Aligned Organization of Synapses and Mitochondria in Auditory Hair Cells

Abstract

Recent studies have revealed great functional and structural heterogeneity in the ribbon-type synapses at the basolateral pole of the isopotential inner hair cell (IHC). This feature is believed to be critical for audition over a wide dynamic range, but whether the spatial gradient of ribbon morphology is fine-tuned in each IHC and how the mitochondrial network is organized to meet local energy demands of synaptic transmission remain unclear. By means of three-dimensional electron microscopy and artificial intelligence-based algorithms, we demonstrated the cell-wide structural quantification of ribbons and mitochondria in mature mid-cochlear IHCs of mice. We found that adjacent IHCs in staggered pairs differ substantially in cell body shape and ribbon morphology gradient as well as mitochondrial organization. Moreover, our analysis argues for a location-specific arrangement of correlated ribbon and mitochondrial function at the basolateral IHC pole.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12264-021-00801-w.

Introduction

In the mammalian cochlea, auditory perception relies on a highly specialized and functionally compartmentalized cell type called the inner hair cell (IHC). At the apical surface of the IHC, deflection of stereocilia in response to sound-induced cochlear fluid motion leads to the influx of potassium-rich endolymph through mechanically-gated ion channels. The resultant membrane depolarization, in turn, modulates fast glutamate neurotransmitter release at the IHC’s basolateral membrane via ribbon synapses [1–4].

It has been reported across different species that synaptic ribbons at the modiolar (neural) IHC face are larger than those at the pillar (abneural) face [5–12]. Furthermore, retrograde-tracing experiments in the cat [13] have linked the function of the postsynaptic auditory nerve fiber (ANF) to the morphology of the presynaptic ribbon, showing that ANFs with lower rates of spontaneous activity and higher thresholds of activation preferentially innervate the modiolar IHC side, where they face predominantly large and more complex presynaptic active zones [13–16]. In addition, the Ca²⁺ channel activation potential as well as the appearance of the postsynaptic terminal along the pillar–modiolar axis further contribute to the heterogeneity of the IHC coding function [5, 8, 10, 17]. This evidence suggests the commonly believed parallel information-processing mechanism behind a group of functionally fractionated ANFs coded simultaneously by a single IHC to cover the full sound intensity information in the IHC receptor potential [1, 18, 19].

To maintain sustained synaptic transmission at different rates, precise organization of the energy-supporting mitochondrial network and vesicle-producing membrane system around the presynaptic active zones is required. Emerging evidence suggests that mitochondrial organization [20] and function [21] contribute to the heterogeneity of ribbon synapses, but how the mitochondrial network and ribbon gradient are spatially aligned remains poorly quantified. Despite recent efforts using state-of-the-art volume electron microscopy to reveal the large-scale subcellular organization in mouse IHCs [9, 12, 20, 22, 23], cell-wide reconstruction of IHC organelles including ribbon synapses and mitochondria with essential volume information is still missing.

Here we developed an automated detection and segmentation pipeline for ribbons and mitochondria using artificial intelligence (AI)-based approaches. In this way, we analyzed 34 mature mid-cochlear IHCs from two published mouse datasets [12] that were acquired using serial block-face electron microscopy (SBEM). Our results suggest two distinct IHC subpopulations that differ substantially in cell body shape, basolateral pole orientation, mitochondrion abundance, and ribbon synapse arrangement. Furthermore, our analysis argued for an aligned pillar-modiolar gradient of ribbon morphology and mitochondrial network only at the basolateral IHC pole, rather than a simple size correlation between synaptic ribbons and the proximal-most mitochondria as commonly thought.

Materials and Methods

SBEM Dataset

Two SBEM datasets were used in this study. They were acquired from the mid-cochlear segment of adult CBA/Ca mice (p49 and p60). Sample preparation followed the established protocol [24]. In brief, fresh auditory bullae were harvested from the animal and the cochlea was fixed by perfusion with ice-cold fixative mixture (2% PFA and 2.5% glutaraldehyde buffered in 0.08 mol/L cacodylate, pH 7.4) through the round and oval windows followed by 5 h of post-fixation at 4 °C. Decalcification was done by 4-h immersion in the same mixture with the addition of 5% EDTA. The decalcified cochleae were washed twice then stained en bloc by sequential incubation in 0.15 mol/L cacodylate buffer (pH 7.4) containing 2% OsO₄, 2.5% ferrocyanide, and 2% OsO₄ again at room temperature (RT) for 2, 2, and 1.5 h, respectively. After two 30-min washes with nanopore-filtered water, the cochleae were incubated at RT in filtered thiocarbohydrazide (saturated aqueous solution) for 1 h, unbuffered OsO₄ aqueous solution for 2 h, and lead aspartate solution (0.03 mol/L, pH 5.0 adjusted by KOH) at 50 °C for 2 h with intermediate washing steps. Dehydration and embedding were done through a graded acetone series (50%, 75%, and 90%, 30 min each, all at 4 °C) into pure acetone (3 × 100%, 30 min at RT) followed by sequential infiltration with 1:1 and 1:2 mixtures of acetone and Spurr’s resin monomer (4.1 g ERL 4221, 0.95 g DER 736, 5.9 g NSA, and 1% DMAE) and the pure resin at RT for 6 h each. The infiltrated cochleae were then placed in embedding molds and incubated in a pre-warmed oven (70 °C) for 72 h.

