Section: New Results

Using Deep Learning and Generative Adversarial Networks to Study Large Scale GFP Screens

Fluorescent imaging of GFP tagged proteins is one of the most widely used techniques to view the dynamics of proteins in live cells. By combining it with different perturbations such as RNAi or drug treatments we can understand how cells regulate complex processes such as mitosis or the cell cycle.

However, GFP imaging has certain limitations. There are only a limited number of different fluorescent proteins available, making imaging multiple proteins at the same time very challenging and expensive. Finally, analyzing complex screens can be very challenging: it's not always obvious a-priori what kind of features will predict the phenotypes we are interested in.

We discuss a new approach to studying large scale GFP screens using deep convolutional networks. We show that by using convolutional neural networks, we can greatly outperform traditional feature based approaches at different kind of prediction tasks. The networks learn flexible representations, which are suitable for multiple tasks, such as predicting the localization of Tea1 in fission yeast cells (blue signal, shown in image) in cells where only other proteins are tagged.

We then show that we can use generative adversarial neural networks to learn highly compact latent representations. Those latent representations can then be used to generate new realistic images, allowing us to simulate new phenotypes, and to predict the outcome of new perturbations (joint work between Federico Vaggi, Anton Osokin, Theophile Dalens).