Artificial intelligence is helping researchers gain a more complete understanding of cell biology by untangling complex data from multiple measurement techniques. Scientists often rely on different methods—such as measuring gene expression, protein levels, or cell structure—to study a cell’s state, especially in diseases like cancer. However, each method captures different aspects of the cell, and combining them traditionally requires time-consuming, manual analysis that can obscure the bigger picture. To solve this, researchers from the Broad Institute of MIT and Harvard, along with ETH Zurich and the Paul Scherrer Institute (PSI), developed a new AI-driven framework that intelligently separates shared information across measurement types from data unique to each method. This allows scientists to see not only what is happening in a cell but also where that information comes from—such as which part of the cell or which biological process is responsible. The framework uses a machine-learning approach that differs from standard methods. Instead of treating each measurement type independently, it creates a shared representation space for overlapping data and separate spaces for modality-specific information. Think of it as a dynamic Venn diagram for cellular data, where the intersections show common signals and the non-overlapping areas highlight unique insights. Lead author Xinyi Zhang, formerly a graduate student at MIT’s Department of Electrical Engineering and Computer Science and now a group leader at AITHYRA in Vienna, explains that this method allows researchers to input data from multiple sources and automatically determine which parts are shared and which are unique. This eliminates the need for repeated experiments and enables faster, more accurate interpretation of cellular states. The model was trained using a two-step process that helps it handle the complexity of distinguishing between overlapping and unique signals. When tested on both synthetic and real-world single-cell datasets, it successfully identified shared gene activity between transcriptomics and chromatin accessibility, and correctly pinpointed which data came from a single modality. It also helped identify which measurement technique captured a key protein marker linked to DNA damage in cancer cells—a critical insight for clinical diagnostics. The researchers believe this tool can guide scientists in deciding which modalities to measure directly and which to predict, optimizing both cost and accuracy. Future work will focus on improving interpretability and expanding the model’s use to broader clinical applications, including neurodegenerative and metabolic diseases. Caroline Uhler, senior author and professor at MIT, emphasizes that true understanding comes not just from combining data, but from carefully comparing how different modalities reflect the same biological processes. “We can learn a lot about the state of a cell if we carefully compare the different modalities to understand how different components regulate each other,” she says. The research was supported by the Eric and Wendy Schmidt Center at the Broad Institute, the Swiss National Science Foundation, the U.S. National Institutes of Health, the U.S. Office of Naval Research, AstraZeneca, the MIT-IBM Watson AI Lab, the MIT J-Clinic for Machine Learning and Health, and a Simons Investigator Award.