Abstract
This paper frames causal structure estimation as a machine learning task. The idea is to treat indicators of causal relationships between variables as 'labels' and to exploit available data on the variables of interest to provide features for the labelling task. Background scientific knowledge or any available interventional data provide labels on some causal relationships and the remainder are treated as unlabelled. To illustrate the key ideas, we develop a distance-based approach (based on bivariate histograms) within a manifold regularization framework. We present empirical results on three different biological data sets (including examples where causal effects can be verified by experimental intervention), that together demonstrate the efficacy and general nature of the approach as well as its simplicity from a user's point of view.
Introduction
Determining causal relationships—identifying "who causes what"—is a fundamental challenge in scientific research, particularly in complex systems like gene regulatory networks. Traditionally, causal discovery has been approached through two primary lenses: constraint-based methods, which rely on statistical tests for conditional independence, and score-based methods, which search for a generative model that best explains the observed data.
However, these traditional approaches often struggle with the complexities of real-world biological data, such as latent variables (unobserved causes), feedback loops (cycles), and the computational explosion inherent in searching through massive graph spaces. Furthermore, researchers often possess "prior knowledge" from previous experiments or literature that is difficult to integrate into standard algorithms. This paper proposes a paradigm shift: framing causal discovery not as a search for a global generative model, but as a semi-supervised classification task. By asking whether we can learn to recognize the "signature" of a causal relationship in data, the authors introduce a method that is both robust and flexible.
The MRCL Framework
The authors introduce Manifold Regularized Causal Learning (MRCL), which treats causality as a labeling problem. The framework consists of four key components:
1. Featurization
For every pair of variables $(i, j)$, the algorithm generates a "feature vector" derived from their joint distribution. In this implementation, the authors utilize simple bivariate histograms of the variables' values. This approach assumes that the geometric properties of the joint distribution contain sufficient information to distinguish causal from non-causal interactions.
2. Semi-Supervised Learning
MRCL assumes that the causal status (Label: +1 for "causal," -1 for "not causal") is known for a small subset of variable pairs. These labels might come from gold-standard interventional data, such as gene knockouts. The vast majority of other pairs remain unlabeled, and the goal is to predict their status.
3. Manifold Regularization
Using a Laplacian-regularized least squares framework, the algorithm operates on the "manifold assumption": if two pairs of variables possess very similar bivariate histograms, they are likely to share the same causal label. The algorithm effectively propagates known labels across a manifold constructed from the distances between these histograms, allowing the model to generalize from a few known examples to the entire network.
4. Structural Agnosticism
Unlike Directed Acyclic Graph (DAG) methods, MRCL does not impose rigid structural constraints. It is indifferent to whether a relationship is direct or ancestral, or whether the underlying system contains cycles. It focuses purely on identifying the statistical patterns associated with causal influence.
Experimental Validation
The efficacy of MRCL was tested across three distinct biological contexts:
- Yeast Knockout Data (D1): A large-scale dataset of gene expression measurements following specific deletions. This provided a definitive "gold standard" for interventional effects to validate the model's predictions.
- Kinase Intervention Data (D2): Measurements of signaling proteins in human cancer cell lines. This dataset included time-course data and chemical inhibitors, providing a complex testbed for the algorithm's ability to handle proteomic dynamics.
- Human Cancer Patient Samples (D3): Observational protein data from 820 breast cancer biopsies (TCGA). In the absence of direct interventions, results were validated against established biochemical literature.
Key Insights and Results
Causality as Pattern Recognition
The success of MRCL suggests that the signature of causality is often embedded in the geometry of the joint distribution. Rather than requiring a full generative model, the method demonstrates that we can detect edges by identifying if a specific pair "looks like" other known causal pairs. This shifts the focus from global model-fitting to local pattern recognition.
Performance and Robustness
On the Yeast dataset, MRCL consistently achieved higher Area Under the Curve (AUC) scores than established causal discovery tools such as the PC algorithm, IDA, and GIES. Notably, the method proved robust to sample size, maintaining high performance even with limited training data ($n=200$). As the proportion of labeled examples ($ ho$) increased, the accuracy scaled predictably, highlighting the benefit of integrating prior knowledge.
Competitive Advantage
In the human cell line experiments, MRCL was competitive with—and in some cases outperformed—specialized Bayesian models that were custom-built for proteomic data. This suggests that the manifold regularization approach is a powerful, general-purpose alternative to domain-specific modeling.
Limitations and Future Directions
While MRCL offers a powerful new tool for causal discovery, it has specific limitations. As a discriminative rather than generative model, it does not provide a full probability distribution of the system and cannot predict the specific numerical outcome of a new intervention. Additionally, the current iteration treats pairs independently, without enforcing global logical constraints like transitivity (e.g., if $A \to B$ and $B \to C$, then $A \to C$ should be recognized as an ancestral effect).
Future work may involve coupling MRCL with global graph constraints to ensure structural consistency. However, even in its current form, MRCL represents a significant step forward in making causal discovery more accessible and effective by leveraging the strengths of semi-supervised machine learning.
Summary updated