Causal gene regulatory network inference from Perturb-seq via adaptive instrumental variable modeling

The paper introduces ADAPRE, an adaptive instrumental variable framework that leverages Perturb-seq data to accurately infer causal, potentially cyclic gene regulatory networks by accounting for heterogeneous CRISPRi knockdown efficiencies, thereby outperforming existing methods in reconstructing biologically relevant interactions.

The Gist

Imagine you are a detective trying to figure out who is boss in a massive, chaotic city called The Cell. In this city, thousands of genes (the citizens) are constantly talking to each other, telling some to work harder, others to slow down, and some to stop working entirely. Your goal is to draw a map of these relationships: a Gene Regulatory Network (GRN).

The problem? Most of the time, you can only watch the city from a window. You see Gene A and Gene B acting up at the same time, but you don't know if A told B to act up, if B told A, or if a third factor (like the weather, or a hidden gang) is making both of them act crazy. This is called confounding, and it makes drawing an accurate map nearly impossible.

The New Tool: Perturb-seq

Scientists recently invented a super-powerful tool called Perturb-seq. Instead of just watching the city, they can now send in a "disruptor" (a CRISPR guide) to specifically silence one citizen (Gene A) and see how the rest of the city reacts.

Think of it like this: You want to know if the Mayor (Gene A) controls the Police Chief (Gene B).

• Old Way: You watch them. They both look stressed. Did the Mayor stress the Chief? Or did the Chief stress the Mayor? Or is there a riot happening that stresses both? You can't tell.
• Perturb-seq Way: You quietly remove the Mayor from the office. If the Police Chief suddenly relaxes, you know the Mayor was stressing them out. If the Chief keeps panicking, maybe they were stressed by something else.

The Problem with the Old Detective Work

The paper introduces a new method called ADAPRE because the previous best detective (a method called inspre) had a major flaw.

Imagine the "disruptors" (the CRISPR guides) sent into the city aren't all equally strong. Some are like a gentle whisper, while others are like a megaphone.

• The Flaw: The old method (inspre) got confused by the volume of the megaphone. If a gene was silenced very loudly (a strong knockdown), the old method assumed that gene must be a "Super Boss" who controls hundreds of other genes. It thought, "Wow, that gene made a huge change, so it must be the most important one!"
• The Reality: The gene wasn't necessarily a boss; it just had a really loud microphone. The method was creating fake "hubs" (super-regulators) just because the experiment was louder for them.

The Solution: ADAPRE (The Smart Detective)

The authors created ADAPRE (ADAptive Penalized inverse REgression) to fix this. Think of ADAPRE as a detective who carries a calibration meter.

Here is how ADAPRE works in two simple steps:

Step 1: Listening to the "Raw Noise" (The PLN Model)

When you listen to the city, the data comes in as "counts" (like how many people shouted a specific word).

• The Old Way: The old method just took the average of the shouting and smoothed it out, ignoring that some people shout louder naturally or that the microphone quality varies.
• ADAPRE's Way: ADAPRE uses a special mathematical lens (a Poisson-Lognormal model) that understands the nature of the noise. It knows that counting people in a crowd is different from measuring the volume of a radio. It separates the "technical noise" (bad microphones) from the "real signal" (actual gene expression). This ensures the detective isn't fooled by bad equipment.

Step 2: Calibrating the Volume (Adaptive Penalties)

This is the magic trick. ADAPRE looks at how loud each "disruptor" was.

• If a gene was silenced with a megaphone (strong knockdown), ADAPRE says, "Okay, this gene made a big change, but that might just be because the mic was loud. I will penalize it. I won't let it become a 'Super Boss' just because it was loud."
• If a gene was silenced with a whisper (weak knockdown), ADAPRE says, "This gene made a change even with a whisper? That must be a real, powerful boss!" It gives this gene more credit.

By adjusting the "penalty" based on the volume of the intervention, ADAPRE stops the loud genes from hogging the spotlight and reveals the true leaders of the cell.

What Did They Find?

