Quantitative comparison of principal component analysis and unsupervised deep learning using variational autoencoders for shape analysis of motile cells

Abstract

Cell motility is a crucial biological function for many cell types, including the immune cells in our body that act as first responders to foreign agents. In this work we consider the amoeboid motility of human neutrophils, which show complex and continuous morphological changes during locomotion. We imaged live neutrophils migrating on a 2D plane and extracted unbiased shape representations using cell contours and binary masks. We were able to decompose these complex shapes into low-dimensional encodings with both principal component analysis (PCA) and an unsupervised deep learning technique using variational autoencoders (VAE), enhanced with generative adversarial networks (GANs). We found that the neural network architecture, the VAE-GAN, was able to encode complex cell shapes into a low-dimensional latent space that encodes the same shape variation information as PCA, but much more efficiently. Contrary to the conventional viewpoint that the latent space is a "black box", we demonstrated that the information learned and encoded within the latent space is consistent with PCA and is reproducible across independent training runs. Furthermore, by including cell speed into the training of the VAE-GAN, we were able to incorporate cell shape and speed into the same latent space. Our work provides a quantitative framework that connects biological form, through cell shape, to a biological function, cell movement. We believe that our quantitative approach to calculating a compact representation of cell shape using the VAE-GAN provides an important avenue that will support further mechanistic dissection of cell motility. AUTHOR SUMMARYDeep convolutional neural networks have recently enjoyed a surge in popularity, and have found useful applications in many fields, including biology. Supervised deep learning, which involves the training of neural networks using existing labeled data, has been especially popular in solving image classification problems. However, biological data is often highly complex and continuous in nature, where prior labeling is impractical, if not impossible. Unsupervised deep learning promises to discover trends in the data by reducing its complexity while retaining the most relevant information. At present, challenges in the extraction of meaningful human-interpretable information from the neural networks nonlinear discovery process have earned it a reputation of being a "black box" that can perform impressively well at prediction but cannot be used to shed any meaningful insight on underlying mechanisms of variation in biological data sets. Our goal in this paper is to establish unsupervised deep learning as a practical tool to gain scientific insight into biological data by first establishing the interpretability of our particular data set (images of the shapes of motile neutrophils) using more traditional techniques. Using the insight gained from this as a guide allows us to shine light into the "black box" of unsupervised deep learning.