Author Correction: Causal disentanglement for single-cell representations and controllable counterfactual generation

Yicheng Gao, Kejing Dong, Caihua Shan, Dongsheng Li, Qi LiuNature Communications 2025 · DOI · OpenAlex

Introduction

This article presents a correction to a research paper originally published in Nature Communications regarding CausCell, a novel framework for causal disentanglement in single-cell data. The correction primarily addresses a labeling error in the evaluation figures, but the underlying research provides significant advancements in separating biological signals from technical noise.

The Challenge of Single-Cell Data

Single-cell RNA sequencing (scRNA-seq) offers high-resolution biological insights but is often plagued by technical noise and overlapping biological variables. Standard machine learning models frequently fail to generalize to Out-of-Distribution (OOD) data—scenarios where the testing data distribution differs from the training data. Furthermore, researchers seek to perform counterfactual generation, predicting cell behavior under hypothetical treatments. This requires a deep understanding of causal relationships rather than mere correlations.

The CausCell Framework

CausCell addresses these challenges through causal disentanglement. The methodology relies on three core pillars:

  1. Causal Graph Modeling: Using Directed Acyclic Graphs (DAGs) to map relationships between variables like Treatment, Cell Type, and Batch effects. This prevents "shortcut learning," where models might mistakenly attribute drug effects to technical confounders.
  2. Disentangled Representation Learning: Mapping high-dimensional data into a latent space where different biological and technical concepts are stored in independent dimensions.
  3. Controllable Generation: Because representations are disentangled, researchers can theoretically "dial" the treatment variable in the latent space to simulate counterfactual biological responses.
Causal graph and experimental setup for in-distribution splitting.

Experimental Evaluation and Results

The authors evaluated CausCell against scDisInFact across multiple metrics, including Accuracy (ACC), Precision, Recall, and F1-score. The evaluation spanned two critical settings:

  • In-distribution (ID): Where training and testing sets share similar concept combinations.
  • Out-of-distribution (OOD): Where the model is tested on novel combinations of concepts (e.g., a specific cell type under a treatment it never saw during training) to test its reasoning capabilities.
Comprehensive performance comparison between CausCell and scDisInFact across ID and OOD settings.

Key Findings

  • Superior Classification: CausCell (represented by red bars in the figures) consistently outperforms the baseline across almost all metrics.
  • Robustness to Distribution Shifts: The performance gap is most pronounced in OOD settings. While other models degrade when faced with unseen data combinations, CausCell maintains high scores, particularly in identifying Cell Type and Treatment effects.
  • Clustering Quality: UMAP projections and metrics like the Adjusted Rand Index (ARI) show that CausCell produces much cleaner clusters, indicating successful separation of biological identities from confounding factors.

Correction Details

The correction clarifies a vital labeling error in the original publication. In Figure 2b, "Experimental setting 2" was incorrectly marked as "In-distribution (ID)" when it should have been labeled "Out-of-distribution (OOD)." This distinction is critical as it underscores the model's primary claim: its superior ability to generalize to unseen biological contexts through causal reasoning.

Summary updated

BibTeX
@inproceedings{gao2025author,
  author = {Yicheng Gao and Kejing Dong and Caihua Shan and Dongsheng Li and Qi Liu},
  title = {{Author Correction: Causal disentanglement for single-cell representations and controllable counterfactual generation}},
  booktitle = {Nature Communications},
  year = {2025},
  doi = {10.1038/s41467-025-63728-0},
}
Figure