跳转至

文章背景与核心概要

虚拟细胞利用机器学习技术来模拟和预测细胞行为，从而助力健康与疾病的研究。虽然注入因果图可以增强模型的可解释性，但现实世界的应用中往往缺乏现成的因果图。尽管自动化的无监督方法可以从数据中构建因果图（将基因分组为概念以提取因果关系），但这些过程常常会引入错误。

为了克服这一挑战，由王鹏程（Pengcheng Wang）等人提出的“虚拟细胞的人机协同因果知识注入”（Human-Guided Causal Knowledge Injection for Virtual Cells）引入了一个人机协同框架。该框架具备以下核心特点：1. 基于基因相似性感知的因果图可视化，并辅以混合优化算法来探索概念关系与基因相似性；2. 依托专属反事实和因果路径可视化的反事实分析策略，以有效验证和精炼因果图。该方法的有效性已通过真实世界的案例研究得到验证，不仅生成了具有科学意义的因果见解，还获得了领域专家的积极反馈。

Human-Guided Causal Knowledge Injection for Virtual Cells

📌 Summary

Virtual cells utilize machine learning to simulate and predict cellular behaviors, aiding in the investigation of health and disease. While injecting causal graphs enhances model interpretability, real-world applications often lack pre-existing graphs. Although automated unsupervised methods can construct causal graphs from data (grouping genes into concepts to extract causal relationships), these processes frequently introduce errors.

To overcome this, Human-Guided Causal Knowledge Injection for Virtual Cells introduces a human-in-the-loop framework designed by Pengcheng Wang et al. It features: * Gene-similarity-aware causal graph visualization backed by a hybrid optimization algorithm to explore concept relationships and gene similarities. * Counterfactual analysis strategies supported by dedicated counterfactual and causal path visualizations to effectively validate and refine causal graphs.

The method's efficacy is validated through real-world case studies, generating scientifically meaningful causal insights and receiving positive feedback from domain experts.

📋 Metadata

  • arXiv ID: arXiv:2608.08430 [cs.HC]
  • Primary Subject: Human-Computer Interaction (cs.HC)
  • Secondary Subjects: Artificial Intelligence (cs.AI)
  • Submitted On: August 9, 2026
  • Authors:
  • Pengcheng Wang
  • Changjian Chen
  • Zhuo Tang
  • You Wu
  • Long Wang
  • Feng Yu
  • Kenli Li

External References