Sample-Specific Molecular Heterogeneity in Atrial Fibrillation
DocDay Poster
The full poster is available below. This poster summarizes an exploratory single-nucleus transcriptomic meta-analysis of human atrial fibrillation, focusing on sample-specific molecular heterogeneity rather than only average AF-versus-SR differences.
Extended Sample Distance Heatmap
This heatmap provides a resolution view of the sample-level structure shown in the poster. This analysis highlights how individual samples group according to their transcriptomic similarity in the cardiomyocyte latent space.
The annotations help evaluate whether sample proximity is associated with disease status, sex, atrial region, study origin, or the inferred cluster group. This is important because atrial fibrillation is highly heterogeneous, and patient-specific molecular patterns may be partly masked in conventional group-level comparisons.
The highlighted samples occupy nearby regions of the embedding across multiple cell-type neighborhoods, suggesting shared sample-level transcriptional structure rather than a signal restricted to a single cell population. The presence of an SR sample close to these AF samples also illustrates why sample-level analysis is useful: molecular similarity may not perfectly follow the clinical AF/SR label. This supports the use of latent-space distance analysis to capture patient-level heterogeneity beyond average case-control comparisons.
Structure of the VAE Model
The input to the model is the integrated single-nucleus RNA-seq dataset, including cells from multiple studies, atrial regions, sexes, and disease conditions.
The model learns a compressed latent representation, referred to here as latent z, that captures major transcriptional patterns while reducing noise from the high-dimensional gene-expression space.
This VAE model generat with MrVI from scvi-tools. More information about scvi-tools is available at the official GitHub repository.