discussed the reviews with one another and the Reviewing Editor has drafted this decision to help you prepare a revised submission.

Summary:

This is an interesting and thoughtful manuscript that advances our understanding of the changes underlying the slower and weaker calcium transients in failing cardiac myocytes. The authors describe changes in the arrangement of ryanodine receptors, using super-resolution microscopy and simulations. The combination of structural cluster imaging, functional Ca2+ imaging and simulations has been used effectively, and the results are convincing. The main finding of the study is that RyR clusters are more dispersed in the failing than in normal cardiomyocytes.

Essential revisions:

1) Cluster fragmentation data (Figure 2) need to be shown more clearly and in greater detail. Some of the magnified unthresholded data and the scaled raw image data should be displayed (taking care not to saturate too much). A clear case should be made that the fragmentation is not a thresholding artifact, as it is known that threshold changes could cause apparent fragmentation for a given data set.

2) It is critical that the authors provide enough implementation detail for the simulation model so that a reader can use the simulation for stably dealing with the nonlinearities contained in a stochastic RyR switching model. If a custom code was used, it should be provided.

It should be clarified if the basic geometry is the same as in the Laver et al. 'induction decay' model, or if not – what is the scope and rationale for the changes made. Also, the authors should explain how they performed the numerical integration (was the time step fixed or adaptive; how were the sudden changes in transition rates incorporated in to the discretized solver?) The authors should also provide rationale for the data exclusions and clearly specify which data were used.

3) The method is based on the assumption that peripheral couplons are effectively 2D objects parallel to the cell surface and hence the imaging plane. It becomes increasingly hard to justify this assumption when moving away from surface where EM data shows that couplons take on complex 3D geometries, often wrapping around t-tubules. The absence of double-rows would imply that the images taken in this paper are not at the cell surface. While looking at the projected area will still give a reasonable estimate of relative numbers, it is likely that the absolute number of RyRs is underestimated for internal couplons. Regarding the projections to a regular 30nm grid, there is very limited evidence that RyRs form a regular grid in normal conditions within cardiomyocytes. Even if they do form a regular grid, it likely does not align with the imposed sub-sampled pixel grid. The authors should specifically discuss these important aspects of their model. It is also suggested to acknowledge that the estimated numbers are not necessarily very precise and discuss the effect that changing the RyR number would have on the simulation results.

4) With regard to the measured parameters, simple t-tests do not suffice. A hierarchical nested model should be used as p<0.05 is not satisfactory when there is large inter-group variability and the number of animals is small. The data needs to be reanalyzed, e.g. using methods described in a recent publication (Sikkel et al., 2017).

https://doi.org/10.7554/eLife.39427.021

Author response

Essential revisions:

1) Cluster fragmentation data (Figure 2) need to be shown more clearly and in greater detail. Some of the magnified unthresholded data and the scaled raw image data should be displayed (taking care not to saturate too much). A clear case should be made that the fragmentation is not a thresholding artifact, as it is known that threshold changes could cause apparent fragmentation for a given data set.

Thank you for your helpful comments, which have improved the manuscript. Figure 2 has now been changed to include a new panel which displays a high magnification “high-low/grayscale” image of the region of interest. This has been done to show that the binarized images give a reasonable estimation of RyR clusters based on the raw images for both Sham and HF groups. The added panel has been given a “high-low” look-up table wherein the top and bottom 3% grey value pixels are colored red and blue respectively. This more clearly illustrates that the grayscale images from which the binarization was derived are not heavily saturated.

2) It is critical that the authors provide enough implementation detail for the simulation model so that a reader can use the simulation for stably dealing with the nonlinearities contained in a stochastic RyR switching model. If a custom code was used, it should be provided.

It should be clarified if the basic geometry is the same as in the Laver et al. 'induction decay' model, or if not – what is the scope and rationale for the changes made. Also, the authors should explain how they performed the numerical integration (was the time step fixed or adaptive; how were the sudden changes in transition rates incorporated in to the discretized solver?) The authors should also provide rationale for the data exclusions and clearly specify which data were used.

We have now elaborated on each of these points in the Materials and methods section of the manuscript. For the reviewers' reference, the time step was fixed at 0.1 microseconds, which is sufficient for the RyR gating events themselves, and is not limiting for solution of the resulting concentration changes on either side of the jSR membrane, because this sub-problem can be solved analytically.

3) The method is based on the assumption that peripheral couplons are effectively 2D objects parallel to the cell surface and hence the imaging plane. It becomes increasingly hard to justify this assumption when moving away from surface where EM data shows that couplons take on complex 3D geometries, often wrapping around t-tubules. The absence of double-rows would imply that the images taken in this paper are not at the cell surface. While looking at the projected area will still give a reasonable estimate of relative numbers, it is likely that the absolute number of RyRs is underestimated for internal couplons. Regarding the projections to a regular 30nm grid, there is very limited evidence that RyRs form a regular grid in normal conditions within cardiomyocytes. Even if they do form a regular grid, it likely does not align with the imposed sub-sampled pixel grid. The authors should specifically discuss these important aspects of their model. It is also suggested to acknowledge that the estimated numbers are not necessarily very precise and discuss the effect that changing the RyR number would have on the simulation results.

