A major challenge in current sequence-to-activity modelling is the reliable interpretation of non-coding variants for clinical prioritisation. As a result, generating system-specific, high-throughput datasets—such as those derived from in vitro differentiation—has become a common strategy to gain clinically relevant insights. However, even with tailored datasets, deep learning models in regulatory genomics face interpretability challenges: it is not yet fully understood how architectural and methodological choices influence learned representations, potentially limiting reliable variant effect predictions.

We addressed these challenges in the context of congenital myopathies, elucidating the early regulatory events of skeletal muscle development and identifying key elements driving progenitor differentiation. We collected single-cell Multiome and bulk histone modification data across four critical time points in an in vitro human skeletal muscle differentiation system. We used these data to train a deep convolutional sequence-to-activity model that predicts cell-type- and timepoint-resolved DNA accessibility profiles at base-pair resolution. We used model explainability techniques to uncover sequence patterns that drive the model's predictions, as well as how multitask learning aids the model's explainability. We leveraged in silico mutagenesis to prioritise non-coding variants that disrupt transcription factor binding sites and utilised complementary 3D-interaction data to further link identified enhancers to target genes. This study provides the first comprehensive map of cis-regulatory drivers in human skeletal myogenesis by combining multi-omic data and explainable deep learning to unravel how time-dependent DNA regulatory activity drives cell fate decisions.

Co-authors: Birthe Lange, Markus Schülke, Uwe Ohler