For data acquisition, serial images were captured in single-tile mode (20,000 × 15,000 pixels) of 11 nm size and 40 nm cutting thickness for the p49 dataset, and 17904 × 21052 pixels of 12 nm size and 50 nm cutting thickness for the p60 dataset; incident beam energy was 2 keV and dwell time 1 μs. Focal charge compensation [25] was set to 100% with a constant vacuum chamber pressure of ~ 2.8 × 10⁻³ mbar. The datasets were aligned to the maximum cross-correlation between consecutive slices by a self-written MATLAB script. This yielded two SBEM volumes of 262 × 194 × 100 μm³ (p49 dataset) and 215× 253 × 154 μm³ (p60 dataset).

3D Reconstruction of IHCs and Mitochondria

A multi-stage method was introduced for the reconstruction of IHC cell bodies including mitochondria (Fig. 1A).

Fig. 1.

Volume reconstruction of IHCs. A Pipeline for reconstructing IHC cell bodies (green box) and mitochondria (orange box). In small image patches (11.26 × 11.26 × 1.28 μm³) 2D binary masks of cell bodies and mitochondria are obtained via automated segmentation by trained CNNs. In-plane watershed and 3D connect operations yield 3D volumes and each mitochondrion is assigned to the maximum overlapped cell body. B Representative volume reconstruction of an IHC (green) with all cellular mitochondria (color). Inset: raw electron micrograph masked by segmented mitochondria with pseudo-colors. Scale bar, 1 μm. C Detection and segmentation of synaptic ribbons. Inset: example electron micrograph (upper left) of a ribbon synapse with annotated synaptic membrane thickening and ribbons (orange), which are used as features for training the region-CNN on 5.63 × 5.63 × 1.28 μm³ image patches (lower left). 3D bounding boxes and segmentation masks of the ribbons are predicted. The shared backbone (purple block) is modified from 3D ResNet50 [54] and used in combination with a region proposal network (RPN, green block) for region proposal generation and follow-up CNN segmentation in the region of interest. D Volume reconstruction of the same IHC (green) as shown in (B) but with all ribbon synapses (color). Inset: electron micrograph of the segmented ribbon (blue) with associated vesicles (red arrowhead). Scale bar, 0.5 μm. Note that the synaptic membrane thickening was used for feature detection but removed from the segmented ribbons.

Network Configuration

To identify IHCs from an SBEM dataset, binary masks of IHC cell bodies were generated by a modified convolutional neural network (CNN) of 3D U-net architecture [26] (Fig. S1A). 3D convolution operations were introduced into the network to help exploit the spatial context information. The skip-connection between the analysis path and the synthesis path combined detailed and abstract features (Fig. S1A).

Network Training

The ground truth annotation was done in a volume of 4000 × 4000 × 200 voxels in the p49 dataset. The split ratio of training, validation, and test sets was 6:2:2. To enlarge the CNN receptive field, the raw images and labels were down-sampled to 22 × 22 × 40 nm³. The training ran 93,500 iterations (71.7 h) for the network for IHC cell body segmentation. Precision of 0.9954 and recall of 0.9889 were achieved when evaluated with the test dataset.

Segmentation of IHC Cell Bodies

Next, the trained network was applied to the full dataset in a blocking scheme (Fig. S1C). The full-size datasets were cut into small blocks with overlaps. Binary cell body masks were generated by the trained network in parallel and then merged by alignment to the overlapping regions to obtain the whole-cell masks.

3D Connection of Segmentation Masks

Once in-plane segmentation was done, we set out to fuse 2D-connected objects into 3D volumes. In order to avoid segmentation errors caused by densely-packed IHCs, we used a distance–transform–watershed algorithm [27] to separate adjacent cells. The obtained super-pixels were then used for 3D IHC reconstruction using our recently published 3D connection algorithm [28], in which the intersection over union (IoU) value was indexed to measure the similarities between organelles in adjacent layers.