The reviewers are correct that the RyR clusters imaged in our work are near the surface, but not at the surface. We agree that using grid-based counting methods for determining RyR numbers at internal locations is likely more error prone and may underestimate cluster sizes. The following section has been added to the Discussion section to address these points:

“However, it should be noted that the grid-based method for RyR counting assumes that the channels lie parallel to the field of view, with a uniform, grid-like configuration; presumptions which are likely less valid when RyRs are visualized internally than at the cell surface. Due to super-imposition of internal RyRs along the z axis, it is therefore likely that present RyR cluster sizes are somewhat underestimated.”

We have made a similar statement in the Discussion section regarding estimates of RyRs / CRU:

“Despite possible underestimation of RyR numbers / CRU due to methodological issues noted above, these estimates are in relatively close agreement with an electron microscopy tomography study of mouse ventricular cardiomyocytes (Hayashi et al., 2009)”.

We have additionally considered the reviewer’s point that underestimation of RyR cluster and CRU sizes by grid-based counting procedures may influence the simulation results. To this end, we have performed additional modeling where RyR number was varied, while maintaining the ratio of jSR to nSR. These new data (Figure 6—source data 1) showed that for CRUs larger than 20 RyRs there was little effect of varying RyR number on Ca2+ spark kinetics. Mean CRU sizes observed experimentally in both Sham and HF cardiomyocytes were within this range (indicated as dotted vertical lines in Figure 6—source data 1). Furthermore, since RyR numbers were likely underestimated experimentally, we do not expect that such counting errors have marked consequences regarding the effects of CRU dispersal on Ca2+ spark parameters. Of note, a lack of sensitivity of spark characteristics to RyR number in average-sized clusters is in agreement with previous work (Cannell et al., 2013). This finding has been described in the Results section:

“Of note, although CRUs were observed to contain fewer RyRs in HF than Sham (Figure 2C), simply reducing the RyR number to an equivalent degree in the mathematical model did not markedly alter Ca2+ spark kinetics (Figure 6—source data 1), further confirming a key role of CRU fragmentation in failing cells.”

We have also addressed this point in the Discussion section:

“Our modeling results also suggest that an altered number of RyRs / CRU, due either to methodological under-estimation or loss of RyRs from CRUs in HF, will not in and of itself promote slowing of sparks kinetics (Figure 6—source data 1). Indeed, we observed that spark kinetics were relatively insensitive to changes in RyR number for medium-sized CRUs, in agreement with previous work (Cannell et al., 2013).”

4) With regard to the measured parameters, simple t-tests do not suffice. A hierarchical nested model should be used as p<0.05 is not satisfactory when there is large inter-group variability and the number of animals is small. The data needs to be reanalyzed, e.g. using methods described in a recent publication (Sikkel et al., 2017).

A hierarchical “nested ANOVA” linear mixed-effects model has been applied to the dataset, which allows RyR configuration to be cumulated and compared within animals. The resulting p-values can be found in a table included in Figure 2—source data 1 and are presented alongside data cumulated within cells. The main differences in RyR configuration between Sham and HF were similar based on these two approaches.

https://doi.org/10.7554/eLife.39427.022

Article and author information

Author details

  1. Terje R Kolstad

    1. Institute for Experimental Medical Research, Oslo University Hospital and University of Oslo, Oslo, Norway
    2. KG Jebsen Center for Cardiac Research, University of Oslo, Oslo, Norway
    Contribution
    Conceptualization, Data curation, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing—original draft, Writing—review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-0589-5689
  2. Jonas van den Brink

    Simula Research Laboratory, Fornebu, Norway
    Contribution
    Conceptualization, Software, Formal analysis, Investigation, Visualization, Methodology, Writing—original draft, Writing—review and editing
    Competing interests
    No competing interests declared
  3. Niall MacQuaide

    Institute of Cardiovascular Sciences, University of Glasgow, Glasgow, Great Britain
    Contribution
    Conceptualization, Software, Formal analysis, Investigation, Visualization, Writing—original draft, Writing—review and editing
    Competing interests
    No competing interests declared
  4. Per Kristian Lunde

    Institute for Experimental Medical Research, Oslo University Hospital and University of Oslo, Oslo, Norway
    Contribution
    Formal analysis, Investigation, Methodology, Writing—original draft, Writing—review and editing
    Competing interests
    No competing interests declared
  5. Michael Frisk

