Abstract
Estimating a causal query from observational data is an essential task in the analysis of biomolecular networks. Estimation takes as input a network topology, a query estimation method, and observational measurements on the network variables. However, estimations involving many variables can be experimentally expensive and computationally intractable. Moreover, using the full set of variables can be detrimental, leading to bias or increasing the variance in the estimation. Therefore, designing an experiment based on a well-chosen subset of network components can increase estimation accuracy and reduce experimental and computational costs. We propose a simulation-based algorithm for selecting sub-networks that support unbiased estimators of the causal query under a constraint of cost, ranked with respect to the variance of the estimators. Three case studies demonstrated the effectiveness of well-chosen network subsets for estimating causal queries from observational data. All the case studies are reproducible and available at https://github.com/srtaheri/Simplified_LVM.
Introduction
In systems biology, researchers often aim to estimate a "causal query"—for example, predicting the downstream effect of inhibiting a specific protein. While extensive maps of biomolecular interactions are available, estimating these effects from observational data is challenging due to cost, complexity, and the counter-intuitive "more is better" fallacy. Including too many variables in a causal model can actually increase bias or variance if "bad controls" (like colliders or descendants of the effect) open spurious statistical paths. This paper introduces a simulation-based framework to identify strategic subsets of network variables that minimize estimation error while adhering to experimental budgets.
Methods: A Simulation-Based Framework
The authors propose a three-step algorithm to identify optimal sub-networks for causal estimation, specifically designed to handle latent (unobserved) variables using Acyclic Directed Mixed Graphs (ADMGs).
- Simulating Future Datasets: Synthetic datasets are generated to mimic planned experiments using historical data or known biological principles (e.g., Hill equations). This provides a "ground truth" for testing.
- Exploring Sub-ADMGs: The framework searches for subsets of nodes that form valid adjustment sets and include mediators (variables on the causal path), staying within a specified budget.
- Estimating Causal Query: Various statistical estimators (e.g., linear regression, AIPW) are applied to the synthetic data. Sub-networks are then ranked based on the empirical variance of their estimates.
Key Insights and Results
The framework was validated across three case studies: the IGF signaling pathway, a SARS-CoV-2 cytokine storm model, and real RNA-seq data from E. coli.
A major finding is the power of mediators. Including variables that sit between the cause and effect can significantly reduce estimator variance. In the IGF signaling case, including a single mediator reduced variance by over 50% compared to standard adjustment sets.
Furthermore, the results consistently demonstrate that "less is more." Well-chosen subsets frequently yield lower variance and less bias than the full network. As shown in the figure above, different sub-networks (1-4) result in varying distributions of the estimated Average Treatment Effect (ATE). Notably, the full network (Sub-ADMG 4) can lead to biased results compared to more focused sub-networks.
Conclusion
This research provides a diagnostic tool for biologists to "test drive" experimental designs before data collection. By identifying sub-networks that provide the most precise and unbiased causal estimates, researchers can optimize their resources. While the method currently requires a known network topology, future work aims to integrate falsification modules and expand the types of causal queries supported.
Summary updated