When the authors used ADAPRE on real data from leukemia cells (K562), they found:

1. No Fake Bosses: The "Super Bosses" that the old method found disappeared. The network looked much more realistic.
2. Real Connections: They found groups of genes that work together to fight leukemia. For example, they found a cluster of genes controlled by YY1 (involved in making RNA) and JUND (involved in stress response).
3. Consistency: Even when they tested the method on different datasets, the map looked the same. It's a reliable map.

The Big Picture

Imagine you are trying to map the influence of every person in a giant office.

• Old Method: If someone screams, you assume they are the CEO.
• ADAPRE: You realize some people just have loud voices. You adjust your map so that the person who actually gives the orders (even if they whisper) gets the credit, and the loud talker gets their proper place.

ADAPRE is a smarter, fairer way to understand how our cells are wired. It helps scientists find the true "wiring diagrams" of diseases like leukemia, which is the first step toward building better treatments.

Technical Summary

1. Problem Statement

Inferring causal Gene Regulatory Networks (GRNs) from observational single-cell data is inherently difficult due to confounding factors and complex feedback loops. While Perturb-seq (CRISPR-based perturbation coupled with single-cell RNA-seq) offers a solution by providing causal leverage, existing computational methods face two critical limitations:

1. Measurement Model Mismatch: Many methods (e.g., inspre) rely on "normalize-then-model" workflows (e.g., z-scoring) that ignore the specific statistical properties of single-cell data, such as Unique Molecular Identifier (UMI) counting processes and the Poisson mean–variance coupling. This risks attributing technical variation to biological expression.
2. Perturbation-Strength Bias: Existing methods often assume perfect interventions or fail to account for heterogeneous CRISPRi knockdown efficiencies. The authors identify a pervasive "perturbation-degree bias": genes with stronger knockdown effects (higher perturbation strength) are disproportionately inferred as network hubs (high out-degree), distorting the true network topology. This bias arises because current estimators do not explicitly model the relationship between instrument strength and network structure.

2. Methodology: ADAPRE

The authors propose ADAPRE (ADAptive Penalized inverse REgression), a two-stage framework designed to infer causal GRNs while correcting for the above biases.

Stage 1: Probabilistic Modeling of UMI Counts (PLN)

Instead of normalizing counts first, ADAPRE models the raw UMI counts ($Y_{ci}$) directly using a Poisson-lognormal (PLN) observation model.

• Observation Layer: $Y_{ci} | X_{ci}, \ell_c \sim \text{Poisson}(\ell_c e^{X_{ci}})$, where $\ell_c$ is the library size and $X_{ci}$ is the latent log-expression. This separates technical sampling (Poisson) from biological variability (Log-normal).
• Latent Structural Equation Model (SEM): The latent expressions follow a linear structural equation: $X_c = X_c B + A_c \Gamma + \epsilon_c$.
  • $B$: The direct-effect matrix (the target GRN).
  • $A_c$: Perturbation indicators (gRNA assignments).
  • $\Gamma$: Diagonal matrix of CRISPRi knockdown strengths ($\gamma_i$).
  • $\epsilon_c$: Latent noise with covariance $\Omega$.
• Identification: Using the instrumental variable (IV) assumption (gRNA assignment is random and affects expression only through the targeted gene), the total effect matrix $T = (I - B)^{-1}$ is identified via the Wald ratio of latent means:
  $$T_{ij} = \frac{E[X_j | A_i=1] - E[X_j | A_i=0]}{E[X_i | A_i=1] - E[X_i | A_i=0]}$$
  ADAPRE estimates these latent means by fitting intercept-only PLN models for each perturbation group.

Stage 2: Adaptive Sparse Inverse Regression

The goal is to recover the direct effect matrix $B$ from the estimated total effect matrix $\hat{T}$ using the relation $B = I - T^{-1}$.