Segmentation of Mitochondria

Like the segmentation of cell bodies, we designed a CNN in 3D U-net architecture containing three categories (mitochondrion, mitochondrial membrane, and background). The size of the input data was 1024 × 1024 × 100 voxels and the split ratio of training, validation, and test sets was 6:2:2. Training of the network took 25.3 h (126,500 iterations). The segmentation performance of mitochondria on test sets achieved 0.9717 in precision and 0.9670 in recall. With the same fusion strategy, 2D masks of mitochondria were obtained and connected in 3D for reconstruction. Finally, each mitochondrion was assigned to the IHC with maximum overlap of segmentation masks.

3D Reconstruction of Ribbon Synapses

A region-CNN [29] was used for ribbon synapse detection (Fig. 1C).

Network Configuration

Image patches of 512 × 512 × 32 voxels were used as inputs for the 3D detection network, which predicted 3D bounding boxes and segmentation masks of ribbons. This in principle allows 3D mask generation directly from raw datasets. In addition to synaptic ribbons, we included synaptic membrane thickening as a feature and proposed a novel two-task learning strategy. The detection branch regressed with the 3D bounding box containing both ribbon and synaptic membrane, while the segmentation branch predicted the 3D mask of ribbon only. The network architecture (Fig. 1C) was modified from a 3D version of the Mask region-CNN [29] which consists of a detection and a mask branch. The configuration of the backbone network was tuned to the anisotropic datasets by setting the kernel size and stride to 7 × 7 × 5 and 2 × 2 × 1 in the first convolution layer, followed by 3D max-pooling layers with 3 × 3 × 3 pool size and 2 × 2 × 2 strides (Fig. S1B). In the other layers, all parameters were kept the same as the original settings but in 3D format. Candidate regions were generated by a region–proposal–network for producing feature maps of a fixed size (7 × 7 × 3 for detection and 14 × 14 × 6 for segmentation). Further, the region–CNN finalized the detection and completed the segmentation of the detected 2D objects. Note that the last transposed convolution layer had a kernel size of 2 × 2 × 1 and a stride of 2 × 2 × 1 to preserve the mask shape (28 × 28 × 6 voxels).

Training of Region-CNN

To train the network and evaluate its detection performance, we construct a large dataset containing 1630 patches of 512 × 512 × 32 voxels with a split ratio of 6:2:2 for training, validation, and test sets. The training process took 87.5 h and in total 32,280 iterations. Evaluation of the trained network with the test sets suggested a precision of 0.9200 and a recall of 0.8846.

Segmentation of Synaptic Ribbons

Due to in-house limitation on GPU and runtime memory, block-wise detection was carried out on the full-size datasets. Ribbon masks were obtained from small blocks (512 × 512 × 32 voxels) with overlaps (32 × 32 × 4 voxels) and neighboring pairs were then fused in three directions using information from overlapping regions. To avoid false positives during structure matching, we fused neighboring blocks with an IoU value >0.5. Finally, each segmented ribbon was assigned to the IHC with maximal overlap.

Experimental Setup

We implemented all the networks on Tensorflow and Keras. Training and testing were run on a server equipped with NVIDIA Tesla V100 GPUs. The networks for segmentation of mitochondria and IHCs were optimized by the Adam algorithm [30] at a learning rate of 0.001. The network for ribbon segmentation was optimized using stochastic gradient descent at a learning rate of 0.001 and a momentum of 0.9. To improve the robustness of the networks, we enriched the training sets by randomly rotating, flipping, and elastic deforming the raw images [31] as well as introducing noise during the training process. To make the network learn robust and strong features, we applied a brightness/contrast transform and random resizing to simulate the changes in grayscale and resolution of SBEM datasets.

IHC Compartmentalization

Supranuclear and Basal Compartments

To divide the IHCs into supranuclear and basal compartments, the centroids of all mitochondria were clustered into two categories using the k-means algorithm [32]. The boundary between two clusters aligned with the virtual plane separating the supranuclear and basal IHC compartments.

Pillar and Modiolar Hemispheres

An automated method based on principal component analysis (PCA) [33] was used to divide the IHC into pillar and modiolar hemispheres. In detail, we started with the center coordinates of one IHC row and fit a line $l_1$ using the least square method. Next, the principal axis $l_2^i$ was computed by PCA for the i-th IHC. The coronal plane of the i-th IHC was then determined by its center point and the direction vector of these two lines ($l_1,l_2^i$). Mitochondria and ribbons were assigned to the pillar or modiolar hemispheres according to their locations relative to this plane.

Correlation Analysis of Ribbon and Mitochondrial Distribution

The formula used for computing the correlation coefficient was:

$$Cor{r}_{\%V_{rib-mod.}vs.\%V_{mito-mod.}}=\frac{E\left[\left(\%V_{rib-mod.}-{\mu }_{\%V_{rib-mod.}}\right)\left(\%V_{mito-mod.}-{\mu }_{\%V_{mito-mod.}}\right)\right]}{{\sigma }_{\%V_{rib-mod.}}{\sigma }_{\%V_{mito-mod.}}}$$

where $\%V_{rib-mod.}=\frac{V_{rib-mod.}}{V_{rib-total}}$ is the volume fraction of modiolar ribbons, while $\%V_{mito-mod.}=\frac{V_{mito-mod.}}{V_{mito-total}}$ is the volume fraction of mitochondria in the modiolar hemisphere. E and $\sigma$ denote the mean and standard deviation.

