bulk2scDiff: Generating Single-Cell Expression Profiles from Pseudobulk Data with Conditional Diffusion Models

In plain English

The method starts with an average gene-expression profile from many cells and generates a collection of individual-cell profiles that could have produced it. The goal is to recover cellular diversity that bulk measurements hide.

What they found

• The authors report coherent held-out cell populations and recurring immune features.
• The swap control tests whether generated profiles respond to the conditioning pseudobulk rather than reproducing a fixed template.

The catch

• The work is a preprint and has not been peer reviewed.
• Generated cells are hypotheses, not substitutes for measuring real single cells.

Methods at a glance

Study type

Generative-model preprint

Biological system

Breast cancer and acute myeloid leukemia

Organism

Human

Dataset

Breast-cancer and acute-myeloid-leukemia single-cell RNA-sequencing datasets

Sample size

Not reported

Model architecture

Conditional diffusion model

Training objective

Generate single-cell expression populations conditioned on pseudobulk profiles

Code

No public code repository is identified in the abstract.

Data

The abstract identifies breast-cancer and acute-myeloid-leukemia single-cell datasets.

Tasks and metrics

Tasks: Population reconstruction; Conditioning sensitivity

Metrics: Population coherence and recurrent feature recovery as reported by the authors

Baselines: Held-out-population comparisons; Pseudobulk-swap control

Principal findings

• The authors report coherent held-out populations and recurrent immune features.

What remains unresolved

External validation: Not reported

• Preprint evidence only.
• Synthetic single-cell profiles require validation against measured cells before biological interpretation.