Bulk RNA-seq deconvolution with Bulk2Single

Overview

Use this skill when a user wants to reconstruct single-cell profiles from bulk RNA-seq together with a matched reference scRNA-seq atlas. It follows t_bulk2single.ipynb, which demonstrates how to harmonise PDAC bulk replicates, train the beta-VAE generator, and benchmark the output cells against dentate gyrus scRNA-seq.

Instructions

1. Load libraries and data
   • Import omicverse as ov, scanpy as sc, scvelo as scv, anndata, and matplotlib.pyplot as plt, then call ov.plot_set() to match omicverse styling.
   • Read the bulk counts table with ov.read(...)/ov.utils.read(...) and harmonise gene identifiers via ov.bulk.Matrix_ID_mapping(<df>, 'genesets/pair_GRCm39.tsv').
   • Load the reference scRNA-seq AnnData (e.g., scv.datasets.dentategyrus()) and confirm the cluster labels (stored in adata.obs['clusters']).
2. Initialise the Bulk2Single model
   • Instantiate ov.bulk2single.Bulk2Single(bulk_data=bulk_df, single_data=adata, celltype_key='clusters', bulk_group=['dg_d_1', 'dg_d_2', 'dg_d_3'], top_marker_num=200, ratio_num=1, gpu=0).
   • Explain GPU selection (gpu=-1 forces CPU) and how bulk_group names align with column IDs in the bulk matrix.
3. Estimate cell fractions
   • Call model.predicted_fraction() to run the integrated TAPE estimator, then plot stacked bar charts per sample to validate proportions.
   • Encourage saving the fraction table for downstream reporting (df.to_csv(...)).
4. Preprocess for beta-VAE
   • Execute model.bulk_preprocess_lazy(), model.single_preprocess_lazy(), and model.prepare_input() to produce matched feature spaces.
   • Clarify that the lazy preprocessing expects raw counts; skip if the user has already log-normalised data and instead provide aligned matrices manually.
5. Train or load the beta-VAE
   • Train with model.train(batch_size=512, learning_rate=1e-4, hidden_size=256, epoch_num=3500, vae_save_dir='...', vae_save_name='dg_vae', generate_save_dir='...', generate_save_name='dg').
   • Mention early stopping via patience and how to resume by reloading weights with model.load('.../dg_vae.pth').
   • Use model.plot_loss() to monitor convergence.
6. Generate and filter synthetic cells
   • Produce an AnnData using model.generate() and reduce noise through model.filtered(generate_adata, leiden_size=25).
   • Store the filtered AnnData (.write_h5ad) for reuse, noting it contains PCA embeddings in obsm['X_pca'].
7. Benchmark against the reference atlas
   • Plot cell-type compositions with ov.bulk2single.bulk2single_plot_cellprop(...) for both generated and reference data.
   • Assess correlation using ov.bulk2single.bulk2single_plot_correlation(single_data, generate_adata, celltype_key='clusters').
   • Embed with generate_adata.obsm['X_mde'] = ov.utils.mde(generate_adata.obsm['X_pca']) and visualise via ov.utils.embedding(..., color=['clusters'], palette=ov.utils.pyomic_palette()).
8. Troubleshooting tips
   • If marker selection fails, increase top_marker_num or provide a curated marker list.
   • Alignment errors typically stem from mismatched bulk_group names—double-check column IDs in the bulk matrix.
   • Training on CPU can take several hours; advise switching gpu to an available CUDA device for speed.

Examples

• "Estimate cell fractions for PDAC bulk replicates and generate synthetic scRNA-seq using Bulk2Single."
• "Load a pre-trained Bulk2Single model, regenerate cells, and compare cluster proportions to the dentate gyrus atlas."
• "Plot correlation heatmaps between generated cells and reference clusters after filtering noisy synthetic cells."

References

• Tutorial notebook: t_bulk2single.ipynb
• Example data and weights: omicverse_guide/docs/Tutorials-bulk2single/data/
• Quick copy/paste commands: reference.md