Causal differential expression analysis under unmeasured confounders with causarray

The paper introduces causarray, a robust causal inference framework that leverages generalized confounder adjustment and semiparametric machine learning to accurately identify causal gene expression effects in single-cell and pseudo-bulk genomic data despite unmeasured confounders, as demonstrated in applications to autism and Alzheimer's disease studies.

The Gist

Imagine you are a detective trying to solve a mystery: What actually causes a specific change in a cell?

In the world of biology, scientists use powerful microscopes (single-cell sequencing) to look at individual cells. They want to know: "If I tweak this gene (the treatment), does it cause the cell to behave differently?"

However, there's a big problem. Cells are messy. They are influenced by hidden factors like their age, their "mood" (cell cycle stage), or even the batch of chemicals used to test them. These hidden factors are confounders. They are like invisible puppet masters pulling strings on both the gene you are studying and the cell's behavior. If you don't account for them, you might think Gene A caused the change, when really it was just the cell being tired or the experiment being slightly off.

This paper introduces a new detective tool called causarray. Here is how it works, explained simply:

1. The Problem: The "Noisy Room" Analogy

Imagine you are trying to hear a specific conversation in a crowded, noisy room.

• The Conversation: The effect of a gene treatment (e.g., "Gene X makes the cell grow").
• The Noise: Unmeasured confounders (e.g., "The cell is from an older mouse," or "The lab temperature was high today").

Old methods tried to solve this by just turning up the volume (looking at more data) or assuming the noise was random. But in biology, the noise is often structured and connected to the conversation. If you don't filter it out correctly, you hear the wrong story.

2. The Solution: The "Smart Noise-Canceling Headphones"

causarray is like a pair of super-smart noise-canceling headphones that doesn't just cancel all noise, but specifically learns to identify and remove the hidden noise that is messing up your results, while keeping the conversation clear.

It does this in three clever steps:

Step A: Finding the Invisible Puppet Masters (Confounder Estimation)

First, causarray looks at the entire dataset to find patterns that aren't explained by the known variables (like age or sex). It uses a statistical trick called a Generalized Factor Model.

• Analogy: Imagine you have a giant spreadsheet of cell data. causarray looks for "ghost columns"—patterns that appear across many genes at once but weren't measured. It realizes, "Ah, these 500 genes are all acting weird because of a hidden factor (like a specific batch effect), not because of the treatment." It creates a map of these invisible factors.

Step B: The "What If" Game (Counterfactuals)

Once it knows what the hidden factors are, it plays a game of "What If?"

• Analogy: It asks, "If this specific cell had received the treatment, but without the hidden noise, what would it look like?" Then it asks, "If it didn't get the treatment, but without the noise, what would it look like?"
• By comparing these two "what if" scenarios, it isolates the true effect of the treatment, stripping away the bias.

Step C: The Flexible Detective (Semiparametric Inference)

Old tools often assumed biology follows simple, straight-line rules (like a ruler). But biology is messy and curved.

• Analogy: causarray uses machine learning (like Random Forests and Neural Networks) as its flexible ruler. It can bend and shape itself to fit the complex, non-linear reality of how genes actually work. This ensures that even if the model isn't perfect, the final conclusion is still trustworthy.

3. Real-World Cases: Solving Two Big Mysteries

The authors tested causarray on two major biological mysteries:

Case 1: Autism and Brain Development (The Perturb-seq Study)

• The Mystery: Scientists used CRISPR (gene scissors) to cut specific genes in mouse brains to see which ones cause autism-like traits.
• The Old Way: Previous methods were too noisy. They found some results, but they were fuzzy and often pointed to general "cell stress" rather than specific brain functions.
• The causarray Way: It cut through the noise. It found that specific autism-risk genes directly caused changes in synapse organization (how brain cells talk to each other) and neuron development. It gave a much clearer picture of why these genes matter.

