Introduction
In the ideal practice of causal inference, researchers perform precise interventions—such as knocking out a specific gene—and observe the resulting downstream effects. However, real-world interventions are often "messy." For example, a CRISPR-Cas9 experiment might successfully hit its intended target but also cause unintended "off-target" mutations. These effects can vary from cell to cell, leaving researchers with a mixture distribution of different, unknown interventions rather than a single, clean data point.
This paper, originally presented at UAI 2021, provides a theoretical and algorithmic framework to solve this problem: identifying individual intervention targets and their mixing proportions from a combined mixture of data. The authors establish the conditions under which these hidden causal actions can be uniquely identified, a property known as identifiability.
Methodology
The researchers model the system using Causal Bayesian Networks (CBNs) and the $do()$ operator. Their approach to disentangling these mixtures rests on two primary pillars:
- The Exclusion Assumption: This is the key theoretical insight. It assumes that for every variable in the network, there is at least one possible value that is never a target of any intervention in the mixture. In a biological context, this is a reasonable assumption; most genes have "normal" or wild-type states that are not the result of a targeted knockout.
- The DISENTANGLE Algorithm: The authors propose a recursive inductive algorithm. For a single node, the problem reduces to a linear system. For a network of $N+1$ nodes, the algorithm marginalizes out a leaf node to solve for $N$ nodes, then "lifts" these solutions back to the $N+1$ case by solving a local system of linear equations to find the specific intervention values and mixing weights for the added node.
Experimental Setup
The paper utilizes synthetic datasets to validate the theory, as standard benchmarks for "mixtures of unknown interventions" do not yet exist. The researchers generated directed acyclic graphs (DAGs) using Scale-Free (SF) and Erdös-Rényi (ER) models, with networks ranging from $N=4$ to $N=32$ nodes. They tested the algorithm's performance with sample sizes ranging from $2^4$ to $2^{20}$ to analyze the data requirements for accurate recovery.
Results and Discussion
The authors prove that under the Exclusion and Positivity assumptions, the set of intervention targets and weights is unique. The DISENTANGLE algorithm runs in polynomial time relative to the number of nodes and mixture components.
As illustrated in the experimental results, the algorithm's performance improves significantly with data volume. The Recall (the percentage of true intervention targets found) approaches 1.0 as the sample size increases. Simultaneously, the Root-Mean-Squared Error (RMSE) of the predicted mixing weights and the false-positive/negative rates decrease sharply. However, the results also highlight a high sample complexity: larger networks (e.g., $N=12$) require significantly more samples to achieve the same level of accuracy as smaller networks ($N=4$).
Conclusion
This work provides a pioneering framework for identifying hidden causal interventions from mixture data. While the algorithm is efficient, it currently assumes the underlying causal structure is known and requires substantial data for larger networks. Future work will focus on extending this to handle unobserved confounders and learning the graph structure directly from the mixture distributions.
Summary updated