3D single-cell shape analysis of cancer cells using geometric deep learning

Matt De Vries, Lucas G Dent, Nathan Curry, Leo Rowe-Brown, Adam Tyson, Chris Dunsby, Chris Bakal

Keywords: 3D, geometric, deep learning, single-cell, shape, cancer

TL;DR: We utilise graph neural networks to learn shape representation of 3D metastatic cancer cells and apply this to drug prediction as well as show the potential for drug screening.

Abstract: 3D cell shape is linked to disease. Despite this, there is a paucity of methods to quantify 3D cell shapes. Current methods utilise pre-defined measures of cell geometry as opposed to data-driven approaches. Furthermore, we do not fully understand the cell shape landscape of cancer cells in 3D and the different shape classes that exist in a dataset or how these are correlated with drug treatments. To address this, we have developed a geometric deep learning method to learn 3D cell shape features and classes simultaneously. This technique combines an existing dynamic graph convolutional neural network encoder with a foldingnet decoder and improved deep embedded clustering. We use this model to learn meaningful shape representations of over 70 000 drug-treated melanoma cells imaged by 3D light-sheet microscopy. We propose describing cells and treatments by a 3D quantitative morphological signature, representing a cell's similarity to the learned shape classes in the dataset. This led to the insight that drugs treated with similar inhibitors share morphological signatures, which can be used to predict the activity of a drug.