Cyclic Causal Discovery from Continuous Equilibrium Data

Joris M. Mooij, Tom Heskes

TL;DR

This paper introduces a novel method for learning causal relationships in biological signaling networks, even when feedback loops are present and data is continuous. By modeling interventions as changes in compound activity rather than abundance and using linearized causal mechanisms, they achieve a more accurate reconstruction of a complex cellular signaling network than previous methods.

What Problem Does This Paper Address?

Understanding cause-and-effect relationships in complex systems, like biological signaling pathways, is crucial for predicting how interventions will affect the system. Traditional causal discovery methods often struggle with:

  • Feedback Loops: Many biological systems have feedback loops, meaning variables can influence each other cyclically. Standard methods like Bayesian networks assume acyclicity, making them unable to detect these crucial relationships.
  • Continuous Data: Biological measurements are often continuous, but many causal discovery methods require data to be discretized (e.g., categorized as "low," "medium," or "high"). This discretization can lead to a loss of valuable information.
  • Intervention Types: Interventions in biological experiments often don't directly change the abundance of a compound but rather its activity – how it influences other compounds. Existing causal discovery formalisms for interventions aren't well-suited for this.

The authors aim to develop a method that addresses these limitations, specifically for analyzing equilibrium data from systems with feedback loops.

Methods

The paper proposes a method based on Structural Causal Models (SCMs), which are a standard framework for causal inference. Their key innovations include:

  • Continuous Data Handling: The method works directly with continuous-valued measurements, avoiding information loss from discretization.
  • Activity Interventions: Instead of modeling interventions as forcing a variable to a specific value (like abundance), they model them as changes to the causal mechanism of a compound. This means the way a compound influences its "children" in the causal graph is altered.
  • Local Linearization: To handle potentially nonlinear causal mechanisms computationally, they approximate them with coupled local linearizations. This means for each experimental condition, they assume a linear relationship between variables, but these linear relationships can differ across conditions.
  • Cyclic Graph Structure: The method explicitly allows for directed cycles in the causal graph, enabling the discovery of feedback loops.
  • Bayesian Approach with Priors: They use a Bayesian framework to learn the causal structure and parameters. They introduce different prior choices for the linearized causal mechanisms, distinguishing between "linear mechanisms prior" (assuming minimal relinearization) and "nonlinear mechanisms prior" (allowing for more significant changes in linearized mechanisms due to interventions).

Datasets

The primary dataset used for evaluation is the multivariate proteomics data set measured and analyzed by Sachs et al. (2005). This dataset consists of flow cytometry measurements of 11 biochemical compounds in human immune cells under various experimental perturbations. The authors use a subset of 8 out of the 14 experimental conditions described in the original paper.

Key Insights and Ideas

  • Feedback is Key: The analysis strongly suggests that feedback loops are prevalent in the cellular signaling network studied, and their method is capable of detecting them, unlike acyclic models.
  • Activity Interventions Matter: Modeling interventions as changes in compound activity rather than abundance is crucial for accurately representing the experimental setup and achieving better model fits.
  • Continuous Data is Informative: Working with continuous data preserves information that would be lost through discretization, leading to more nuanced and accurate causal models.
  • Parameter Identifiability Challenges: In cyclic cases, while the causal structure can be robustly identified, the specific parameters of the model are often not uniquely identifiable from the data. This has implications for the precision of quantitative predictions.

Key Results

  • The proposed method successfully reconstructs a cellular signaling network, providing evidence for feedback loops that were missed by previous acyclic methods.
  • Their reconstructed model offers a more accurate quantitative description of the data compared to existing methods at comparable model complexity.
  • When allowing for cycles, their method achieves significantly better (lower negative log-evidence) scores than the consensus network and the reconstruction by Sachs et al. (2005), indicating a better fit to the data. For example, with a cyclic, nonlinear, non-Gaussian prior and 17 edges, their model achieved a negative log-evidence of 72.9 (x 10^3), compared to 79.2 (x 10^3) for Sachs et al.'s reconstruction.
  • They found evidence for specific two-cycles, including Mek⇄PKC, Akt⇄Erk, and Mek⇄PKA.

Limitations and Future Work

  • Equilibrium Assumption: The method assumes that the cells have reached equilibrium when measurements are taken, which is an approximation necessary due to the absence of time-series data.
  • Parameter Identifiability: As mentioned, in cyclic cases, parameters can be difficult to estimate reliably, impacting the precision of quantitative predictions. The authors suggest that more experimental data might be needed to overcome this.
  • Future Directions:
    • Compare their local linearization approach with other approximations (e.g., FITC).
  • Incorporate information about the sign of activity interventions (inhibitor vs. activator) to potentially further improve results.
  • Find collaborators for experimental validation of their findings.

Summary updated

Abstract

We propose a method for learning cyclic causal models from a combination of observational and interventional equilibrium data. Novel aspects of the proposed method are its ability to work with continuous data (without assuming linearity) and to deal with feedback loops. Within the context of biochemical reactions, we also propose a novel way of modeling interventions that modify the activity of compounds instead of their abundance. For computational reasons, we approximate the nonlinear causal mechanisms by (coupled) local linearizations, one for each experimental condition. We apply the method to reconstruct a cellular signaling network from the flow cytometry data measured by Sachs et al. (2005). We show that our method finds evidence in the data for feedback loops and that it gives a more accurate quantitative description of the data at comparable model complexity.
BibTeX
@inproceedings{mooij2013cyclic,
  author = {Joris M. Mooij and Tom Heskes},
  title = {{Cyclic Causal Discovery from Continuous Equilibrium Data}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2013},
  url = {https://openreview.net/forum?id=G0nhb1JbOp},
}
Figure