bulk-to-single-deconvolution_skill

Overview

This skill turns bulk RNA-seq cohorts into synthetic single-cell datasets using omicverse's Bulk2Single workflow. It combines cell-fraction estimation, beta-VAE generation, and quality-control comparisons against a matched reference scRNA-seq atlas. The output is an AnnData object of generated cells plus plots and tables for benchmarking and downstream analysis.

How this skill works

The workflow first harmonises gene identifiers and estimates sample-wise cell fractions from bulk counts using an integrated deconvolution estimator. It then aligns features between bulk and reference single-cell data, trains a beta-variational autoencoder to generate per-cell expression profiles, and filters low-quality synthetic cells. Final steps produce composition plots, correlation heatmaps, and embeddings to compare generated profiles to the reference atlas.

When to use it

• You have bulk RNA-seq from heterogeneous tissue and a matched scRNA-seq reference atlas.
• You want per-sample synthetic single-cell profiles for downstream single-cell pipelines.
• You need cell-type proportion estimates for cohort-level reporting.
• You plan to benchmark synthetic data quality against a known single-cell reference.
• You want to augment limited single-cell data with generated cells for method testing.

Best practices

• Provide raw counts for both bulk and reference single-cell data to allow the lazy preprocessing steps to run correctly.
• Confirm bulk_group names match the column IDs in your bulk matrix before initialising the model.
• Start training on GPU (gpu set to a CUDA device) to reduce run time; use gpu=-1 only for CPU runs.
• Save intermediate outputs (fraction table, trained VAE weights, generated AnnData) to enable resuming and reproducibility.
• If marker selection fails, increase top_marker_num or supply a curated marker list to improve alignment.

Example use cases

• Estimate cell fractions for PDAC bulk replicates and generate synthetic single cells for cell-type-specific differential expression.
• Load pre-trained Bulk2Single weights to quickly regenerate and benchmark cells against the dentate gyrus atlas.
• Create composition plots and correlation heatmaps to validate how closely generated cells match reference clusters.
• Filter noisy generated cells and export a filtered AnnData for downstream trajectory or RNA-velocity analyses.