Deep Learning Tool Empowers Biologists, Speeds Screening

June 25, 2018

By Allison Proffitt

June 25, 2018 | 2018 Bio-IT World Best Practices Award-Winner | Traditional high content screening is labor- and time-intensive, and many groups are working to apply deep learning to image analysis. Genedata worked with AstraZeneca to validate a deep learning workflow tool for biologists that can be “unsupervised”—it works without re-engineering across different assay types and settings. The work earned the pair a Bio-IT World Best Practices Award in the Informatics & Knowledge Management category.

Genedata has been developing an upcoming deep learning solution—Imagence—that will automate data workflows and accelerate phenotypic image analysis to empower biologists to control the screening process without a bioinformatician’s help. Genedata started with TensorFlow and its wrappers, and designed the network from scratch. The solution was engineered to work on commodity hardware: in this case, one Intel Xeon processor with 4 cores, 128 GB RAM, 500 GB SSD hard drive and a gaming card.

The deep learning workflow has three parts—none of which require a bioinformatician, explained Stephan Steigele, Genedata’s head of science.

First, biologists design their assay, and produce and review images. The biologist identifies distinct “clouds” of data points originating from the same phenotype class, Steigele explained. This step takes from 20 min to a few hours for the scientist to complete, depending on overall number of phenotypes and how easy a biologist can spot differences visually, he said. “This ‘supervised review’ stage needs to be done only once—input can then be applied automatically to later production analysis or applied to other assays,” Steigele said.

The biologist then curates example images per class, which serve as the training data for the neural networks. The need for curation depends on the heterogeneity of cell populations. “For instance,” Steigele said, “we have worked in situations where a manual curation step was not even required due to a quite homogeneous cell phenotype. Other cases required a more extensive curation—those had strongly mixed populations. A strong focus of our product development was rendering a very efficient curation process, such that even in complex situations a biologist can produce sufficient training data within 3-4 hours.”

Finally, the trained network can move into the production stage for fast, reliable classification of dozens of phenotypes. Genedata reports processing images from a 384-well microtiter plate (about 6,500 images) in about 50 minutes.

But the real time gains come when the workflow is applied to another assay. Imagence has been optimized to work across different assay types and settings without re-engineering, Genedata says. The company calls this automatic domain adaptation, or transfer learning. For example, the system can easily move between small and large molecules or different human donors and the adjustments take about 200 seconds.

Real Data, Real Testing

To validate the Imagence solution, Genedata worked on more than eight industry projects with real world data. Steigele emphasized the importance of working with real-world datasets, instead of only validating the tool against public benchmarking datasets. Genedata used the Broad Bioimage Benchmark Collection 022 (BBBC022, “Cell Painting Assay”) dataset, for instance, to validate its deep learning frameworks, Steigele said, but there’s a “serious problem” with only using public datasets.

“When we started developing this deep learning solution, we wanted to ensure that we delivered pharmacologically-relevant parameters. That means we needed to have some reference results,” he said. “We needed something real. We needed something that was produced in pharma in the quality that pharma expects.”

AstraZeneca was the closest of the industry partners, Steigele said, contributing a well-validated dataset of images testing compounds with the potential to block fibroblast activation in chronic kidney disease.

“It represented a very typical workflow for assay development,” explained AstraZeneca’s James Pilling, associate principal scientist. It was a good fit for high content screening work, Pilling explained, because the resulting screen would be the type applied to a number of different experiments and a number of different donor conditions. “It was a very relevant piece of data for this use.”


Genedata compared Imagence results to AstraZeneca’s findings for the dataset, looking at classification accuracy, plate statistic, compound potency results, and other metrics. We “produced in all cases a result quality with our deep learning solution that matched or exceeded those found by classical image analysis,” Steigele said.

True Value

The limits of the transfer learning haven’t been fully explored. “We pushed it looking at different cellular donors; we changed the type of treatment we were doing. Those were domain differences within the assay,” explained Pilling.

The network handles changes to the assay settings well, Steigele said. “The experiment settings can be quite different: different microscopes with different optical characteristics can make a huge difference on the resulting image quality or resolution. Limited material like human donor samples coming [for example] from the clinic can hinder extensive tuning of assay parameters like optimizing staining protocols etc.,” he said. “In classical HCS image analysis, any such change requires an extensive recalibration, which is a huge time investment. Our benchmarking results show that with deep learning we can adapt to such new settings within a few hundred seconds, by a hands-free automated re-computation.”

For a completely different assay, Pilling expects to retrain the network. “That’s where we need to do further work to test what needs to be done to adapt from one assay to a completely different assay, he said. “The expectation is that the approach that’s being taken wouldn’t do it in a completely unsupervised way, and the network would need to be retrained against the cloud of communities and classes that the new assay would produce. But then once that retraining had happened, it could be applied in an unsupervised fashion.”

Steigele is hopeful. “We have not stretched how far we can go. In my personal thinking—and we see this from initial experiments—I think we can stretch this quite far. For instance, people can go from bright field microscopy to florescence microscopy and try to translate the knowledge embedded in a trained network from bright field microscopy to florescence microscopy.”

AstraZeneca isn’t using Imagence widely yet—it’s not commercially available. But Pilling said the tool offers real advantages for working biologists.

“To me the impact of this was not necessarily about the application of deep learning to classify phenotypes from images, because I think there are lots groups doing this. For me the real impact was being able to do that, and being do it in a way that was part of sort of a “business-as-usual” workflow… as a more accessible technology for biologists.”