Quantification of cristae architecture reveals time-dependent characteristics of individual mitochondria

Abstract

Recent breakthroughs in live-cell imaging have enabled visualization of cristae, making it feasible to investigate the structure-function relationship of cristae in real time. However, quantifying live-cell images of cristae in an unbiased way remains challenging. Here, we present a novel, semi-automated approach to quantify cristae, using the machine-learning Trainable Weka Segmentation tool. Compared with standard techniques, our approach not only avoids the bias associated with manual thresholding but more efficiently segments cristae from Airyscan and structured illumination microscopy images. Using a cardiolipin-deficient cell line, as well as FCCP, we show that our approach is sufficiently sensitive to detect perturbations in cristae density, size, and shape. This approach, moreover, reveals that cristae are not uniformly distributed within the mitochondrion, and sites of mitochondrial fission are localized to areas of decreased cristae density. After a fusion event, individual cristae from the two mitochondria, at the site of fusion, merge into one object with distinct architectural values. Overall, our study shows that machine learning represents a compelling new strategy for quantifying cristae in living cells.

Figure 1.. Trainable Weka Segmentation protocol more effectively segments mitochondrial cristae from living cells compared with various thresholding standards.

(A, B, C, D) Step-wise comparison of different segmentation workflows: Conventional (manual) (A); MaxEntropy (B); Shanbhag (C); and Trainable Weka Segmentation (D). Note that without a probability map, Conventional, MaxEntropy, and Shanbhag thresholding, which depend exclusively on pixel intensities, cannot effectively segment cristae. (A, B, C, D) Blue arrowheads denote successful segmentation of cristae (column D) versus ineffective segmentation (columns A, B, C). Scale bars = 500 nm.

Figure 2.. Trainable Weka Segmentation (TWS) protocol offers significant advantages over standard thresholding techniques for the quantification of cristae in living cells.

Quantification of cristae parameters using TWS protocol versus standard thresholding. (A, C, E) Quantification of cristae density (cristae#/μm²) in HeLa cells stained with NAO. Note: TWS is significantly more effective at measuring cristae density compared with Conventional, MaxEntropy, and Shanbhag segmentation methods, respectively. N = 8 independent experiments. (B, D, F) Quantification of cristae area (μm²) in HeLa cells stained with NAO. Note: Conventional and MaxEntropy segmentation are unable to segment cristae effectively, resulting in significantly higher values in cross-sectional area, compared with Weka segmentation. Conversely, Shanbhag segmentation shows average cristae areas similar to those of Weka segmentation, but this results from Shanbhag being overly restrictive, subsequently underestimating or entirely missing a large proportion of cristae structures (see Fig 1). N = 8 independent experiments. Data information: Data are presented as mean ± SD. P-values are shown in panels (t tests).

Figure 3.. Trainable Weka Segmentation protocol enables segmentation of cristae in a variety of cell types.

(A, B, C, D) Live-cell Airyscan images of HeLa (A), L6 (B), H1975 (C), and HUH7 (D) mitochondria, stained with NAO. Note that top row shows original images (scale bars = 500 nm). Bottom row, including circular, zoomed-in regions, show probability maps of cristae in respective cell types (red scale bars = 200 nm). Blue arrowheads denote cristae. (E) Quantification of cristae density between HeLa, L6, H1975, and HUH7 cells, stained with NAO. N ≥ 3 independent experiments. (F) Quantification of cristae area between HeLa, L6, H1975, and HUH7 cells, stained with NAO. N ≥ 3 independent experiments. (G) Quantification of cristae aspect ratio between HeLa, L6, H1975, and HUH7 cells, stained with NAO. N ≥ 3 independent experiments. Note that HeLa cells tend to have lower cross-sectional area of cristae together with increased aspect ratio. Data information: Data are presented as mean ± SD. P-values are shown in panels (ANOVA).

Figure S1.. Live-cell images and probability maps of cristae using different mitochondrial dyes.

