(E-175) Cell shape-force coupling in dynamic fibrous environment computed via novel label-free deep learning enabled technique

Introduction:: Contractility driven cell migration is pivotal to many physiological processes such as wound healing, angiogenesis and tissue morphogenesis while also playing crucial role in disease progression such as cancer metastasis. This is facilitated through force induced remodeling of the fibrous extracellular matrix which in turn, dynamically modulates cell shape in feedback process. Mapping the deformation field is impeded by cells occluding the matrix and additionally, the time dependence of the matrix properties poses a challenging problem for accurate force estimation. To quantify the extracellular matrix-mediated dynamic cell shape-force interplay, we leveraged our previously established crosshatch fiber networks generated through STEP technique (Spinneret based Tunable Engineered Parameters) capable of modulating cell shapes. We integrate it with machine learning(ML) and inverse formulation of fiber mechanics to generate a label free approach. The underlying fibrous deformations are captured through deep learning based pix2pix model that alleviates deleterious photobleaching/phototoxicity effects and cells removal from fibers for deformations mapping required in most other force reconstruction methods. The displacement fields thus obtained are used to solve an inverse formulation using beam mechanics. The discrete planar architecture and non-linear mechanics make the problem ill-posed which required development of a new framework as existing models are not applicable as is. Overall, we advance ML by tracking deformable backgrounds and biology by quantifying forces of single-cell differentiation and force-dipoles as migrating cells change shape or encounter other cells, as expected in vivo.


Materials and Methods:: Cell Culture and Fiber Manufacturing: Human bone marrow derived mesenchymal stem cells derived were used cultured at 37˚C and 5% CO₂. Crosshatch fiber scaffolds made of polystyrene, were generated using our previously STEP technique. Fibers were made of polystyrene with ~250nm in diameter and interfiber spacing of ~25µm for square grids and ~25µm in X-direction and 5-8µm in Y-direction for rectangular grids. Imaging was done with AxioObserver-Z1 with 20x objective. Fibers were tagged with rhodamine fibronectin to generate fluorescent images for training of pix2pix. Adipogenic and Osteogenic media cocktails were used to induce differentiation.

Deep Learning: The deep learning based pix2pix model (conditional Generative Adversarial Network) was trained on sets of pristine phase contrast and fluorescent images obtained experimentally. RetinaNet in conjunction with a bi-directional tracking algorithm was used to detect grid intersections from the synthetically generated fluorescent images. All codes were written in Python.

Computational Mechanics: For the finite element(FE) modeling of nanofiber mechanics, the nanofibers are modeled as beams undergoing elastic deformations with pretension generated from the manufacturing process. We used corotational beam formulation that accounts for the geometrical nonlinearities due to large rigid body motions and small elastic strains in a corotating frame. The inverse problem is formulated as a regularization-based minimization problem constrained by the nanonet mechanics. The inverse problem solver minimizes the difference between measured and FE model deformations by tweaking the unknown nodal forces. Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithm is used for minimization of objective function. All codes were written in MATLAB 2019.


Results, Conclusions, and Discussions:: Results:

The overall schematic of crosshatch Nanonet Force Microscopy (cNFM) is shown in (Fig1). Cells on STEP fibers apply forces to remodel the underlying fibrous network which in turn affects the cell shapes in feedback loop (Fig1A). Both phase contrast and fluorescent images are generated to train the image translation pix2pix model. The model then predicts the fluorescent images which contain the fiber information, directly from phase contrast images. Grid intersections are tracked using a RetinaNet model in conjunction with a bidirectional tracking algorithm. Displacement fields are calculated using the extracted nodal and forces are computed from obtained displacement fields.

We demonstrate cNFM efficacy by calculating the force reduction as well as subsequent recovery by addition and removal of Y27632 drug(Fig1B). We next observed that stem cells differentiating toward adipogenic lineage exerted lower forces while cells differentiating toward osteogenic lineage exerted higher forces (Fig1C). On hybrid fiber networks, cells transitioned from spread shape on square regions to elongated shapes on rectangular regions (Fig1D). As the cells migrated across regions, the migration cycle entailed low and high contractility. We also captured the force dynamics between interacting cells as well as dividing cells.

Discussion:

Our novel automated cNFM technique has several advantages: 1.ECM mimicking fibers allows cell shape and focal adhesion arrangement control 2.Fluorescent-free method alleviates the deleterious phototoxicity/photobleaching effects 3. Automated ML routines that account for deformation of the background, thus, allowing us to interrogate force feedback loops in cell migration. The limitations we envision to overcome in the future, include generating fiber profiles underneath fully opaque cells that can occlude the background fiber grids using masked autoencoders.

Conclusions:

We present a new fiber-based cell-force computation technique in which the complex fiber architecture is reproduced using deep learning. Cell shapes and the associated force polarity in migration is controlled through fiber architecture, facilitating interrogation of force feedback loops in cell-fiber force and shape coupling. Deformable background tracking advances ML, inverse formulation advances fiber mechanics and quantifying forces of single-cell differentiation and force-dipoles in interacting and diving cells provide new biological insights.


Acknowledgements (Optional): :

References (Optional): :