Deep Learning for Biology

Chapter 6. Learning Spatial Organization Patterns Within Cells

In this chapter, we shift focus from classifying high-level cell states—such as distinguishing cancerous from healthy tissue—to something more low level and foundational: understanding the spatial organization inside individual cells. Specifically, we’ll train a deep learning model to analyze microscopy images and learn where exactly in the cell different proteins are located, a task known as protein localization.

Protein localization plays a crucial role in cell biology. A protein’s position within the cell—for example, whether it’s in the nucleus or the mitochondria—often determines its function. Mislocalization of proteins is implicated in many diseases, even when the protein’s structure is normal (i.e., not mutated or altered). Thanks to modern fluorescence microscopy, we can observe a protein’s location in a cell directly, but the resulting images are often high dimensional, noisy, and hard to interpret at scale.

Unlike earlier chapters, the goal here isn’t to strictly optimize a metric like accuracy, recall, or precision on a specific classification or regression task. Instead, we’ll train a model to learn a latent representation of protein localization directly from raw microscopy images. You can think of a latent space as the model’s internal map—a compressed representation where proteins with similar localization patterns are grouped together, even without explicit labels. This approach falls under representation ...