Statistics

All data analysis and statistical tests were performed using self-written scripts in MATLAB (MathWorks, Inc.) including built-in functions and the Statistics Toolbox. Groups were compared using the two-sample t-test (ttest2) for Figs. 2C–E, 3A, B, D, E, 4B–F, and S2, while the paired t-test (ttest1) was used for Fig. S3B. The two-sample Kolmogorov-Smirnov test was used for Fig. 3D, E, and S3A. Pearson linear correlation analysis was used in Fig. 5A, B, D, E. All data are represented as the mean ± SD and the significance level of statistical tests was denoted as n.s. P > 0.05, *P < 0.05, **P < 0.01, ***P < 0.001.

Fig. 2.

IHC subpopulations in the mid-cochlear segment. A Sagittal view of volume reconstructed IHCs in a staggered arrangement. The type-A IHC has a basolateral pole tilted towards the pillars (large intersection angle between the reticular lamina (RL) and IHC middle line, α), whereas that of the type-B IHC tilts more towards the modiolus (small RL-to-IHC intersection angle, β). Scale bars, 20 μm. B Longitudinal view of the seven type-A IHCs (blue) and the eight type-B IHCs (orange) in the p49 dataset. Scale bars, 20 μm. C Comparison of the RL-to-IHC intersection angles in pooled IHCs from the two datasets (triangles: p49 dataset, dots: p60 dataset). Larger angles occur in the type-A (53.1 ± 9.1°, n = 18) than in the type-B IHCs (36.0 ± 11.2°, n = 16, ***P < 0.001, two-sample t-test). Note that one IHC in the p49 dataset has an intermediate tilt angle and post-hoc classification of this cell was based on cell length and volume. D Comparison of the cell body length. Type-A IHCs (37.07 ± 2.32 μm, n = 18) are significantly longer than type-B IHCs (32.14 ± 2.41 μm, n = 16, ***P < 0.001, two-sample t-test). E Comparison of cell body volume in pooled IHCs from the two datasets. Cell bodies are larger in type-A (2861.86 ± 358.32 μm³, n = 18) vs type-B IHCs (2476.07 ± 463.60 μm³, n = 16, *P  = 0.010, two-sample t-test).

Fig. 3.

Quantification of synaptic ribbons in IHCs. A Comparison of ribbon numbers per cell between type-A (19.7 ± 2.7, n = 18) and type-B IHCs (18.1 ± 2.9, n = 16, P = 0.103, two-sample t-test). B Comparison of cumulative ribbon volume per cell between type-A (0.233 ± 0.028 μm³, n = 18) and type-B IHCs (0.232 ± 0.035 μm³, n = 16, P = 0.942, two-sample t-test). C Normalized basal position (NBP) of ribbons. Left: schematic of the definition of NBP. The NBP value is the projection of the ribbon position on the IHC radial axis (c) over the total cell length (L). Positive NBP values indicate the modiolar IHC face, while negative values refer to the pillar face. Right: illustration of one IHC pair (type-A, blue; type-B, orange) divided into modiolar and pillar hemispheres. D Cumulative probability distribution of NBP values from pooled ribbons. Ribbons of type-A IHCs (blue, 0.064 ± 0.119, n = 354) are located predominantly on the modiolar IHC face, while more pillar ribbons are found in type-B IHCs (orange, –0.033 ± 0.103, n = 289, ***P < 0.001, two-sample t-test and two-sample Kolmogorov-Smirnov test). E Cumulative probability distribution of volumes of pillar (thick blue line, 0.008 ± 0.002 μm³, n = 79) and modiolar ribbons (thin blue line, 0.013 ± 0.002 μm³, n = 275, ***P < 0.001, two-sample t-test and two-sample Kolmogorov-Smirnov test) from type-A IHCs, as well as from those from type-B IHCs (pillar: thick orange line, 0.013 ± 0.002 μm³ [n = 220] vs modiolar: thin orange line, 0.014 ± 0.004 μm³ [n = 69]; P = 0.061, two-sample t-test, *P = 0.013, two-sample Kolmogorov-Smirnov test).

Fig. 4.