(A) Top row: Live-cell Airyscan image of HeLa cells stained with Rho123, showing original image (left) and cristae probability map (right). Blue arrowheads denote cristae. Scale bar = 500 nm. (B) Bottom row: Live-cell Airyscan image of HeLa cells stained with MTG, showing original image (left) and cristae probability map (right). Blue arrowheads denote cristae. Scale bar = 500 nm.

Figure 4.. Trainable Weka Segmentation protocol is sufficiently sensitive to detect differences in cristae density, area, and shape in cardiolipin-deficient (sh-PTPMT1) H1975 cells, a model of inner mitochondrial membrane dysregulation.

(A) Image of sh-Scramble H1975 cell, stained with NAO and Hoechst. Scale bar = 5 μm. (B) Image of sh-PTPMT1 H1975 cell, stained with NAO and Hoechst. Scale bar = 5 μm. (C) Zoomed-in region of (A), showing normal lamellar cristae (blue arrowheads) in original image (left) compared with cristae probability map (right). Scale bars = 500 nm. (D) Zoomed-in region of (B), showing deranged cristae (blue arrowheads) in original image (left) compared with cristae probability map (right). Scale bars = 500 nm. Note that these images appear to corroborate previously published EM data, showing loss of PTPMT1 results in swelling and disruption of cristae structure. (E) Quantification of cristae density between sh-Scramble and sh-PTPMT1 in H1975 cells, stained with NAO. N = 3 independent experiments. (F) Quantification of cristae area between sh-Scramble and sh-PTPMT1 in H1975 cells, stained with NAO. N = 3 independent experiments. (G) Quantification of cristae aspect ratio between sh-Scramble and sh-PTPMT1 in H1975 cells, stained with NAO. N = 3 independent experiments. Data information: Data are presented as mean ± SD. P-values are shown in panels (t tests).

Figure 5.. Trainable Weka Segmentation is effective at segmenting cristae in live-cell images obtained with structured illumination microscopy (SIM), highlighting ultrastructural heterogeneity within the same mitochondrion.

(A) Live-cell SIM image of HeLa cell stained with MTG. Scale bar = 5 μm. (B) Mitochondrion cropped from (A), showing fine structure of mitochondrion. Scale bar = 500 nm. Note that the different mitochondrial regions appear to encompass a single fused structure. (C) Cristae probability map of cropped SIM image from (B). White scale bar = 500 nm. Note that the zoomed-in regions show heterogeneous cristae architecture (blue arrowheads) within the same organelle: the gold circle highlights a region of lamellar cristae; the red square shows a variety of arched cristae, running either parallel or perpendicular to the long axis of the organelle; the green rectangle shows a jigsaw configuration; and the blue circle shows cristae spanning adjacent mitochondrial structures. Red scale bars = 100 nm.

Figure S2.. Trainable Weka Segmentation (TWS) protocol is more effective than conventional thresholding at segmenting cristae from SIM images.

(A) SIM image of mitochondrion cropped from HeLa cells stained with MTG. Yellow outlines show segmented cristae regions of interest (ROIs) using TWS protocol. Scale bar = 500 nm. (B) SIM image of mitochondrion cropped from HeLa cells stained with MTG. Yellow outlines show segmented cristae ROIs using conventional thresholding. Scale bar = 500 nm. Note that ROIs produced from conventional thresholding are less effective at segmenting cristae, having more merged structures, which overestimate the area and underestimate the density of cristae. (C) Quantification of cristae density from live-cell SIM images in (A, B), showing more effective segmentation via TWS protocol. N = 3 independent experiments. (D) Quantification of cristae area from live-cell SIM images in (A, B), showing more effective segmentation via TWS protocol. N = 3 independent experiments. Data information: Data are presented as mean ± SD. P-values are shown in panels (t tests).

Figure 6.. Trainable Weka Segmentation protocol can segment different cristae structures found in a range of mitochondrial morphologies.

