DECONVOLVING THE TRANSCRIPTOMIC SIGNATURES OF SOMATIC EXPANSION IN HUNTINGTON’S DISEASE

Huntington’s disease (HD) is a monogenic neurodegenerative disorder characterized by progressive damage to striatal neurons, primarily driven by somatic CAG repeat expansions – a dynamic, cell type-dependent process that leads to their degeneration. However, generating paired single-cell data including both expansion length alongside gene expression is experimentally challenging and not commonly done. Nevertheless, a clear understanding of somatic expansion processes is vital to understand experimental models and accurately evaluate new therapeutic approaches. Here, we present a novel methodology, DEconvolution of CAG Signatures (DECAG), based on a deep generative model that predicts somatic expansion stage from single-cell RNA-seq data using a recently published paired dataset as a reference panel. In addition, DECAG explicitly disentangles expansion-associated transcriptional variability from confounding factors such as cell-type and sequencing technology, and will continue to improve as more paired datasets become available. Using paired data with ground-truth, we show that DECAG outperforms the baseline models, such as scANVI, by a margin of 10% reaching a test balanced accuracy of72%, while producing interpretable latent representations. More importantly, DECAG also shows competitive performance in the semi-supervised setting, where we train our model using scRNA-seq data with partial CAG information. In this case, the model enables propagation of CAG expansion labels to standard scRNA-seq datasets, facilitating mutation-length-aware analyses of existing and future experiments. In conclusion, DECAG provides an evolving framework for accurately inferring somatic CAG expansion stages from single-cell transcriptomes, enabling mutation-length-aware analyses that can accelerate research and therapeutic development in HD and other polyglutamine diseases.