Quantification of IHC mitochondria. A Scatterplot of mitochondria in two representative IHCs. Colormap represents the normalized mitochondrial size. B Comparison of mitochondrion counts per IHC. The type-A IHCs (1777.39 ± 277.89, n = 18) have slightly more mitochondria than type-B (1526.44 ± 359.42, n = 16; *P = 0.029, two-sample t-test). C Comparison of mean mitochondrial volume per IHC. The wo subpopulations show no significant difference (type-A: 0.065 ± 0.007 μm³ [n = 18] vs type-B: 0.068 ± 0.020 μm³ [n = 16]; P = 0.675, two-sample t-test) except for one outlier in the type-B group. D Comparison of the cumulative mitochondrial volume per IHC. The type-A IHCs (114.80 ± 11.64 μm³, n = 18) have larger mitochondrion volume than type-B (99.11 ± 18.47 μm³, n = 16; **P = 0.005, two-sample t-test). E, F Comparison of the per-IHC cumulative mitochondrial volume in the supranuclear (E) and basal compartments (F). Supranuclear: 37.10 ± 5.35 μm³ (type-A, n = 18) vs 36.99 ± 9.31 μm³ (type-B, n = 16). P = 0.964, two-sample t-test; basal: 77.70 ± 9.33 μm³ (type-A, n = 18) vs 62.13 ± 14.11 μm³ (type-B, n = 16). ***P < 0.001, two-sample t-test.

Fig. 5.

Correlation between ribbons and mitochondria. A Scatterplot of ribbon volume vs distance from the nearest mitochondrion (Pearson correlation: r = 0.04, P = 0.355). B Scatterplot of ribbon volume vs volume of the nearest mitochondrion (Pearson correlation: r < 0.01, P = 0.992). C Correlation coefficients between the volume fraction of modiolar ribbons and that of modiolar mitochondria in the basal IHC compartment ranging from the basolateral pole. A pole distance-dependent correlation is evident with a maximum at 4.4 μm. Right: schematic of the ribbons and mitochondria included in the analysis when changing the pole distance from 4.4 μm to 10 μm. D Scatterplot of the volume fraction of modiolar ribbons vs that of the mitochondria within 4.4 μm from the basolateral pole (Pearson correlation: r = 0.67, ***P < 0.001). E As in (D) but with the mitochondria within 10.0 μm from the basolateral pole (Pearson correlation: r = 0.28, P = 0.105).

Results

Machine Learning-Based Segmentation of IHC Organelles

Several electron microscopy (EM) volumes of rodent cochleae have recently been reported [9, 12, 20, 22, 23]. As these datasets were acquired using different imaging parameters and analyzed manually, it remained to be determined whether machine learning-based image-processing tools could be applied to extract structural information in an automated fashion. For this, we applied CNNs to two large EM datasets of mid-cochlear segments from adult mice (p49 and p60, Table 1). The CNNs were trained on ground truth annotations of IHC surface membrane (635.8 μm², 1.96%), mitochondria (n = 319, 0.56%), and synaptic ribbons (n = 239, 37.17%) via webKNOSSOS (a browser-based annotation tool [34]) in different subsets (see Material and Methods). A two-step approach was used for 3D segmentation of IHC cell bodies and mitochondria (Fig. 1A). First, the trained CNNs were used to produce binary masks of the target structures in each image slice. This resulted in separate detection of the cell bodies (Fig. 1B, precision and recall: 0.9954 and 0.9889) and the mitochondria (precision and recall: 0.9939 and 0.9897). Subsequently, a 3D connection algorithm was applied to generate 3D masses and finally each mitochondrion was assigned to the IHC according to their spatial overlap (Fig. S1C). As for ribbons, we proposed a 3D instance segmentation network (Fig. 1C) for end-to-end feature detection and segmentation (Fig. 1D, precision and recall: 0.95 and 0.95). All automatically segmented structures were manually proofread before being used for follow-up quantification.

Table 1.

SBEM datasets of two CBA mice.

	CBA-1	CBA-2
Sex	female	female
Age	P49	P60
Pixel	20,000 × 15,000	17,904 × 21,052
Resolution (nm3)	11 × 11 × 40	12 × 12  × 50
Volume (μm3)	262 × 194 × 100	215 × 253 × 154
# IHCs	15	19
# ribbons (per cell)	20.1 ± 3.1	17.9 ± 2.2

Morphological Quantification of IHC Subpopulations

Structural quantification was carried out on 34 reconstructed IHCs from two datasets. As previously reported, a staggered arrangement of IHCs was prevalent in the mouse mid-cochlear segment [35]. We classified the IHCs based on distinct orientations of their basolateral poles, with type-A tilted more towards the pillar and type-B towards the modiolus (Fig. 2A). This yielded seven type-A and eight type-B IHCs in the p49 dataset (Fig. 2B) as well as eleven type-A and eight type-B IHCs in the p60 dataset.

