Understanding complex tissues and tumors is a major challenge in biology, especially when different types of cells are mixed together in bulk samples. Traditional methods often rely only on RNA sequencing data to separate these cell types, which can limit accuracy when looking at other types of biological information, such as proteins or DNA modifications.

To address this, researchers from St. Jude Children’s Research Hospital developed MOADE, a computational tool that uses a multimodal autoencoder to better separate bulk multi-omics data. Unlike older methods, MOADE can combine information from multiple biological sources and predict both cell composition and personalized molecular profiles.

Overview of MOADE framework

a MOADE resolves the developmental origins or tumor microenvironment (TME) of target bulk tissues using scRNA-seq reference from fetal samples or external tumors. The dissociated multi-omic profiles may improve disease subtyping and prognosis prediction. b The input of MOADE consists of RNA and non-RNA bulk omics data paired by tissues. MOADE initializes a personalized and internal signature matrix via Joint Non-negative Matrix Factorization and shared cell counts fractions. c Generate large-scale training data based on an external scRNA-seq reference and the internal signature matrices from Step 1. A proportional shift strategy coupled with modality-specific sparsity is used to simulate ground truth cellular compositions in RNA and non-RNA molecules. d A multimodal autoencoder with one encoder for the concatenated input data and two decoders for reconstructed bulk multiome. An updated personalized multi-omics signature matrix is extracted from two decoders adapted to each target tissue. e MOADE outputs the predicted cell proportions and high-resolution purified multi-omic profiles, enabling downstream cts-differential analysis

The team tested MOADE with simulated data and real samples from a variety of tissues. Compared to nine other pipelines, MOADE showed stronger accuracy and better adaptability across different conditions. This means scientists can now gain a clearer picture of how cell types contribute to health and disease, even when analyzing complex tissue mixtures.

MOADE offers a promising step forward for researchers working on cancer biology, developmental biology, and personalized medicine by helping them make sense of complicated multi-omic datasets.

Availability – MOADE is available on GitHub at https://github.com/jsuncompubio/MOADE