    1. Institute for Experimental Medical Research, Oslo University Hospital and University of Oslo, Oslo, Norway
    2. KG Jebsen Center for Cardiac Research, University of Oslo, Oslo, Norway
    Contribution
    Investigation, Visualization, Writing—original draft, Writing—review and editing
    Competing interests
    No competing interests declared
  6. Jan Magnus Aronsen

    1. Institute for Experimental Medical Research, Oslo University Hospital and University of Oslo, Oslo, Norway
    2. Bjørknes College, Oslo, Norway
    Contribution
    Investigation, Methodology, Writing—original draft, Writing—review and editing
    Competing interests
    No competing interests declared
  7. Einar S Norden

    1. Institute for Experimental Medical Research, Oslo University Hospital and University of Oslo, Oslo, Norway
    2. Bjørknes College, Oslo, Norway
    Contribution
    Formal analysis, Methodology, Writing—original draft, Writing—review and editing
    Competing interests
    No competing interests declared
  8. Alessandro Cataliotti

    1. Institute for Experimental Medical Research, Oslo University Hospital and University of Oslo, Oslo, Norway
    2. KG Jebsen Center for Cardiac Research, University of Oslo, Oslo, Norway
    Contribution
    Resources, Supervision, Writing—original draft, Writing—review and editing
    Competing interests
    No competing interests declared
  9. Ivar Sjaastad

    1. Institute for Experimental Medical Research, Oslo University Hospital and University of Oslo, Oslo, Norway
    2. KG Jebsen Center for Cardiac Research, University of Oslo, Oslo, Norway
    Contribution
    Conceptualization, Resources, Supervision, Methodology, Writing—original draft, Writing—review and editing
    Competing interests
    No competing interests declared
  10. Ole M Sejersted

    Institute for Experimental Medical Research, Oslo University Hospital and University of Oslo, Oslo, Norway
    Contribution
    Resources, Supervision, Writing—original draft, Writing—review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-8817-3296
  11. Andrew G Edwards

    Simula Research Laboratory, Fornebu, Norway
    Contribution
    Conceptualization, Formal analysis, Investigation, Visualization, Methodology, Writing—original draft, Writing—review and editing
    Competing interests
    No competing interests declared
  12. Glenn Terje Lines

    Simula Research Laboratory, Fornebu, Norway
    Contribution
    Conceptualization, Software, Formal analysis, Investigation, Methodology, Writing—original draft, Writing—review and editing
    Competing interests
    No competing interests declared
  13. William E Louch

    1. Institute for Experimental Medical Research, Oslo University Hospital and University of Oslo, Oslo, Norway
    2. KG Jebsen Center for Cardiac Research, University of Oslo, Oslo, Norway
    Contribution
    Conceptualization, Resources, Data curation, Formal analysis, Supervision, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing—original draft, Project administration, Writing—review and editing
    For correspondence
    w.e.louch@medisin.uio.no
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-0511-6112

Funding

Horizon 2020 Framework Programme (Consolidator grant for WEL 647714)

  • Terje R Kolstad
  • William E Louch

Helse Sør-Øst RHF (2012072)

  • Terje R Kolstad
  • Ole M Sejersted
  • William E Louch

Helse Sør-Øst RHF (2013014)

  • Terje R Kolstad
  • Ole M Sejersted
  • William E Louch

Oslo University Hospital Ullevål

  • Terje R Kolstad
  • Ole M Sejersted
  • William E Louch

University of Oslo

  • William E Louch
  • Terje R Kolstad

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

The authors thank the Section of Comparative Medicine, Oslo University Hospital Ullevål for animal care and Dina Behmen and Almira Hasic at the Institute for Experimental Medical Research, Oslo University Hospital Ullevål for assistance with Western blotting. We additionally thank Vigdis Sørensen at the Advanced Light Microscopy Core Facility, Oslo University Hospital, Montebello, for assistance with dSTORM imaging.

Ethics

Animal experimentation: All experiments were approved by the Norwegian National Animal Research Authority (project license no. FOTS 5982, 7786), and were performed in accordance with the National Institute of Health guidelines (NIH publication No. 85-23, revised 2011) and European Directive 2010/63/EU.

Senior Editor

  1. Harry C Dietz, Howard Hughes Medical Institute and Institute of Genetic Medicine, Johns Hopkins University School of Medicine, United States

Reviewing Editor

  1. Gordana Vunjak-Novakovic, Columbia University, United States

Publication history

  1. Received: June 21, 2018
  2. Accepted: October 24, 2018
  3. Accepted Manuscript published: October 30, 2018 (version 1)
  4. Version of Record published: November 20, 2018 (version 2)

Copyright

© 2018, Kolstad et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 535
    Page views
  • 104
    Downloads
  • 0
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

  1. Further reading

Further reading

    1. Cell Biology
    2. Neuroscience
    Manuela D'Alessandro et al.
    Research Article Updated
    1. Cell Biology
    2. Neuroscience
    Nickolai V Vysokov et al.
    Research Article