• Optimization Problem: ADAPRE solves a constrained optimization problem to find a sparse inverse $V \approx T^{-1}$:
  $$\min_{U, V} \frac{1}{2} \| W \odot (\hat{T} - U) \|_F^2 + \sum_{i \neq j} \lambda \phi(\hat{\gamma}_i) |V_{ij}| \quad \text{s.t.} \quad VU = I_p$$
• Adaptive Penalty: The key innovation is the row-specific penalty weight $\phi(\hat{\gamma}_i) = |\hat{\gamma}_i| / \text{mean}(|\hat{\gamma}|)$.
  • Mechanism: Regulators with stronger knockdowns ($\hat{\gamma}_i$) receive a higher penalty, preventing them from being over-inferred as hubs. Conversely, weaker instruments receive lower penalties.
  • Result: This adaptively corrects the strength-degree bias, ensuring that network topology is driven by regulatory logic rather than perturbation efficacy.
• Algorithm: Solved using the Alternating Direction Method of Multipliers (ADMM).

3. Key Contributions

1. Explicit Count Modeling: ADAPRE is the first framework to integrate a Poisson-lognormal observation layer with instrumental variable modeling for Perturb-seq, properly disentangling technical noise from biological signal.
2. Bias Correction via Adaptive Penalization: It introduces a novel adaptive penalty scheme that mathematically corrects for the "perturbation-degree bias," ensuring that genes with strong knockdowns are not falsely identified as network hubs.
3. Cyclic Network Recovery: Unlike methods assuming acyclicity (e.g., DoTEARS), ADAPRE allows for feedback loops (cycles) in the GRN, which are common in biological systems.
4. Scalability: The method is computationally efficient, handling genome-scale datasets (thousands of cells and hundreds of genes) with distinct trade-offs between memory (Stage 1) and runtime (Stage 2).

4. Results

• Simulation Studies:
  • Bias Removal: In simulations with heterogeneous instrument strengths, ADAPRE eliminated the correlation between knockdown strength and estimated out-degree (slope $\approx 0$), whereas the baseline method (inspre) showed a strong negative correlation (stronger knockdown $\to$ higher out-degree).
  • Performance: ADAPRE achieved higher F1 scores, lower Structural Hamming Distance (SHD), and lower Mean Absolute Error (MAE) compared to inspre, LiNGAM, GIES, and IGSP, particularly in settings with weak instruments and confounding.
• Real-World Validation (K562 & teloHAEC):
  • Bias Correction: Applied to the Replogle et al. K562 dataset and Schnitzler et al. endothelial cell dataset, ADAPRE successfully removed the spurious strength-degree dependence observed in inspre.
  • Biological Enrichment: Networks inferred by ADAPRE showed significantly stronger enrichment for known biological interactions (CORUM protein complexes, STRING interactions, and ChIP-seq TF binding) across various sparsity levels compared to baselines.
  • Stability & Replicability: The inferred networks were stable across data splits and highly replicable between independent Perturb-seq datasets (GWPS vs. Essential datasets), with high sign consistency (>80%) on overlapping edges.
• Biological Insights:
  • ADAPRE identified coherent, leukemia-associated subnetworks centered on hubs like YY1, JUND, and E4F1.
  • It revealed specific regulatory mechanisms, such as the antagonistic control of YBX1 (receiving negative inputs from YY1/JUND and positive from E4F1), providing testable hypotheses for experimental follow-up.

5. Significance

ADAPRE represents a critical advancement in the field of causal inference from single-cell perturbation data. By bridging the gap between high-throughput perturbation assays and rigorous statistical modeling, it addresses the "black box" nature of previous methods that ignored count data properties and instrument heterogeneity.

• Scientific Impact: It enables the reconstruction of more accurate, cyclic, and biologically plausible GRNs, which are essential for understanding cellular identity and disease mechanisms (e.g., leukemia).
• Methodological Impact: It sets a new standard for handling heterogeneous CRISPRi efficiencies, demonstrating that adaptive regularization is necessary to recover true network topology from noisy, real-world perturbation screens.
• Future Outlook: The framework is designed to be extensible to multi-omics data (multiome) and multiplex perturbations, positioning it as a foundational tool for next-generation functional genomics.