Analysis of the pooled datasets confirmed distinct tilt angles between the IHC habenular-cuticular axis and the reticular lamina [Fig. 2C, 53.1 ± 9.1° (type-A, n = 18) vs 36.0 ± 11.2° (type-B, n = 16), P < 0.001]. It was notable that the two datasets had different cut-off angles for the two IHC subpopulations (53.3° for the p49 and 35.8° for the p60 dataset), between which a consistent intersection angle was found (17.7° for p49 and 20.3° for p60 dataset). Therefore, we compared all quantified structures between type-A and type-B IHCs in the p49 and p60 datasets with statistical tests (Table 2). Moreover, the type-A IHC cell bodies were found to be 15.6% larger [Fig. 2E, 2861.86 ± 358.32 μm³ (type-A, n = 18) vs 2476.07 ± 463.60 μm³ (type-B, n = 16), P = 0.010] and 15.3% longer along the habenular-cuticular axis [Fig. 2D, 37.07 ± 2.32 μm (type-A, n = 18) vs 32.14 ± 2.41 μm (type-B, n = 16), P < 0.001] than the type-B IHCs. When divided into supranuclear and basal compartments, IHCs showed comparable volumes of the supranuclear compartment [Fig. S2A, 956.71 ± 190.91 μm³ (type-A, n = 18) vs 1015.31 ± 209.96 μm³ (type-B, n = 16), P = 0.400] but significantly larger basal compartments in the type-A IHCs [Fig. S2B, 1904.14 ± 265.66 μm³ (type-A, n = 18) vs 1459.91 ± 336.73 μm³ (type-B, n = 16), P < 0.001].

Table 2.

Comparison of Type-A and Type-B IHCs in the two datasets.

	P49 dataset			P60 dataset
	Type-A	Type-B	ttest2	Type-A	Type-B	ttest2
Intersection angle (°)	63.6 ± 3.2	45.9 ± 6.3	***P < 0.001	46.4 ± 3.0	26.1 ± 2.1	***P < 0.001
IHC length (μm)	38.1 ± 2.8	31.1 ± 2.4	***P < 0.001	36.4 ± 1.8	33.2 ± 2.1	**P = 0.002
IHC volume (μm3)	2592 ± 360	2172 ± 472	P = 0.077	3033 ± 240	2781 ± 162	*P = 0.020
Nrib. per IHC	21.7 ± 2.6	18.8 ± 3.0	P = 0.065	18.4 ± 1.8	17.4 ± 2.8	P = 0.36
Vrib. per IHC	0.25 ±0.02	0.24 ± 0.05	P = 0.65	0.22 ± 0.03	0.23 ± 0.02	P = 0.88
Nmito. per IHC	1632 ± 254	1308 ± 373	P = 0.075	1870 ± 261	1745 ± 168	P = 0.25
Mean Vmito.	0.067 ± 0.006	0.072 ± 0.028	P = 0.60	0.065 ± 0.008	0.063 ± 0.004	P = 0.57
Vmito. per IHC	107.8 ± 13.1	88.9 ± 20.4	P = 0.056	119.3 ± 8.4	109.3 ± 8.9	*P = 0.023
Vmito. per IHCsupra. (μm3)	34.2 ± 4.2	37.1 ± 10.7	P = 0.52	39.0 ± 5.3	36.9 ± 8.4	P = 0.52
Vmito. per IHCbasal (μm3)	73.6 ± 10.4	51.8 ± 11.7	**P = 0.002	80.3 ± 8.0	72.4 ± 7.0	*P = 0.038

Data are presented as the mean ± SD.

Quantification of Ribbon Synapse Organization

Volume and location information on the synaptic ribbons was extracted for each IHC. In total, 643 ribbons were found and almost every one of them (639 out of 643) was paired with one postsynaptic terminal. The two IHC subpopulations did not differ in ribbon counts [Fig. 3A, 19.7 ± 2.7 (type-A, n = 18) vs 18.1 ± 2.9 (type-B, n = 16), P = 0.103] and total volume [Fig. 3B, 0.233 ± 0.028 μm³ (type-A, n = 18) vs 0.232 ± 0.035 μm³ (type-B, n = 16), P = 0.942]. In order to quantify the pillar-modiolar gradient of ribbon synapses, we computed for each ribbon the normalized basal position (NBP, Fig. 3C) as described in [36]). Ribbons of the type-A IHCs were found more frequently on the modiolar IHC side, whereas the type-B IHCs had more pillar ribbons [Fig. 3D, mean NBP: 0.064 ± 0.119 (n = 354 ribbons in type-A IHCs) vs –0.033 ± 0.103 (n = 289 ribbons in type-B IHCs), P < 0.001]. Next, we compared the volume distribution between modiolar and pillar ribbons in the two subpopulations. A prominent pillar-modiolar size gradient of ribbons was only found in the type-A IHCs [Fig. 3E, type-A: 0.008 ± 0.002 μm³ (n = 79 pillar ribbons) vs 0.013 ± 0.002 μm³ (n = 275 modiolar ribbons), P < 0.001], while in the type-B IHCs large ribbons were found on both the pillar and modiolar sides [0.013 ± 0.002 μm³ (n = 220 pillar ribbons) vs 0.014 ± 0.004 μm³ (n = 69 modiolar ribbons), P = 0.061]. Taken together, these results argue for a distinct modiolar-pillar gradient of ribbon abundance and size in the two IHC subpopulations.

