From Cells to Populations: A Variational Autoencoder Approach to Pseudo-Bulk Alignment

<p xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" class="first" dir="auto" id="d2561877e98">Single-cell technologies can provide unique resolution of cellular heterogeneity. The issue is that large-scale studies currently exist mostly in bulk transcriptomics. Bulk RNA-seq captures robust population-level averages but masks cellular heterogeneity, whereas single-cell RNA-seq resolves diversity at the cost of increased noise. This discrepancy limits integrative analyses, where both fine-grained variability and population-level robustness are required. We propose a computational framework that bridges single-cell and bulk modalities using a Coupled Variational Autoencoder (Dual-VAE) designed to generate high-fidelity pseudo-bulk profiles. <p xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" dir="auto" id="d2561877e100">Using the Cancer Cell Line Encyclopedia single cell RNA sequencing dataset, we compared naïve aggregation approaches to our Dual-VAE method. While standard methods of developing pseudo-bulk structures can partially recover the data structure evinced by bulk datasets, they consistently diverged in all cases from actual bulk transcriptomes. The Dual-VAE instead aligns single-cell–derived pseudo-bulks with matched bulk RNA-seq samples by coupling their latent representations. The advantage in performing this alignment is that many of the platform-specific biases are reduced while still maintaining some biologically interpretable structure, so it facilitates cross-comparison of the different modalities in a way that reflects true similarities in the transcriptional signatures and not technical differences across platforms. <p xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" dir="auto" id="d2561877e102">Preliminary evaluations indicate that the Dual-VAE is producing pseudo-bulk profiles that are quantitatively closer to ground-truth bulk compared to those obtained by naive approaches and that support biologically meaningful clustering on a shared latent space. This framework provides harmonized inputs for gene regulatory network inference, combining single-cell granularity with the larger statistical power of bulk RNAseq datasets available for larger cohorts.