Live-cell SIM images of HeLa cells stained with MTG (left) and cristae probability maps (right) showing various mitochondria with differing cristae structures. (A) Thin and distended mitochondria, showing relatively small and large arched cristae, respectively (blue arrowheads). Scale bars = 500 nm. N = 3 independent experiments. (B) Ouroboros mitochondrion, showing cristae radiating from central gap (blue arrowheads). Scale bars = 500 nm. N = 3 independent experiments. (C) Elongated mitochondrion with ouroboros-like end, containing cristae radiating from center (blue arrowhead); note the spherical mitochondrion filling the gap within this ouroboros-like structure. Scale bars = 500 nm. N = 3 independent experiments. (D) Fragmented mitochondria, showing arched cristae structures (blue arrowhead). Scale bars = 500 nm. N = 3 independent experiments. (E) Mitochondria of intermediate length, containing various netlike and/or curving cristae (blue arrowheads). Scale bars = 500 nm. N = 3 independent experiments.

Figure 7.. Trainable Weka Segmentation protocol can capture changes in cristae density, area, and shape after acute treatment with FCCP.

(A) Live-cell SIM images of HeLa cells stained with MTG (top row) together with cristae probability maps (bottom row). Note that the left-most column shows a representative image of control, whereas the center and right-most columns show representative images of cells treated with 10 μM FCCP between 0 and 30 min and 30 and 60 min, respectively. Scale bars = 500 nm. N = 3 independent experiments. (B, C, D) Quantification of SIM images from (A), showing a time-dependent decrease in cristae density (B) and cristae area (C), as well as an increase in cristae circularity (D) as a result of the FCCP treatment. N = 3 independent experiments. Data information: Data are presented as mean ± SD. P-values are shown in panels (ANOVA).

Figure S3.. FCCP treatment can result in formation of markedly elongated cristae structures.

(A) Live-cell SIM image (upper panel) of HeLa cells treated with 10 μM FCCP can result in mitochondrial swelling, accompanied by elongation of cristae (blue arrowheads). The lower panel shows cristae probability map. Scale bar = 500 nm. (B) Quantification of cristae aspect ratio, showing no difference between control and FCCP-treated conditions. Note that elongated cristae structures in FCCP-treated cells were accompanied by smaller, circular cristae structures, in other cells; on average, these divergent morphologies resulted in similar values as control. N = 3 independent experiments. Data information: Data are presented as mean ± SD. There were no significant differences between conditions (ANOVA).

Figure 8.. Trainable Weka Segmentation protocol can quantify real-time cristae remodeling within the same mitochondrion.

Zoomed-in, time-lapse SIM images of HeLa cells stained with MTG (top rows) and cristae probability maps (bottom rows). (A) Example of quantification of shape changes before and after cristae fusion event. Note parallel cristae (red and green arrowheads) at ∼28 s are separate structures with aspect ratios of 3.7 and 4.3, respectively; but, after fusing into an arched structure (gold arrowhead), the aspect ratio of the resulting crista is altered to 1.9. Scale bars = 500 nm. (B) Example of quantification of shape change before and after cristae fission events. Note 3 parallel cristae (gold, red, and green arrowheads) at ∼12 s have a circularity of 0.4, 0.3, and 0.4, respectively. After fission of these cristae at ∼14 s, however, membrane fragments (blue, brown, and purple arrowheads) have a circularity of 1, 0.7, and 0.9, respectively. At ∼16 and 18 s, these fragments appear to fuse into a branched and arching structure (red arrowheads), respectively, with a circularity of 0.2 and 0.3. Scale bars = 500 nm. (C) Example of quantification of multiple, consecutive cristae remodeling events. Note that separate cristae (red and green arrowheads) at ∼20 s have a circularity of 0.2 and 0.5, respectively. At ∼22–24 s, these cristae appear to fuse (gold arrowheads), showing alterations in circularity to 0.1 and 0.3, at respective time points. This structure, at ∼26 s, then appears to divide into a smaller crista (blue arrowhead) with a circularity of 0.5 and a larger forked crista (brown arrowhead) with a circularity of 0.2. At ∼28 s, the smaller crista appears to have fused with the upper region of the previously forked crista, resulting in a new structure (purple arrowhead) with a circularity of 0.4, whereas the lower region of the previously forked crista appears to have fused with the crista to the right, generating a more complex network (red arrowhead) with a circularity of 0.2. Scale bars = 500 nm.

