Machine learning helps reveal cells' inner structures in new detail

Machine learning helps reveal cells' inner structures in new detail

Open any introductory biology textbook, and you’ll see a familiar diagram: A blobby-looking cell filled with brightly colored structures – the inner machinery that makes the cell tick. 

Cell biologists have known the basic functions of most of these structures, called organelles, for decades. The bean-shaped mitochondria make energy, for example, and lanky microtubules help cargo zip around the cell. But for all that scientists have learned about these miniature ecosystems, much remains unknown about how their parts all work together. 

Now, high-powered microscopy – plus a heavy dose of machine learning – is helping to change that. New computer algorithms can automatically identify some 30 different kinds of organelles and other structures in super high-resolution images of entire cells, a team of scientists reports in the journal Nature

The detail in these images would be nearly impossible to parse by hand throughout the entire cell, says the study lead. The data for just one cell is made up of tens of thousands of images; tracing all a cell’s organelles through that collection of pictures would take one person more than 60 years. But the new algorithms make it possible to map an entire cell in hours, rather than years. 

“By using machine learning to process the data, we felt we could revisit the canonical view of a cell,” the author says. 

In addition to two companion articles in Nature, the scientists also released a data portal, OpenOrganelle, where anyone can access the datasets and tools they’ve created.

These resources are invaluable for scientists studying how organelles keep cells running, says, a senior group leader. “What we haven’t really known is how different organelles and structures are arranged relative to each other – how they’re touching and contacting each other, how much space they occupy,” the author says. 

For the first time, those hidden relationships are visible.

A graduate student had previously developed machine learning tools that could pinpoint synapses, the connections between neurons, in electron microscope data. For COSEM, the student adapted those algorithms to instead map out, or segment, organelles in cells. 

The segmentation algorithms worked by assigning each pixel in an image a number. The number reflected how far the pixel was from the nearest synapse. An algorithm then used those numbers to ID and label all the synapses in an image. The algorithms work in a similar way, but with more dimensions. They classify every pixel by its distance to each of 30 different kinds of organelles and structures. Then, the algorithms integrate all of those numbers to predict where organelles are positioned. 

Using data from scientists who have manually traced organelle boundaries and assigned numbers to pixels, the algorithm can learn that particular combinations of numbers are unreasonable, the author says. “So, for example, a pixel can’t be inside a mitochondrion at the same time it’s inside the endoplasmic reticulum.” 

To answer questions like how many mitochondria are in a cell, or what their surface area is, the algorithms need to go even further, says group leader. The team built algorithms that incorporate prior knowledge about organelles’ characteristics. For example, scientists know that microtubules are long and thin. Based on that information, the computer can make judgments about where a microtubule begins and ends. The team can observe how such prior knowledge affects the computer program’s results – whether it makes the algorithm more or less accurate – and then make adjustments where necessary.  

After two years of work, the team has landed on a set of algorithms that generate good results for the data that have been collected so far. Those results are important groundwork for future research.