New AI method helps researchers see the bigger picture in cell biology

Modern biology can measure a single cell in many different ways at once: which genes are active, how its DNA is packaged, which proteins it carries. Each technique captures something real, but stitching those views together is hard, because no one knows in advance which signals are genuinely shared and which belong to just one method. Researchers at the Broad Institute of MIT and Harvard, working with collaborators at ETH Zurich and the Paul Scherrer Institute, have built an artificial-intelligence framework that addresses exactly this problem, in work published on February 25, 2026.

The method learns to separate information that is captured by a single measurement modality from information that is shared across several of them. In tests on synthetic datasets, where the right answer was known, the framework correctly identified which signals were shared and which were modality-specific. Applied to real data, it distinguished gene activity that was jointly captured by transcriptomics and chromatin-accessibility measurements, and pinpointed which technique had detected specific protein markers indicating DNA damage.

“Each technique captures something real, but stitching those views together is hard, because no one knows in advance which signals are genuinely shared and which belong to just one method.”

The practical payoff is efficiency and clarity. "By putting the information from all these measurement modalities together in a smarter way, we could have a fuller picture," said lead author Xinyi Zhang. Senior author Caroline Uhler, an MIT professor, framed the deeper question the tool can answer: "Which modalities should we measure and which should we predict? Our method can answer that." That guidance could spare labs from running expensive, redundant experiments and help them choose the most informative measurements.

The researchers are careful about limits. They plan further work to make the cellular information the model surfaces more interpretable, and additional experiments to confirm the framework truly disentangles the data before it is used in clinical settings. They also note that the approach was validated first on synthetic data, where ground truth is known, and that extending the same confidence to messy real-world measurements will take more testing across different tissues and disease states. Published in Nature Computational Science, the study is a reminder that some of AI's most useful contributions are quiet ones: not replacing biologists, but helping them ask sharper questions, spend their limited budgets on the most informative experiments, and waste less effort on the long, patient path toward understanding disease.