Scholarlink

<p>Conducting disentanglement learning on single-cell omics data offers a promising alternative to traditional black-box representation learning by separating the semantic concepts embedded in a biological process. We present CausCell, which incorporates the factual information about causal relationships among disentangled concepts within a diffusion model to generate more reliable disentangled cellular representations, with the aim of increasing the explainability, generalizability and controllability of single-cell data, including spatial-temporal omics data, relative to those of the existing black-box representation learning models. Two quantitative evaluation scenarios, i.e., disentanglement and reconstruction, are presented to conduct the first comprehensive single-cell disentanglement learning benchmark, which demonstrates that CausCell outperforms the state-of-the-art methods in both scenarios. Additionally, CausCell can implement controllable generation by intervening with the concepts of single-cell data when given a causal structure. It also has the potential to uncover biological insights by generating counterfactuals from small and noisy single-cell datasets.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Causal disentanglement for single-cell representations and controllable counterfactual generation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月23日

查看全文
  • Yicheng Gao,
  • Kejing Dong,
  • Caihua Shan,
  • Dongsheng Li,
  • Qi Liu

摘要

Conducting disentanglement learning on single-cell omics data offers a promising alternative to traditional black-box representation learning by separating the semantic concepts embedded in a biological process. We present CausCell, which incorporates the factual information about causal relationships among disentangled concepts within a diffusion model to generate more reliable disentangled cellular representations, with the aim of increasing the explainability, generalizability and controllability of single-cell data, including spatial-temporal omics data, relative to those of the existing black-box representation learning models. Two quantitative evaluation scenarios, i.e., disentanglement and reconstruction, are presented to conduct the first comprehensive single-cell disentanglement learning benchmark, which demonstrates that CausCell outperforms the state-of-the-art methods in both scenarios. Additionally, CausCell can implement controllable generation by intervening with the concepts of single-cell data when given a causal structure. It also has the potential to uncover biological insights by generating counterfactuals from small and noisy single-cell datasets.