Researchers at MIT, the Broad Institute, and ETH Zurich unveiled an AI framework that disentangles what cellular information is unique to each measurement modality and what is shared across them, promising a more complete picture of cell state. The model, published in Nature Computational Science, uses a shared-plus-private representation approach and a two-step training scheme to parse data from techniques such as transcriptomics, chromatin accessibility, protein markers, and morphology. In tests on synthetic and real single-cell datasets, the system identified cross-modal signals and modality-specific readouts, including pinpointing which assay best captures a DNA-damage protein marker in cancer samples—guidance that could streamline experiment planning and clinical measurement choices. The work, led by MIT’s Caroline Uhler with collaborators including G.V. Shivashankar and first author Xinyi Zhang, is aimed at clarifying disease mechanisms in cancer, neurodegeneration, and metabolic disorders. Funding came from the Eric and Wendy Schmidt Center, Swiss National Science Foundation, U.S. NIH, Office of Naval Research, AstraZeneca, the MIT-IBM Watson AI Lab, MIT J-Clinic, and a Simons Investigator Award.
Related article:
MultiVI: deep generative modeling for multi-modal single-cell RNA-seq and ATAC-seq
