Brightfield Is Back: A 17th Century Cell Imaging Technique is Making a Comeback Thanks to Machine Learning

A rudimentary method of detecting cells in samples using light called brightfield microscopy – first developed in the 17th century – is proving to be a powerful tool for analyzing changing cell states for tech-enabled drug discovery thanks to advances in machine learning. And unlike fluorescence-based approaches that are typically used for phenomics observations, including Cell Painting, brightfield microscopy does not perturb the cells, allowing scientists to examine multiple layers of biology on the same sample – a game-changer when it comes to improving cost, speed, and experimental consistency.???

With a brightfield microscope, white light is transmitted through a cell sample onto a detector.?To the human eye, the resulting image looks like a vague collection of gray blobs and black dots – nowhere near as detailed as the multi-colored Cell Painting images that Recursion scientists had been using prior.?

But during the company’s 2021 “Hack Week” – when employees are given free reign to test out-of-the-box theories – a team at Recursion decided it was a method worth exploring. They reasoned that brightfield would require fewer steps, adding to the speed and efficiency of the high throughput process. The same cell plates could be used for multiple imagings, and it would remove human bias from the equation entirely.?

Although brightfield images appear simplistic, the team suspected that the information was in there – and Recursion’s machine learning models could be trained to extract that information that was undetectable to the human eye.?

In the words of Charles Baker , VP of Technology Development, “It’s not just that humans couldn’t analyze these images, but that it is only now with modern machine learning approaches that you can unlock the underlying value. When you train these models, it turns out that the information is present in brightfield images.”

Using Brightfield to Predict Cell Painting Output

Recursion’s earlier method of gathering cell imaging data for training models, Cell Painting, involves staining cells with fluorescent dyes in order to capture biological changes to the cell known as phenomics. Our automated wet lab can run over 2 million experiments per week, and we’ve built one of the largest phenomics datasets in the world using primarily this method.?

To test whether brightfield could produce similar usable data, scientists developed a curated dataset of Cell Painting and brightfield images from the same time point. They then trained a small vision transformer on top of our latest phenomics foundation model to reconstruct Cell Painting images from their corresponding brightfield images. When given a brightfield image as an input, they found that the model could predict the equivalent Cell Paint output with near-perfect accuracy. Additionally, they found that similar relationships between perturbations were recovered when comparing Cell Painting experiments to brightfield experiments.?

And it gets better, says Imran Haque , Senior VP of AI and Digital Sciences. “This model was only trained on three cell types, but when we ran it on cell types it was not trained on, it worked on those as well.”?

To adapt the wet lab to brightfield imaging, Baker says, “We’ve added the ability to incubate our cells during the imaging process.” With brightfield, he adds, “we can take images of living cells multiple times to observe dynamics which can be used for inferring causal relationships.”

Once scientists capture the phenomics of these living cells – or biological changes happening under different conditions – they can run these same cells through additional assays.

“Since we don’t kill the cells, we can move them to other assays and don’t have to duplicate the experiment,” says Jordan Christensen , SVP of Technology. Once they’ve undergone phenomics testing, scientists can test for transcriptomics, or RNA transcript activity. “In order to do transcription, if we had to redo the experiment it would be less useful than using the same cells,” he adds.?

An upcoming paper will delve further into the specifics of one of the modeling breakthroughs that is enabling brightfield. In the meantime, we are accelerating the transition to brightfield as the company’s primary method for extracting information from cells – making us likely one of the only companies exploring its use in AI drug discovery at scale.

Author: Brita Belli

#brightfield #cells #tech #techbio #science #innovation #discovery #drugdiscovery #ai #ml