Researchers at the University of Pittsburgh have developed a powerful new computational method called EMixed that can more precisely uncover the cellular composition of complex tissues. Understanding which cell types make up a tissue is essential for studying how organs function, how diseases progress, and how treatments might work at the cellular level.

Traditionally, scientists have relied on a technique known as cellular deconvolution, which uses bulk data to estimate what proportion of different cell types are present in a sample. Most existing deconvolution methods focus on one type of molecular data at a time. For example, RNA sequencing (RNA-seq) reveals which genes are active in cells, providing a snapshot of gene expression, while DNA methylation (DNAm) reflects chemical modifications that influence how genes are turned on or off. Each of these approaches offers valuable insights, but when used alone, they can miss important layers of biological complexity.

Overview of the proposed EMixed framework

Bulk RNA and bulk DNA methylation (DNAm) data are derived from the same biological tissue sample, consisting of mixed cell types. These two data modalities are modeled separately using latent Dirichlet allocation (LDA) to estimate cellular fractions. The multi-omics cellular fraction estimates, as illustrated on the right, represent the deconvolution results for different samples, with each color corresponding to a different cell type. EMixed integrates both RNA and DNAm data to provide a more accurate and comprehensive estimate of cellular composition.

EMixed takes a major step forward by integrating both RNA and DNAm data in what’s known as a multi-omics approach. By combining gene expression and epigenetic information, EMixed provides a more complete and accurate view of the cellular makeup of tissues. This is especially important when studying samples that contain a mix of different cell types, such as tumors, where variations in cellular composition can influence disease progression and response to therapy.

The tool works by modeling how RNA transcripts and methylation reads are distributed across cell types, then using an expectation–maximization (EM) algorithm to iteratively refine its estimates. This algorithm helps the system learn from the data and improve its accuracy with each cycle of analysis. In benchmark tests across multiple datasets, EMixed consistently outperformed existing deconvolution methods. It proved effective not only when using a single molecular modality but also when combining multiple types of data, demonstrating broad versatility and robustness.

By enabling researchers to see a more detailed picture of cellular diversity, EMixed offers a valuable resource for exploring how different molecular signals work together in health and disease. Its ability to integrate multiple layers of biological information makes it a powerful tool for advancing research in cancer biology, immunology, neuroscience, and many other fields.

Availability – R package EMixed is publicly hosted on GitHub (https://github.com/manqicai/EMixed)