AI tool predicts potential drug targets by analyzing cell images

A new AI model developed by researchers at the Eric and Wendy Schmidt Center at the Broad Institute and ETH Zurich’s Department of Health Science and Technology can identify genes that have been altered, such as ones that might be causing a disease, in a cell just by analyzing an image of the cell’s chromatin — the dense package of chromosomes inside the cell’s nucleus. The machine learning tool, called Image2Reg, promises to accelerate both research on the genetic causes of disease and drug discovery by predicting drug targets and mechanisms. 

In a study published recently in Cell Systems, a team led by longtime collaborators Caroline Uhler of the Schmidt Center and GV Shivashankar of ETH Zurich describe how their model can make predictions of genetic perturbations it has never encountered before. The scientists say that by bridging the gap between imaging data and molecular biology, Image2Reg offers a simple, rapid, and inexpensive alternative to more traditional sequencing-based approaches for mapping how cells respond to genetic or chemical changes. 

“This is a prime example of what we aim to do at the Schmidt Center,” said Uhler, who is director of the Schmidt Center and the Andrew and Erna Viterbi Professor of Engineering in the Department of Electrical Engineering and Computer Science and Institute for Data, Systems, and Society at MIT. “By combining machine learning with biology at scale, we can reveal new layers of information from data that’s already being collected — and use it to guide therapeutic development.”

“As we continue to explore the potential of chromatin imaging, we’re seeing that it can serve as a powerful window into the regulatory state of the cell,” said Shivashankar, full professor at ETH Zurich and head of the Laboratory of Nanoscale Biology at the Paul Scherrer Institute.

How Image2Reg works

Disruptions in gene networks can lead to many diseases such as cancer and neurodegeneration. Uhler and Shivashankar have long been interested in the structure of chromatin because it can influence how these gene networks are regulated and contribute to disease. Fortunately for scientists, they have a rapid and inexpensive way to study chromatin: using fluorescent dyes to stain the chromatin and microscopes to capture images of it. 

In previous work, Uhler and Shivashankar have shown that simple chromatin staining images combined with machine learning algorithms can yield a lot of information about the state and fate of a cell in health and disease. While machine learning has helped identify cell states, researchers hadn’t been able to trace them back to specific genes or regulatory programs. 

To make these connections, the research team designed Image2Reg to learn from two types of data: chromatin images of cells with known genetic or chemical perturbations, and molecular profiles of gene interactions. First, the model uses a convolutional neural network to learn how different perturbations change chromatin structure. Then, it uses a graph-based model to learn how genes relate to each other in a specific cell type, using transcriptomics and protein-protein interaction data. Finally, a third component aligns these two embeddings, effectively translating between the physical organization of DNA and its biochemical regulation.

By successfully aligning chromatin structure with gene regulatory function, Image2Reg confirms a strong, predictive link between how DNA is physically organized and how genes behave — a connection that could help explain how diseases take hold at the molecular level.

This alignment enables Image2Reg to infer which genes are perturbed in new images it has never seen before. “By learning to map between representations of cell images and genes, our model can generalize to unseen perturbations, and that’s what makes it so powerful,” said Adityanarayanan Radhakrishnan, co-first author of the new study and postdoctoral fellow at the Schmidt Center.

Predicting drug effects

To test Image2Reg’s ability to generalize, the team trained it on chromatin images of cells that each had one gene turned off or turned up to a high level. These image datasets — such as Cell Painting data from the Carpenter-Singh lab and perturbation screens from the Broad’s Cancer Dependency Map and JUMP-Cell Painting Consortium efforts — were generated at the Broad Institute and provided a rich foundation for training and validating the model. The model was able to predict the genetic targets of these drugs with 60 percent accuracy, even when it had never seen those compounds before. 

“These results show that chromatin images can reveal how a compound affects the cell,” said Daniel Paysan, a co-first author, postdoctoral fellow at Novartis, and former Schmidt Center visiting PhD student. “We’re essentially using imaging to understand which genes a drug is targeting.”

While this study focused on a single cell type and specific perturbation conditions, the researchers say the approach is broadly applicable. As large-scale optical perturbation screens become more common, Image2Reg could be adapted to other experimental contexts — enabling scientists to study how gene regulation shifts in different cell types, disease states, or treatment responses.

Ultimately, the team hopes Image2Reg will become a foundation for linking chromatin structure to gene function at scale — helping researchers uncover the molecular mechanisms underlying disease and identify the genes that could be most effective to target with new or existing treatments.