TL;DR
This paper introduces a novel two-step framework that combines empirical Bayes matrix factorization (EBMF) with causal discovery algorithms to tackle the challenges of high dimensionality and multicollinearity in biological datasets. This approach successfully identifies latent factors that represent underlying biological mechanisms and enables more accurate causal inference, leading to improved predictions and biological insights in real-world datasets.
What Problem Does This Paper Address?
Modern biological research, particularly genomics, generates massive datasets with tens or hundreds of thousands of features. Analyzing these datasets to understand disease mechanisms and predict outcomes is crucial. However, two major hurdles exist:
- Curse of Dimensionality: The sheer number of features makes it computationally challenging and statistically difficult to find meaningful patterns.
- Multicollinearity: Biological features (like gene expression) are often highly correlated due to underlying regulatory networks (e.g., genes regulated by the same transcription factors). This makes it hard to disentangle the independent effects of individual features and can lead to unstable causal inference.
Existing causal discovery algorithms struggle with these issues, limiting their application to complex, high-dimensional biological data. This paper aims to bridge this gap by developing a method that can effectively reduce dimensionality and handle multicollinearity while performing causal discovery.
Methods
The proposed framework consists of two key steps:
- Latent Factor Discovery (EBMF): The authors employ a recently developed Empirical Bayes Matrix Factorization (EBMF) method. EBMF is a non-orthogonal, unsupervised technique that decomposes the high-dimensional data into a smaller number of "latent factors." Unlike methods like Principal Component Analysis (PCA) that enforce orthogonality, EBMF allows for correlated latent factors, which better reflects biological reality. This step effectively reduces dimensionality and addresses multicollinearity by representing complex interactions within these latent factors. The paper uses both greedy and backfitting algorithms to fit the EBMF model.
- Causal Discovery on Latent Factors: Once latent factors are extracted, causal discovery algorithms are applied to these factors and any available clinical or outcome variables. The authors use Mixed Graphical Model (MGM) with StEPS for learning undirected graph skeletons and then PC-Max with StARS for learning a stable causal graph. This approach allows for causal inference among the reduced set of latent factors and their relationships with clinical outcomes.
Datasets
The framework was evaluated on:
- Simulated Datasets: Generated using Conditional Gaussian (CG) and Lee & Hastie (LH) models to mimic the high-dimensional and multicollinear nature of clinical and multi-omics data. These simulations included both continuous (latent factors) and discrete (clinical features) variables.
- Real-World Datasets:
- METABRIC Breast Cancer Dataset: A microarray dataset with gene expression data and 14 clinical features.
- SARS-CoV-2 RNA-seq Dataset (GSE157103): A dataset containing gene expression data from COVID-19 and non-COVID-19 patients, along with clinical features.
Key Insights and Ideas
- Latent Factors as Biological Proxies: The core idea is that latent factors derived from gene expression data can act as proxies for underlying biological processes (e.g., transcription factor activity, signaling pathways) that are not directly measured.
- EBMF's Advantage over PCA: EBMF's ability to learn non-orthogonal latent factors is crucial. This allows it to capture more biologically relevant relationships compared to PCA, which forces orthogonality and can obscure these interactions.
- Causal Discovery on a Reduced Space: Performing causal discovery on a smaller set of interpretable latent factors is more robust and computationally feasible than on the original high-dimensional, multicollinear data.
- Bridging Mechanistic and Predictive Goals: The framework aims to achieve both mechanistic understanding (identifying causal relationships) and predictive power (forecasting outcomes).
Key Results
- Synthetic Data:
- EBMF (especially with backfitting) successfully recovered the number of latent factors and achieved significantly higher Mean Correlation Coefficient (MCC) with true latent factors compared to PCA.
- The causal discovery on EBMF latent factors showed comparable or better performance than using true latent factors in recovering relationships between continuous and discrete variables.
- The approach demonstrated higher adjacency recall and precision than FCI for identifying confounded features.
- Breast Cancer Dataset:
- Identified 328 latent factors from gene expression data.
- Latent factors in the Markov blanket of clinical features (e.g., ER status, PR status, tumor size, distant relapse) were associated with biologically meaningful pathways (e.g., immune response, protein targeting, RNA translation).
- These factors were used to stratify patients into distinct survival groups, demonstrating their clinical relevance.
- SARS-CoV-2 Dataset:
- Identified 40 latent factors.
- The framework could predict COVID-19 disease state and ICU admittance with high accuracy (AUCs around 0.93 for ICU and 0.81 for disease state using linear models on latent factors).
- Latent factors associated with disease state showed significant differences between COVID-19 and non-COVID-19 patients and were linked to immune response pathways.
- Latent factors associated with ICU admittance also showed significant differences and were linked to pathways relevant to disease severity.
Limitations and Future Work
- mRNA-mRNA Interactions: The model's assumption of limited direct mRNA-mRNA interactions might overlook some regulatory mechanisms, especially when considering non-coding RNAs.
- Linearity Assumption: While the framework uses linear models for prediction, biological relationships can be non-linear. The authors suggest that non-parametric models could further improve predictive accuracy.
- Cyclic Relationships: The causal discovery algorithms used assume Directed Acyclic Graphs (DAGs), which might not fully capture all biological feedback loops.
- Future Work: The authors suggest exploring non-parametric models for prediction and investigating the impact of RNA-RNA interactions. They also highlight the potential for this framework to be applied to other high-dimensional, multicollinear biomedical datasets.
Summary updated