Microscopic Image filaments' Curvature Extractor using Neural Networks in live cells – MICENN

Microtubules are semi-flexible and dynamic filaments that play major roles in vivo. During mitosis, they appear highly curved compared to in vitro. It suggests that microtubule bending rigidity may play a role and be regulated. While it is established that microtubule dynamics contribute to cell division robustness, their mechanics are yet to be investigated. Forces and tension are proven essential for faithful cell division, ensuring correct chromosome attachment, e.g. So, an optimal microtubule-rigidity may exist. Rigidity regulation was investigated in neuronal cells, where microtubules are stable. Proposed mechanisms encompass bundling, protofilament coupling or post-translational modifications. In dividing cells, these mechanisms are likely combined with additional ones involving lattice defects or mechanical coupling with the crowded cytoplasmic environment. The proposed project explores these mechanisms in dividing cells, first in the established-model nematode zygote and then in cultured human cells. We foresee that they participate in ensuring a faithful cell division.
Microtubule flexural rigidity was investigated in vitro, taking advantage of bright labelling, isolated microtubules and controlled bending. However, live images are low-contrasted and microtubule networks dense. We propose taking advantage of recent deep learning progress to create an in vivo mapping of microtubule curvature as a proxy to rigidity. We will input images of fluorescently labelled microtubules. In particular, we will directly extract geometrical parameters, the curvature, from semi-super-resolved live images (Airyscan). It contrasts with the usual two steps: segmenting the filaments then computing the longitudinal curvatures from their positions. Indeed, segmenting step requires well-contrasted images, not feasible in vivo. We will generate a filament-curvature map by solving the corresponding pixel-wise continuous labelling problem.
Our tool, called MICENN, will first extract filament curvatures in 2D images. In the axisymmetric C. elegans embryo, it is enough to offer a complete view of the network. To design it, we will benchmark the architecture among GAN-based, U-Net variants or combinations using in silico images, as annotating microtubule images appears too time-consuming to create a learning set. We will generate these synthetic images using Cytosim and confocalGN tools. We will then implement a cross-training strategy using simulated images, followed by transfer learning or fine-tuning on a few semi-automatically annotated fixed-cell or live images.
Combining 2D-MICENN with functional studies established in the lab, we will ask whether and how microtubule mechanics participate in cell division. We will correlate microtubule-rigidity readout (microscopic level) to cell-scale phenotypes using the C. elegans embryo – a highly appropriate model by its stereotyped mitotic choreography. We will then investigate the regulatory mechanisms.
Since the cell environment plays an essential role in shaping the microtubule network, we foresee that MICENN will be instrumental in extracting filament curvatures in 3D images, tissues or organoids. We will stay to cultured cells in this proposal for simplicity. Merely extending 2D-MICENN will likely be insufficient. We will again benchmark architectures of 3D neural networks, still using synthetic images and cross training. Importantly, we will consider block processing to save memory and processing costs. As a use-case, we will apply 3D-MICENN to human cultured cells and challenge the nematode findings about microtubule-rigidity regulation.
With two additional applications of MICENN to non-microtubule cases, using published images, particularly of blood vessels in the eye fundus, we will showcase MICENN universality and suitability. We will document and disseminate this unique software to deep learning experts and novices to support original research involving linear structures and even diag