3D single-cell shape analysis using geometric deep learning

Abstract

Aberrations in cell geometry are linked to cell signalling and disease. For example, metastatic melanoma cells alter their shape to invade tissues and drive disease. Despite this, there is a paucity of methods to quantify cell shape in 3D and little understanding of the shape-space cells explore. Currently, most descriptions of cell shape rely on predefined measurements of cell regions or points along a perimeter. The adoption of 3D tissue culture and imaging systems in medical research has recently created a growing need for comprehensive 3D shape descriptions of cells. We have addressed this need using unsupervised geometric deep learning to learn shape representations of cells from 3D microscopy images of metastatic melanoma cells embedded in collagen tissue-like matrices. We used a dynamic graph convolutional foldingnet autoencoder with improved deep embedded clustering to simultaneously learn lower-dimensional representations and classes of 3D cell shapes from a dataset of more than 70,000 drug-treated melanoma cells imaged by high throughput light-sheet microscopy. We propose describing cell shape using 3D quantitative morphological signatures, which represent a cell's similarity to shape modes in the dataset, and are a direct output from our model. We used the extracted features to reveal the extent of the cell shape landscape and found that the shapes learned could predict drug treatment (up to 86% accuracy) and cell microenvironment, and are explainable. In particular, we found strikingly similar deep learning shape signatures between cells treated with microtubule polymerisation inhibitors and branched actin inhibitors. Finally, we implemented our methods as a Python package for ease of use by the medical research community.

Data-modeling the interplay between single cell shape, single cell protein expression, and tissue state

“…On the other hand, we argue that the ability to consistently measure the contribution of cell shape to a variety of applications and datasets indicates a strong biological signal that goes beyond the noise levels introduced by the coarse segmentation. Indeed, high resolution and 3D cell shapes provide discriminative information regarding the cell’s microenvironment (Dent et al, 2024; Segal et al, 2022), molecular organization (Mazloom-Farsibaf et al, 2023; Viana et al, 2023) and treatment (De Vries et al, 2023). Also here, future technologies are expected to have improved resolution and maybe even 3D information leading to information-rich segmentations.…”

Interpretable representation learning for 3D multi-piece intracellular structures using point clouds

“…Previous studies have introduced unsupervised representation learning approaches for cell images using autoencoders with geometric deep learning 44,45 . Our work complements these approaches in three ways: first, by incorporating the notion of orientation invariance into our intracellular structure morphologydependent framework for representation learning; second, by providing a systematic multi-task benchmark to evaluate the utility of each model that goes well beyond traditionally assessed reconstruction quality; third, by focusing our analysis on 3D multi-piece intracellular structures with complex morphology and spatial distribution.…”

“…Dynamic graphs are computed by constructing k-nearest neighbor graphs on points. We used k=20 based on previous works as a balance between computational complexity and local structure information 45 . We concatenated the cross-product of the neighbor features and input points as well as the input points themselves to the hidden representation.…”

“…We took the norm of the final vector embedding to get a rotation invariant representation. We also trained classical point cloud autoencoders with DGCNN encoders as described in Vries et al 45 , where VNLinear and VN LeakyReLU layers are replaced with edge convolutions and ReLU layers.…”

Machine learning enhanced cell tracking

“…Indeed, the advantage of time-lapse imaging is that temporal models of cell behavior can encode (or learn) the transitions of states over time (Held et al, 2010;Bove et al, 2017;Soelistyo et al, 2022;Gallusser et al, 2023). Incorporating features such as local (or collective) motion, neighbourhood embeddings or cell state classification can be used to generate rich representations (Bove et al, 2017;Driscoll et al, 2019;Andrews et al, 2021;Gradeci et al, 2021;De Vries et al, 2022;Ko et al, 2022;Malin-Mayor et al, 2022;Yamamoto et al, 2022;Viana et al, 2023). Increasingly, selfsupervised methods (such as variational autoencoders) are being used to learn explainable representations directly from the image data ( (Zaritsky et al, 2021;Soelistyo et al, 2022;Wu et al, 2022)).…”