Convolutional neural network segmentation of bare and decorated actin filaments in cellular cryo-tomography

Developing neurons have a highly dynamic structure at the tip of growing axons, the growth cone, that guides the neuron to its synaptic target. Rapid remodeling of actin filaments and bundles is vital to the movement of these structures. Recently, an actin binding protein that canonically plays a severing role in filaments, cofilin, has been shown to play a role in filament stabilization and flexibility within growth cones when fully decorated at a 1:1 ratio. Decorated filaments, called cofilactin, are visibly distinct from f-actin within cryo-electron tomograms due to the helical pitch tightening that cofilin induces along the length of the decorated portion. To better characterize the spatial arrangement and role of cofilactin within growth cones, we are training convolutional neural networks (CNNs) to identify and segment f-actin, cofilactin, and the transition zones between decorated and undecorated regions within cryo-tomographic data. Due to the low signal-to-noise ratio and native image aberrations of cryo-electron tomography (cryo-ET), distinguishing filaments accurately by hand within very noisy data is both tedious and time consuming. Actin networks in growth cones can be densely packed and annotating a single tomogram can take upwards of a week for an expert hand segmenter. Convolutional neural networks excel at distinguishing true signal from noise. With proper training, it is feasible that a CNN, in this case a U-Net, will be able to segment filaments rapidly and accurately within growth cones. We have used simulated tomographic data to create ‘perfect’ ground truth training sets and generated trained U-Nets that can accurately identify cofilin and actin in both in vitro and cellular tomograms.