Case 2: Alzheimer's Disease (The Human Brain Study)

• The Mystery: Scientists looked at brain tissue from people with Alzheimer's to find which genes are "broken" by the disease.
• The Old Way: Different studies gave different answers because of hidden differences in the patients (age, sex, how the tissue was stored).
• The causarray Way: It acted as a universal translator. It found the same consistent set of broken genes across three different datasets. It revealed that the disease specifically attacks pathways related to synaptic signaling and cell development, giving doctors a clearer target for new drugs.

Why This Matters

Before this, scientists often had to guess which results were real and which were just statistical flukes caused by hidden variables.

causarray is like upgrading from a blurry, old camera to a high-definition, AI-enhanced lens. It allows researchers to:

1. See the truth: Separate the real cause from the background noise.
2. Be confident: Know that the genes they find are actually causing the disease, not just hanging out with it.
3. Move faster: Accelerate the discovery of treatments for complex diseases like Alzheimer's and Autism.

In short, causarray helps us stop guessing and start knowing exactly how our genes control our health.

Technical Summary

1. Problem Statement

The advent of single-cell RNA sequencing (scRNA-seq) and CRISPR-based perturbation technologies has enabled high-resolution studies of gene regulation. However, inferring causal relationships from observational genomic data remains a significant challenge due to:

• Unmeasured Confounders: Biological factors (e.g., cell cycle, developmental stage) and technical artifacts (e.g., batch effects, library size) often correlate with both the treatment (perturbation/disease status) and the outcome (gene expression).
• Limitations of Existing Methods:
  • Matching/Optimal Transport methods (e.g., CINEMA-OT, CoCoA-diff): Assume causal structures are transferable between groups and often fail when covariate distributions differ significantly, leading to biased estimates.
  • Linear Factor Models (e.g., RUV, SVA): Assume additive linear relationships between covariates and outcomes. These fail to capture the sparsity, zero-inflation, and over-dispersion inherent in single-cell count data.
  • Standard Differential Expression (e.g., DESeq2): Assumes treatment assignment is unconfounded after adjusting for observed covariates, ignoring latent confounding.

The core problem is the lack of a robust framework that can simultaneously model the specific distributional properties of single-cell count data and adjust for unmeasured confounders to provide valid causal inference.

2. Methodology: The causarray Framework

The authors propose causarray, a robust causal inference framework operating at both pseudo-bulk and single-cell levels. It integrates three main components:

A. Generalized Confounder Adjustment (GCATE)

Instead of linear models, causarray employs a Generalized Linear Model (GLM) based factor model tailored for count data (specifically Negative Binomial distribution).

• Model: The natural parameters of the gene expression distribution are decomposed as $\Theta = \tilde{X}B^\top + U\Gamma^\top$.
  • $\tilde{X}$: Observed covariates and treatment indicators.
  • $U$: Latent variables representing unmeasured confounders.
  • $B, \Gamma$: Unknown coefficients.
• Estimation: It uses an iterative algorithm (based on joint likelihood maximization with adaptive penalization) to estimate the latent confounders ($\hat{U}$) while accounting for gene-wise dispersion and zero-inflation. This allows the model to capture non-linear nuisance variation better than linear methods.

B. Semiparametric Inference (Doubly Robust Estimation)

Once confounders are estimated, causarray estimates the causal effect using the Potential Outcomes Framework.

• Target Estimand: Log Fold Change (LFC), defined as $\tau_j = \log(E[Y_j(1)] / E[Y_j(0)])$.
• Doubly Robust Estimator: It combines an outcome model ($\mu_j$) and a propensity score model ($\pi$). The estimator remains consistent if either the outcome model or the propensity score model is correctly specified.
  • Outcome Model: Generalized Linear Models (Negative Binomial with log link) to handle count data.
  • Propensity Score Model: Flexible machine learning methods (Logistic Regression or Random Forests) to model the probability of treatment assignment given covariates $W = [X, \hat{U}]$.
• Counterfactual Imputation: The method generates estimated potential outcomes ($\hat{Y}(0)$ and $\hat{Y}(1)$) for each cell, effectively denoising the data and balancing the treatment groups.

