Visual Proteomics of Neuronal Growth Cones: Deep Learning Segmentation for Large Scale Analysis of Cryo-Electron Tomograms
Restricted (Penn State Only)
- Author:
- Heebner, Jessica
- Graduate Program:
- Biomedical Sciences
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- July 20, 2023
- Committee Members:
- David Degraff, Co-Chair & Outside Field Member
Richard Mailman, Outside Unit Member
Matthew Swulius, Chair & Dissertation Advisor
Maria Bewley, Major Field Member
Fang Tian, Major Field Member
Lisa Shantz, Program Head/Chair - Keywords:
- cryoEM
cryoET
Deep Learning
Image Segmentation
Growth Cones
Proteomics
Actin
Cofilactin
TriC/CCT
Tomography - Abstract:
- The “Resolution Revolution” of cryo-electron microscopy (cryoEM) is well and truly underway. From sub-atomic molecular structures to correlative light and electron microscopy studies, cryoEM has bridged the resolution gap and opened a window into an entirely new world for in situ biology. This makes it an ideal technique with which to study neuronal growth cones. Discovered over a century ago, these highly dynamic structures at the tips of outgrowing neurites are the guidance machinery of the developing neuron. They are responsible for synthesizing a host of different extracellular signals to steer neurons huge distances to appropriate synaptic targets. A complicated crosstalk of diffusible guidance molecules, extracellular matrix ligands, cellular adhesion molecules, cytoskeletal dynamics, and second messenger cascades coordinate this incredibly intricate task. This dissertation aims to create and apply new tools to contribute to a visual proteome of the growth cone by pairing cryo-electron tomography with deep learning image analysis to describe the spatial arrangements of large macromolecules within each region of the growth cone. Because the gold standard method of expert hand segmentation for cryoEM data analysis is tedious, time-consuming, and labor-intensive, we chose to adapt and improve existing deep learning models for our needs. Deep learning models, specifically convolutional neural networks (CNNs), excel at image analysis and numerous architectures exist for a myriad of tasks, including a handful of tools designed specifically for cryoEM data, but no existing tool could accomplish the type of large-scale segmentation of multiple features that this project required. With this in mind, a workflow for training a U-Net was created as this CNN architecture was originally developed for bioimaging analysis and can be trained with a very small amount of data. Training any CNN requires high quality training data. Unfortunately, due to the low signal to noise and imaging aberrations inherent to cryoET, obtaining this high-quality data is a challenging and limiting first step. To help address this limitation and reduce the requirement for an expert segmenter, our workflow includes simulated cryoET images as part of the training data. The simulation software we developed and present below allows these images to be generated and fully segmented in a matter of minutes, dramatically improving the amount and quality of training data with very little effort required. Using this simulated data paired with a small amount of real data, we trained a 2.5D U-Net to simultaneously segment nine different features in cryo-tomographic images: membrane, f-actin, cofilactin, microtubules, ribosomes, TriC/CCT, fiducial markers, carbon edge, and background. The trained network, NeuroSeg 1.0, was then applied to four years of compiled growth cone tomograms and the segmentations were quantified and analyzed. The dataset consists of 106 tomograms that represent 20 individual cells. Tomograms were each assigned to a region of the growth cone based on gross cytoskeletal morphology, allowing us to quantify protein distribution per region. Segmentations revealed region-specific distributions of f-actin, cofilactin, ribosomes, TriC/CCT, and microtubules. Interestingly, we found that TriC/CCT favors self-association with other TriC particles in the growth cone, a behavior that is reminiscent of polyribosomes. This novel behavior for the chaperonin has not been described before in the literature and was heavily featured in all images of the growth cone. This large-scale analysis of growth cone tomographic data is the first of its kind both in scope and in scale. It represents a significant leap forward in our ability to analyze cryo-electron tomography data quickly and is the start of a comprehensive visual proteome of neuronal growth cones. Future directions will include adding more data from new tomograms, collecting larger fields of view with montage tomography to give more cellular context, and determining the mechanism underlying TriC/CCT clustering in growth cones.