Un-Freezing Time: Electron Microscopy-Based Reconstructions of Cardiomyocyte Contraction Dynamics Enabled by Deep Learning

Cardiomyocyte function is profoundly linked to its ultrastructure, which undergoes significant mechanical deformation during each heartbeat. Characterization of these dynamics is crucial for understanding cardiac function in health and disease. Conventional methods for imaging dynamic processes rely on light microscopy, which lacks spatial resolution to resolve nanoscopic detail. Electron microscopy offers orders of magnitude higher resolution but lacks temporal information. Here, we overcome the latter limitation by combining action potential-synchronized high-pressure freezing to capture defined time-points during the cardiomyocyte contraction and relaxation cycle (effectively adding 1 ms-resolved temporal information) with electron tomography-based reconstructions (offering ~1 nm³ spatial resolution). Deep learning-enabled processing not only greatly accelerates analyses of these images but also uncovers information that would remain inaccessible when using conventional analysis methods. Using our approach, we characterize the nanoscopic structure and dynamics of cardiomyocytes, focusing on key organelles involved in excitation-contraction coupling.

 

Cardiomyocytes were isolated from vibratome-cut living left-ventricular rabbit tissue and treated with pharmacological agents to modulate microtubule stability: colchicine (12.5 µM and 1.25 mM) for destabilization or paclitaxel (1.2 µM) for stabilization, using untreated cells as controls. Cells were high-pressure frozen at prescribed intervals post-electrical stimulation, freeze-substituted, heavy metal-stained, resin-embedded, and cut into 300 nm sections. Dual- and single-axis electron tomography resulted in ~850 nanoscopic reconstructions with an average reconstructed volume of ~2.5×2.5×0.2 µm³. Detailed 3D models were created with neural networks and custom semi-automatic segmentation tools (Figure 1). We created an accessible web-based visualization to explore and share these extensive datasets. Finally, we used neural networks to synthesize nanoscopic dynamics of virtual cells during contraction and relaxation. Our evaluation of the segmented volumes revealed that contraction is associated with increased cross-sectional axis ratio (eccentricity) of transverse tubules in control cells. While we observed a similar effect in paclitaxel-treated cells, we could not observe a significant change in cross-sectional axis ratio in colchicine-treated cells. These observations indicate a microtubule-dependent contribution to transverse tubular deformation. At the same time, there was a remarkable absence of significant effects of contractile activity or microtubule modulations on dyadic distances. This suggests that the tight control of dyadic architecture is not acutely affected by microtubule integrity. 

 

In conclusion, our study presents an innovative framework that combines action potential-synchronized high-pressure freezing, electron tomography, and deep learning-enabled processing, providing unique insights into the nanoscopic dynamics of cardiomyocytes and their modulation. We believe that understanding cardiac nanoscopic dynamics is key to advancing our knowledge of the ultrastructural foundations of cardiac health and disease.


Figure 1: 3D model of the sarcoplasmic reticulum, transverse tubules, microtubules, mitochondria, and myosin (depth-position color-coded). The reconstruction has a dimension of 2.3×2.3×0.16 µm³, here at a sarcomere length of 1.72 µm.