Causal effect estimation from trans-regulatory single-cell CRISPR screens

Authors: Oliver P. Christensen, Alex Markham, Hyunseung Kang, Erin Gabriel, and Tune H. Pers.

Abstract: Recent advances in single-cell transcriptomics and CRISPR-based genome editing have enabled large-scale perturbation experiments with genome-wide expression readouts. Single-cell CRISPR screens therefore offer the opportunity to move beyond correlation and estimate causal effects of genetic perturbations on gene expression at scale. These approaches promise to substantially deepen insights into cellular functions and disease mechanisms. However, interpreting statistical associations as causal effects requires additional assumptions beyond those needed for standard statistical analyses. In this review, we introduce key concepts and principles for causal effect estimation in trans-regulatory single-cell CRISPR studies. We describe a set of assumptions under which estimates from existing statistical methods admit a causal interpretation and provide a concise overview of these approaches. Finally, through an illustrative example, we demonstrate how violations of these assumptions can bias estimated effects.

---

Worked Example: Violations of Causal Assumptions in a Simple Simulation

Overview

This simulation study illustrates how violations of key causal inference assumptions can bias estimated treatment effects in a setting resembling a trans-regulatory single-guide single-cell CRISPRi screen.

We consider:

• A binary perturbation indicator t (gRNA targeting gene X present vs absent)
• A count outcome Y (expression of gene Y)
• A Poisson data-generating process
• True causal log fold change = −1 (fold change ≈ 0.368)

For each scenario, data are simulated, a Poisson regression model (Y ~ t) is fitted, and estimates are averaged across Monte Carlo replicates.

---

Scenarios

Baseline (No Assumption Violations)

Treatment assignment is randomized, and there are no confounders, effect modifiers, or heterogeneous treatment versions.

All causal assumptions are satisfied, and the estimated effect recovers the true causal fold change.

---

Consistency Violation

Multiple gRNAs targeting gene X have different knockdown efficiencies but are combined into a single treatment category.

Specifically:

• gRNA1: moderate effect
• gRNA2: strong effect
• gNT1: non-targeting control

Aggregating these guides into one indicator means that “treatment” does not correspond to a single well-defined intervention, leading to biased estimates.

---

No Interference Violation

An environmental variable modifies the treatment effect.

The outcome depends on the interaction between treatment and environment:

effect = beta1 × t × env

This creates heterogeneous treatment effects across cells, violating the assumption that outcomes depend only on a unit’s own treatment.

---

Conditional Ignorability Violation From Confounder

An unobserved binary variable C:

• Increases the probability of receiving treatment
• Directly increases the outcome Y

Because C is omitted from the analysis model, treatment assignment is confounded, producing biased effect estimates.

---

Conditional Ignorability Violation From Collider

An observed variable C:

• Is affected by both treatment and outcome Y

Because the analysis conditions on C, this opens a backdoor path and induces bias in the estimated treatment effect.

---

Output

For each scenario, the script reports:

• Estimated baseline mean expression
• Estimated treatment fold change

---

Requirements

R packages:

• tidyverse
• broom

Install if needed:

install.packages(c("tidyverse", "broom"))

---

Reproducibility

A fixed random seed is used to ensure reproducible results.

---

Interpretation

The simulations demonstrate that even simple violations of causal assumptions can produce substantial bias in estimated effects, emphasizing the importance of carefully considering these assumptions when interpreting results from CRISPR screening experiments or similar observational analyses.