Steering cell-state and phenotype transitions by causal disentanglement learning

Chengming Zhang, Zexi Chen, Yuanxiang Miao, Yun Xue, Deyu Cai, Weifeng Guo, Hongbin Ji, Kazuyuki Aihara, Luonan ChenbioRxiv 2024 · DOI

TL;DR

This paper introduces CauFinder, a novel framework that uses causal disentanglement and network control to precisely identify and manipulate the key drivers of cell state and phenotype transitions. CauFinder can not only reveal natural biological processes like cell differentiation but also steer them, offering new avenues for therapeutic interventions.

What Problem Does This Paper Address?

Understanding and controlling how cells change their state or phenotype is fundamental to biology and medicine, from development to disease. While many methods can identify correlations between biological factors, pinpointing the causal drivers of these transitions has been a significant challenge. Existing approaches often struggle to disentangle true causal relationships from mere associations, especially in complex, high-dimensional biological data. This limitation hinders our ability to precisely intervene and steer cellular processes for therapeutic purposes.

Methods

CauFinder is built on two core components:

  1. Causal Decoupling Model: This component uses a dual variational autoencoder (DVAE) architecture. It learns to separate "causal" features from "spurious" ones in a latent space. By maximizing "causal information flow" (a measure of how much a latent variable influences the outcome), it identifies the truly influential factors. SHAP values and gradient calculations are then used to map these latent causal factors back to the original feature space, identifying key causal features and their direction of influence.
  2. Network Control Mechanism: This component integrates the identified causal features with prior gene interaction network knowledge. It employs a Minimum Feedback Vertex Set (MFVS) algorithm, weighted by the causal importance of each node, to pinpoint "master regulators" – the key genes that can be manipulated to steer transitions. A counterfactual generation algorithm is then used to simulate interventions on these master regulators to achieve desired state transitions with minimal perturbation.

Datasets

CauFinder was evaluated on a variety of datasets:

  • Simulated Datasets: Used to benchmark performance against traditional methods under controlled conditions, varying noise levels and causal strengths in linear and non-linear scenarios.
  • Real-World Datasets:
    • Embryonic Stem Cell (ESC) Data: Single-cell RNA-seq data from human and mouse ESCs to identify driver regulatory genes involved in cell fate commitment and differentiation.
  • The Cancer Genome Atlas (TCGA): Bulk RNA-seq data from eleven cancer types to identify cancer drivers.
  • Lung Adenosquamous Carcinoma (LUAS) Data: Human and mouse datasets to study transdifferentiation from lung adenocarcinoma to squamous cell carcinoma.
  • Lung Cancer Cycling Persister Cells Data: Single-cell RNA-seq data with lineage tracing to study drug resistance and response.
  • Colorectal Cancer (CRC) and Liver Metastasis (LM1) Data: Spatial transcriptomics data to analyze intercellular interactions and spatial transitions.

Key Insights and Ideas

  • Causality is Key for Control: The paper strongly emphasizes that identifying true causal drivers, not just correlations, is essential for effectively controlling biological processes.
  • Disentangling the Complex: CauFinder's ability to separate causal from spurious features in a latent space is a critical step in handling the inherent noise and interdependencies in biological data.
  • Master Regulators for Targeted Intervention: By combining causal inference with network control, CauFinder can identify specific "master regulators" that, when manipulated, can precisely steer cell state transitions.
  • Reversing Biological Transitions: A significant finding is CauFinder's capability to not only identify drivers of natural transitions but also to reverse them (e.g., drug-resistant to drug-sensitive states, or squamous cell carcinoma back to adenocarcinoma).
  • Broad Applicability: The framework demonstrates robust performance across diverse biological contexts, from stem cell differentiation and cancer progression to drug response and spatial tissue analysis.

Key Results

  • Superior Performance: CauFinder consistently outperformed traditional methods (PC algorithm, RF, T-test, VAE, MI, PCC, SCC) in identifying causal drivers on simulated datasets, achieving higher AUC and F1-scores.
  • Accurate Driver Identification: On real-world datasets, CauFinder showed superior precision in identifying driver genes for ESC differentiation and cancer progression compared to benchmark methods like CEFCON, CellOracle, and WMDS.net.
  • Lung Cancer Transdifferentiation: CauFinder successfully identified key drivers for LUAD to LUSC transdifferentiation and, importantly, demonstrated the ability to reverse this process by manipulating identified genes like TP63 and SOX2.
  • Drug Resistance Insights: In lung cancer persister cells, CauFinder identified ROS and fatty acid metabolism genes as key drivers of drug resistance and showed potential for guiding strategies to improve drug sensitivity.
  • Colorectal Cancer Metastasis: The framework identified key drivers and intercellular interactions involved in colorectal cancer progression and liver metastasis, suggesting potential therapeutic targets.

Limitations and Future Work

  • Unobserved Variables: The paper acknowledges that unobserved variables can confound causal inference. However, they argue this is less of an issue in high-dimensional spaces like scRNA-seq data where observed data often encapsulates relevant causal information.
  • Multi-Class Phenotypes: The current framework is primarily designed for binary state transitions. Extending it to handle multi-class phenotypes would broaden its applicability.
  • Continuous Variables: Integrating continuous variables directly without discretization could improve theoretical robustness.
  • Many-to-Many Relations: Developing methods to model many-to-many relationships in high-dimensional data could enhance its ability to capture hierarchical controls.
  • Advanced Network Control: Incorporating more sophisticated network control algorithms could further refine intervention strategies.

Summary updated

Abstract

Understanding and manipulating cell-state and phenotype transitions is essential for advancing biological research and therapeutic treatment. We introduce CauFinder, an advanced framework designed to accurately identify causal regulators of these transitions and further precisely steer such transitions by integrating causal disentanglement modelling with network control based solely on observed data. By leveraging do-calculus and optimizing information flow metrics, CauFinder can distinguish causal factors from spurious ones, ensuring precise control over desired state transitions. One significant advantage of CauFinder is its ability to identify those variables causally affecting the cell-state/phenotype transitions among all observed variables, both theoretically and computationally, leading to the identification of their master regulators when combined with network control. Consequently, by employing a counterfactual algorithm, CauFinder is able effectively to facilitate desirable state transitions or steer these transitional trajectories/paths by modulating these causal drivers. Beyond its theoretical advantages, CauFinder outperforms existing approaches computationally in both simulated and real-world settings. CauFinder is able to not only reveal natural biological transition processes such as (a) cell differentiation, (b) lung adenocarcinoma (LUAD) to lung squamous cell carcinoma (LUSC) transdifferentiation and (c) drug-sensitive to drug-resistant transitions but also identify the causal regulators of their reverse transition processes, such as (A) cell dedifferentiation, (B) LUSC to LUAD transdifferentiation and (C) drug-resistant to drug-sensitive transitions. These findings highlight its superior ability to causally uncover essential regulatory mechanisms and accurately steer cell-state/phenotype transitions, thus providing novel therapeutic strategies.
BibTeX
@misc{zhang2024steering,
  author = {Chengming Zhang and Zexi Chen and Yuanxiang Miao and Yun Xue and Deyu Cai and Weifeng Guo and Hongbin Ji and Kazuyuki Aihara and Luonan Chen},
  title = {{Steering cell-state and phenotype transitions by causal disentanglement learning}},
  year = {2024},
  doi = {10.1101/2024.08.16.607277},
  howpublished = {bioRxiv},
}
Figure