Deep learning for spatial multi-omics: predicting cardiomyocyte differentiation efficiency at single-cell resolution

Abstract

Cardiovascular diseases remain the leading cause of global mortality, with limited regenerative capacity of adult cardiac tissue presenting significant therapeutic challenges. The primary cause of death worldwide is still cardiovascular diseases, and treating these conditions is extremely difficult due to the adult heart tissue's limited capacity for regeneration. Cardiomyocytes derived from human induced pluripotent stem cells (hiPSC CMs) present promising potential for cardiac regenerative medicine; however, existing differentiation protocols are highly inconsistent and do not have accurate predictive evaluation techniques. By integrating the analysis of temporal gene expression data and spatial transcriptomics, this study developed a novel hybrid deep learning architecture that combines Graph Neural Networks (GNNs) and Recurrent Neural Networks (RNNs) to predict the outcomes of cardiomyocyte differentiation. RNN components analysed temporal gene expression trajectories across 800 samples and 10 time points, while GNN components processed spatial transcriptomics data from 752 tissue spots to capture spatial relationships. Three fusion strategies - concatenation, attention-based, and ensemble approaches - were meticulously evaluated. With an accuracy of 96.67%, the ensemble fusion approach outperformed the state-of-the-art computational approaches by a significant margin (+13.47% compared to the top GNN approaches and +6.97% compared to specialised biological models). Keywords: Cardiomyocyte differentiation; Spatial transcriptomics, Spatial multi-omics; Single-cell biology; Deep learning; Graph Neural Networks; Recurrent Neural Networks; Stem cells; Artificial Intelligence; Cardiac biology