IHC-wide Quantification of Mitochondria

The AI-based segmentation approach facilitated structural quantification of thousands of mitochondria per IHC. We analyzed their volume and distribution in both the supranuclear and basal compartments of the two IHC subpopulations (Fig. 4A). The number of mitochondria per IHC ranged from 840 to 2,261 and on average the type-A IHCs had 16.4% more mitochondria than their type-B counterparts [Fig. 4B, 1,777.39 ± 277.89 (type-A, n = 18) vs 1,526.44 ± 359.42 (type-B, n = 16), P = 0.029], while they were comparable in size in both IHC subpopulations [Fig. 4C, 0.065 ± 0.007 μm³ (type-A, n = 18) vs 0.068 ± 0.020 μm³ (type-B, n = 16), P = 0.675]. This was confirmed by a larger total volume of mitochondria in the type-A IHCs [Fig. 4D, 114.80 ± 11.64 μm³ (type-A, n = 18) vs 99.11 ± 18.47 μm³ (type-B, n = 16), P = 0.005] and this inter-population difference was contributed by the basal IHC compartment [Fig. 4F, basal: 77.70 ± 9.33 μm³ (type-A, n = 18) vs 62.13 ± 14.11 μm³ (type-B, n = 16), P < 0.001] rather than the supranuclear compartment [Fig. 4E, supranuclear: 37.10 ± 5.35 μm³ (type-A, n = 18) vs 36.99 ± 9.31 μm³ (type-B, n = 16), P = 0.964]. Furthermore, the intracellular comparison in pooled IHCs revealed a larger mean volume of mitochondria in the basal than the supranuclear compartment [Fig. S3A, 0.063 ± 0.017 μm³ (supranuclear, n = 34) vs 0.068 ± 0.013 μm³ (basal, n = 34), P < 0.001] as well as a higher mitochondrial volume fraction in the basal compartment [Fig. S3B, 3.79 ± 0.39% (supranuclear, n = 34) vs 4.18 ± 0.27% (basal, n = 34), P < 0.001]. Interestingly, large IHCs, which had more mitochondria (Fig. S4A), did not host more ribbon synapses (Fig. S4B), arguing against a universal scale-up of organelles with cell body size.

Structural Linkage Between Synaptic Ribbons and Local Mitochondria

A recent study has suggested that the size of ribbons is regulated by synaptic mitochondrial function [21]. But whether the mitochondrial organization is fine-tuned to individual ribbons and to what extent it contributes to the characteristic ribbon number and size gradient along the IHC pillar-modiolar axis remained elusive. In pooled ribbon synapses from both IHC subpopulations, we did not find notable correlations between ribbon size and the distance from the proximal-most mitochondrion (Fig. 5A) or the volume of the mitochondrion (Fig. 5B), arguing against the existence of a local regulatory effect at the single-synapse level. Next, we explored a potential structural linkage between the spatial gradient of ribbon morphology along the pillar-modiolar axis and the distribution of “associated” mitochondria by mapping the correlation between the volume fraction of modiolar ribbons and that of local mitochondria recruited with increasing distance from the basolateral pole (Fig. 5C). This led to a peak correlation between ribbon and mitochondrial distribution in the proximity of the basolateral pole (≤ 4.4 μm, Fig. 5D), beyond which the correlation rapidly decayed (e.g., ≤ 10.0 μm, Fig. 5E). Note that the structural correlation was mainly attributable to the intertype difference (Fig. 5D), suggesting distinct structural tuning of these two IHC subpopulations. Taken together, these results reveal a distance-dependent alignment of the pillar-modiolar ribbon gradient and the local mitochondrial network.

Discussion

Over decades, compartmentalized IHC cytoarchitecture has been studied primarily using transmission electron microscopy (TEM), which provides nanometer-resolution in 2D thin slices, whereas serial-section TEM requires intense human labor and so 3D reconstructions were usually limited to small volumes which often contained only few IHCs [15, 16, 37]. Recent advances in volume EM techniques enable high-throughput large-scale structural mapping with improved isotropic spatial resolution (see reviews [38–40]. Despite modern computational power, nowadays, processing of big 3D image data is still a bottleneck for volume EM-based structural investigations. In this study, an AI-based approach was applied to reliably extract volume information on IHC mitochondria and ribbon synapses. To the best of our knowledge, automated ribbon detection in a 3D EM dataset has not yet been established elsewhere. This method will not only facilitate comprehensive structural quantification of IHCs under different pathological conditions but be applicable to other sensory receptor cells and neuron types (such as vestibular hair cells, retinal photoreceptors, and bipolar cells) where ribbon-type synapses are also found.