C. Statistical Inference

• Variance Estimation: Uses the influence function to compute standard errors for the causal effects.
• Multiple Testing: Implements False Discovery Rate (FDR) control using the Benjamini-Hochberg procedure and, for highly dependent test statistics, a Gaussian Multiplier Bootstrap to control the False Discovery Exceedance (FDX).

3. Key Contributions

1. Novel Framework: Introduction of causarray, the first method to explicitly combine generalized linear modeling for unmeasured confounder estimation with semiparametric causal inference for single-cell data.
2. Robustness to Distribution: By using Negative Binomial GLMs rather than linear models, it accurately models the mean-variance relationship and sparsity of scRNA-seq data.
3. Doubly Robust Inference: Ensures valid statistical conclusions even if one of the nuisance models (outcome or propensity) is misspecified.
4. Counterfactual Generation: Provides estimated potential outcomes, enabling downstream analyses such as Conditional Average Treatment Effects (CATE) stratified by covariates (e.g., age).

4. Results

A. Simulation Studies

• Benchmarking: causarray was compared against DESeq2, Wilcoxon, RUV, RUV-III-NB, CINEMA-OT, CoCoA-diff, and Mixscape.
• False Positive Control: In the presence of unmeasured confounders, causarray maintained the False Positive Rate (FPR) close to the nominal level (0.1) across various sample sizes and confounding levels. In contrast, methods like RUV and CINEMA-OT exhibited inflated FPRs (up to 0.5) as sample size increased.
• Power: causarray achieved the highest True Positive Rate (TPR), consistently reaching 0.80–0.90, significantly outperforming non-confounder-adjusted methods and other confounder-adjusted methods.
• Confounder Recovery: It successfully disentangled treatment effects from unmeasured confounders in both response and confounder spaces, showing better mixing of treatment groups in the latent confounder space compared to competitors.

B. Real-World Applications

1. In Vivo Perturb-seq (Autism Risk Genes):

   • Dataset: Developing mouse brains with CRISPR perturbations of autism spectrum disorder (ASD) risk genes.
   • Finding: causarray identified gene-level causal effects missed by module-level analyses. For the Satb2 perturbation, it enriched for biologically relevant pathways (neuronal projection development, synapse organization) consistent with known Satb2 functions.
   • Comparison: RUV identified spurious enrichment in mitochondrial processes, likely due to residual confounding. causarray provided higher specificity.

2. Alzheimer's Disease (AD) Case-Control Study:

   • Dataset: Three independent human brain snRNA-seq datasets (ROSMAP, SEA-AD MTG, SEA-AD PFC).
   • Finding: causarray identified consistent causal gene expression changes across all three datasets, whereas RUV showed dataset-specific inconsistencies.
   • Biological Insights: Uncovered pathways related to synaptic signaling and cell development.
   • Conditional Analysis: Enabled the estimation of age-dependent treatment effects (CATE), revealing that certain genes (e.g., SLC16A6) showed stronger causal effects at extreme ages.

5. Significance

• Paradigm Shift: Moves single-cell analysis from simple association (differential expression) to rigorous causal inference, essential for understanding disease mechanisms and identifying therapeutic targets.
• Handling Complexity: Addresses the critical issue of unmeasured confounding in observational genomic studies, a major source of bias that previous methods failed to resolve effectively.
• Reproducibility: Demonstrated high reproducibility across independent datasets (AD study) and robustness to varying experimental designs.
• Future Directions: The framework is extensible to other omic modalities (ATAC-seq, proteomics) and spatial transcriptomics, offering a unified approach for causal discovery in complex biological systems.

In summary, causarray provides a statistically rigorous, flexible, and robust tool for uncovering causal gene regulatory networks in single-cell data, overcoming the limitations of linear assumptions and unmeasured confounding that have plagued previous approaches.