Figure 9.. Quantification of dynamic ranges of cristae parameters within individual mitochondria inside the same HeLa cell.

(A, B, C, D, E) Measurement of time-dependent changes in cristae density (A), cristae area (B), cristae circularity (C), cristae aspect ratio (D), and cristae number per mitochondrion (E). Note that each colored line (red, green, and blue) represents time-dependent changes in cristae parameters within a whole and separate mitochondrion inside the same cell. (F) Representative mitochondrion from (A, B, C, D, E). Scale bars = 500 nm. Note that the different values associated with this mitochondrion are displayed by the green curves. (G) Table showing SDs in cristae density, area, circularity, aspect ratio, and cristae number per mitochondrion per min. Note that the SDs in these cristae parameters reflect typical, time-dependent changes in cristae density and architecture. N = 3 independent experiments. Values are shown with associated SDs.

Figure 10.. Trainable Weka Segmentation protocol shows fission sites containing decreased cristae density.

(A, B, C) Representative time-lapse SIM images of HeLa cells stained with MTG (top rows) with cristae probability maps (bottom rows), showing effective segmentation of cristae during mitochondrial fission events. Note the narrowing of the inner boundary membranes (blue arrowheads) before mitochondrial fission into two daughter mitochondria. Also note the decreased cristae density at fission sites. Scale bars = 500 nm. N = 3 independent experiments.

Figure 11.. Trainable Weka Segmentation protocol can quantify cristae remodeling during mitochondrial fusion events.

(A) Representative time-lapse SIM images of HeLa cells stained with MTG. Scale bars = 500 nm. Note at ∼4 s, a finger-like region of inner mitochondrial membrane extends from the tip of the mitochondrion on the left (blue arrowhead), before fusion with mitochondrion on the right in the following frame. (B) Cristae probability maps of time-lapse images from (A). Scale bars = 500 nm. Note at ∼24 s, the separate cristae (red and green arrowheads) of the adjoining mitochondria have circularities of 0.7 and 0.5, respectively; however, after fusion in the following frame, the crista shows a circularity of 0.2, marking a transition from a less to more branching structure.

Figure S4.. Before mitochondrial fusion, finger-like extension of inner mitochondrial membrane appears to bridge membranes of separate organelles.

(A) Time-lapse SIM images of mitochondria from HeLa cells stained with MTG. Note that, at ∼4 s, the smaller mitochondrion on the left, appears to extend a finger-like projection toward the mitochondrion on the right (blue circle). At subsequent time points, the membranes of the two mitochondria merge into a single network. Scale bars = 500 nm. (B) Zoomed-in region of finger-like extension (blue arrowhead) from (A). Red scale bar = 100 nm.

Figure 12.. Machine-learning approach using Trainable Weka Segmentation can segment mitochondrial cristae in living cells.

Approach for quantifying live-cell imaging of cristae using the LSM 880 with Airyscan. (A) Image of cristae (blue arrowheads) in HeLa cells stained with 10-N-nonyl acridine orange (NAO). Scale bar = 500 nm. (B) Cristae probability map depicting areas likely to be cristae (white pixels) versus background (black pixels). Blue arrowheads denote cristae. (C) Thresholding of the cristae probability map in (B). Selected cristae are shown in red (denoted by blue arrowheads). (D) Binary mask resulting from application of thresholding step in (C). Blue arrowheads denote cristae (black pixels). (E) Regions of interest (ROIs) marked by yellow borders around cristae (black pixels). Blue arrowheads denote cristae. (F) Superimposition of ROIs onto original image in (A). Subsequent measurement yields data relating to various parameters, such as cristae density, area, and shape.

Figure S5.. Images of Trainable Weka Segmentation (TWS) window during training of classifier.

(A) TWS window showing training of classifier on two classes of objects, “Cristae” and “Background.” Note that the traces marking cristae are red lines (denoted by blue arrowheads); and the traces marking background are green lines (denoted by white arrowheads). (B) Overlay of trained classifier, showing early step in training process, where cristae are marked as reddish brown areas (blue arrowheads) and background is marked as greenish areas (white arrowheads).