GANs for Biological Image Synthesis

[gwaybio]

https://arxiv.org/abs/1708.04692

In this paper, we propose a novel application of Generative Adversarial Networks (GAN) to the synthesis of cells imaged by fluorescence microscopy. Compared to natural images, cells tend to have a simpler and more geometric global structure that facilitates image generation. However, the correlation between the spatial pattern of different fluorescent proteins reflects important biological functions, and synthesized images have to capture these relationships to be relevant for biological applications. We adapt GANs to the task at hand and propose new models with casual dependencies between image channels that can generate multichannel images, which would be impossible to obtain experimentally. We evaluate our approach using two independent techniques and compare it against sensible baselines. Finally, we demonstrate that by interpolating across the latent space we can mimic the known changes in protein localization that occur through time during the cell cycle, allowing us to predict temporal evolution from static images.

https://github.com/aosokin/biogans

cc @brettbj for pointing this one out

[agitter]

This could fit alongside #372 if we want to add it.

[gwaybio]

General Summary

The authors develop 2 distinct GAN architectures (DCGAN from Radford et al. and their own "separable" generator) and test their method with 2 different losses (Jensen-Shannon and Wasserstein) (4 total GANs). The objective is to study protein co-localization of yeast cell growth over time using 2-channel fluorescent microscopy images of yeast cells. The red channel always tags Bgs4, a marker of active cell growth, while the green channel tracks 1 of 44 different polarity factors. The authors train their models using 6 factors.

Biological Aspects

LIN Dataset - 170,000 images of individual yeast cells (S. pombe) with red and green color channels tracking 44 different polarity factors during cell growth. The authors train using 26,909 images and 6 factors. Goal is to use factors to observe protein localization throughout growth. Other goals include infering multi-channel images with more than 2 proteins and interpolating through time to understand how proteins move throughout cell growth.

Computational Aspects

Nice description of GANs and how they must be modified to microscopy domain. Their separable model has one-way connections from the red channel to the green channel and therefore learns the green protein (polarity factors) dependence on growth phase. They evaluate their method using classifier two-sample tests (C2ST) and assess held out sample reconstruction. They additionally evaluate their method by running two held out proteins through their model, applying C2ST, and observing better scores for predicting similar functioning proteins. The authors end by interpolating through points in their learned latent space to simulate cell-cycle progression. They validate their interpolation through the known movements of Arp3. Very cool movies of this interpolation are provided in the repository README https://github.com/aosokin/biogans