Morphological and Functional Heterogeneity of Mid-Cochlear IHCs

Prior studies using light or electron microscopy documented the staggered organization of mid-cochlear IHCs in mice [12, 20, 35]. This feature was believed to be crucial for the high sensitivity of mid-frequency sound perception [35], as it effectively increases the packing density of sensory IHCs. Although these studies, including ours, cannot completely rule out the possibility of fixation artifacts, the observed structural heterogeneity, i.e., a distinct spatial gradient of ribbons in the two IHC subpopulations, argues for a putative functional tuning. Complementary to recent studies [20, 23], adjacent IHCs were found not only to differ in length and volume (Fig. 2D, E), but also tend to have alternating orientated basolateral poles (Fig. 2C). Such IHC positioning, however, is supposed to compromise the spatial gradient of ribbon morphology that is well-described in a single IHC row, as some synapses on the pillar IHC face may spatially overlap with modiolar synapses from a neighboring IHC. In fact, small ribbons were found exclusively on the pillar face of the type-A IHCs but were absent from both the pillar and modiolar faces of the type-B IHCs (Fig. 3D), providing a plausible explanation for the observation of a continuous modiolar-pillar gradient of ribbon morphology in tissue [35]. Considering that large and multiple ribbons were found preferentially presynaptic to ANFs with higher thresholds of activation and thereby likely to be responsible for the encoding of loud sounds, our results indicate structural tuning of adjacent IHCs in the mid-cochlear segment to complement the sound-encoding function which collectively covers the entire audible range. In line with this idea, neighboring IHCs were recently found to not only be coupled electrically and metabolically [41] but also co-innervated by a small fraction of branched ANFs [12]. They may thereby form a “mini-syncytium” for enhanced sensitivity and reliability of cochlear sound encoding over a wide dynamic range [41].

Organization of Synaptic Mitochondria in IHCs

It is critical for normal hearing to maintain diversified afferent connections, which are set by the expression of distinct complements of transcription factors in the postsynaptic spiral ganglion neurons [42–44] and presynaptic Gαi activity [45]. Mild noise exposure, ageing, and cochlear de-efferentation can cause synaptopathy preferentially on the modiolar IHC face, where large synaptic ribbons are prevalent [35, 46–51]. Both pre- and postsynaptic mechanisms involving mitochondrial dysfunction have been proposed, but less is known about how the presynaptic mitochondrial network contributes to the maturation of heterogeneity in IHC afferent synapses. It was shown recently in developing zebrafish IHCs that Ca²⁺ uptake by synaptic mitochondria lowers the NAD⁺/NADH ratio and in turn may function to restrict the ribbon size [21], while at the whole-cell level, a prior study using SBEM and point-counting stereology revealed matched asymmetric distributions of ribbons and mitochondria [20]. However, we found that in mature mouse IHCs, small ribbons were not necessarily accompanied by closer (Fig. 5A) or larger mitochondria (Fig. 5B) than large ribbons. The absence of local structural correlation may explain why the ribbon size is no longer tightly regulated by mitochondria after IHC maturation, at least in zebrafish [21]. Strikingly, in the basal compartment of IHCs, mitochondrion-associated membrane cisternae have been reported in gerbil and mouse [20, 52, 53]. Coexistence of this structure and cytoplasmic vesicles imply an atypical endoplasmic reticulum system for IHC vesicle biogenesis, which might be critical for maintaining synaptic ribbon function by preventing exhausted vesicle cycling during sustained high-speed neurotransmission. In line with the idea, our finding of a location-specific correlation between the volume fraction of modiolar ribbons and local mitochondria (Fig. 5C) argues for aligned gradients of these two functional organelles only at the basolateral IHC pole (≤4.4 μm from the basal end) where the terminals of mitochondrion-associated cisternae—putative vesicle generating sites—are located [52]. Furthermore, our cell-wide quantification of mitochondria may serve as the structural basis for the precise functional compartmentalization of the IHC into a basolateral ‘synaptic pole’ for high-speed neurotransmitter release and a supranuclear ‘synthetic pole’ for biomolecule synthesis, as well as a membrane cistern-enriched peri-nuclear region [20, 52].

Supplementary Information

Below is the link to the electronic supplementary material.

Data Availability

The SBEM datasets and segmentation supporting the current study have been deposited in a public repository (https://wklink.org/9870). All codes used for analysis were uploaded to github (https://github.com/MiRA-lab-dev/IHC_organelles_detection) for free download.

Conflict of interest

The authors have declared that no conflicts of interest exist.