Visualizing Cellular Context: Improved Interpretability through Latent Embeddings of Cancer Cell Morphology

A cell's morphology contains a wealth of information about a cell such as its health, function, or behavior. While a given snapshot of morphology may convey different possibilities of what a cell is experiencing, tracing how the morphology changes over time can begin to clarify the state of the cell. As an example. perimeter changes may indicate cytoskeletal reorganization. Tracking different morphological features also begins to complicate the picture since the features are not all independent. Dimensionality reduction can distil the features to structurally relevant components, but may also leave a complicated relationship between the reduced components and the morphological features. Motivated by the fact that 13 morphological features have been shown to predict the four most significant morphological phenotypes of MDA-MB-231, consistent with their actin cytoskeletal organization, with an autoencoder we were able to encoded this feature space into a two dimensional space of structurally relevant morphological quantities. We then investigate the degree of interpretability of this space for cells in different situations. For a spheroid invasion, we map invasion related information to this space to see the motility patterns of different morphologies. While for a drug treatment, we track the temporal behavior of the morphologies to understand the pharmacological impacts on the cell shape. We also apply the model to histological samples to understand how different stages of metastasis are reflected in the space, and if any novelty detection methods can detect anomalies in the morphological point clouds.