🧬 genomics

PETRO enables robust estimation of on-target perturbation effects in Perturb-seq

The paper introduces PETRO, a method that enables robust estimation of on-target perturbation effects in Perturb-seq by leveraging guide agreement to effectively filter out off-target biases that persist even after standard pooling and filtering.

Original authors: Sun, Z., Kang, H., Keles, S.

Published 2026-10-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Sun, Z., Kang, H., Keles, S.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the modern study of biology, scientists often need to understand how a single change in a cell's genetic code ripples out to alter the cell's behavior. To do this, they use a powerful technique called Perturb-seq. Imagine a library where every book represents a different gene. Researchers can use a molecular tool to "knock out" or switch off specific books, one by one, to see what happens when that information is missing. In a typical experiment, they might target a single gene with several different molecular guides, which act like precise scissors cutting the DNA at that specific spot. By reading the cell's entire instruction manual, known as its transcriptome, after the cut, scientists can map out the gene's true function. The logic is straightforward: if you cut the same gene with three different pairs of scissors and get the same result, you can be confident that the result comes from cutting that gene.

However, the molecular scissors are not always perfectly precise. Sometimes, a guide meant for one gene accidentally latches onto a different, similar-looking sequence elsewhere in the genome. This is called an off-target effect. It is a bit like trying to edit a specific sentence in a novel, but your editing tool accidentally changes a similar-looking word on a different page. If this happens, the cell's reaction might be a mix of the intended change and this accidental one. When researchers combine the results from multiple guides to get a clearer picture of the target gene, these accidental changes can skew the final answer, leading to incorrect conclusions about what the gene actually does. The challenge has been to separate the intended signal from the noise without throwing away valuable data.

A team of researchers has developed a new method called PETRO to solve this problem. Rather than simply discarding guides that might be causing trouble or averaging all the results together, PETRO looks for agreement among the guides. The method operates on a simple but powerful principle: if a majority of the guides targeting a gene are working correctly, their shared result represents the true effect of that gene. The method assumes that while some guides might be acting up due to off-target effects, the majority will be honest. By focusing on this majority consensus, PETRO can identify the true biological signal even when a few guides are behaving erratically. It does not need to know in advance which guides are faulty; it simply lets the data reveal which ones are in agreement and which ones are outliers.

The researchers tested this approach using data from several large-scale experiments involving human cells. In one set of experiments, they looked at how guides with specific sequence features, which are known to increase the risk of accidental binding, influenced the results. They found that when they simply pooled all the cells together, a bias remained that was directly linked to these risky sequences. Even when they removed guides that had been previously flagged as suspicious, the bias persisted and sometimes even grew stronger because the remaining guides were forced to carry more weight. In contrast, when they applied PETRO, this sequence-linked bias nearly vanished. The method successfully filtered out the noise caused by the faulty guides while keeping the valid data, all without needing to know which guides were the problem beforehand.

To ensure their method was reliable, the team also ran computer simulations where they knew the exact truth. They created scenarios where most guides were correct, but one was deliberately corrupted with a strong off-target effect. As they increased the number of cells in the simulation, the traditional method of pooling data became more precise but remained stubbornly wrong, converging on a biased answer. PETRO, however, continued to home in on the correct value, with its error shrinking as more data became available. The simulations also showed that PETRO could control the rate of false alarms, ensuring that scientists did not mistakenly claim a gene had an effect when it did not, even in complex situations where many outcomes were tested at once.

The power of this new approach became clear when the researchers re-examined a real-world study on brain development. In the original analysis, scientists had concluded that a specific gene, ZNF219, played a key role in repressing the development of certain types of neurons. This conclusion was based on pooling the results from three different guides targeting the gene. However, when the researchers looked closer, they found that the guides were actually telling different stories. One guide suggested the gene reduced neuron numbers, while another suggested it increased them. The pooled result was a confusing average that masked these contradictions. By applying PETRO, which required a majority agreement to declare a result, the researchers found that the strong consensus for the gene's role disappeared. The method revealed that the original conclusion was an artifact of combining conflicting signals, whereas the guides that did agree on other genes produced consistent and reliable results.

This work highlights a fundamental shift in how we interpret complex genetic data. It shows that simply having more data or more guides does not automatically lead to a clearer picture if those guides are not in agreement. The researchers demonstrated that by respecting the disagreement between guides and focusing on the majority, it is possible to recover the true biological effect. This does not mean that the guides causing trouble are useless; rather, it means their influence can be managed so they do not distort the final answer. The method works across different types of genetic experiments and does not require scientists to have prior knowledge of which guides might fail.

The implications of this finding extend beyond just one experiment. It suggests that many previous conclusions drawn from pooled genetic data might need to be re-evaluated, especially in cases where the guides did not fully agree. The researchers showed that their method could preserve strong biological signals while removing the noise, offering a more robust way to understand how genes function. In the end, PETRO provides a way to trust the consensus of the majority, ensuring that the story told by the data is one of the gene's true nature, rather than a mixture of its intended function and accidental side effects. This clarity allows scientists to build a more accurate map of the genetic landscape, one where the signals are clear and the noise is kept at bay.

Technical Summary: PETRO for Robust Perturb-seq Analysis

Problem Statement
Perturb-seq combines pooled CRISPR perturbations with single-cell RNA sequencing to profile transcriptome-wide responses, typically employing multiple guide RNAs (gRNAs) per intended target to serve as replicates. Standard analytical approaches (e.g., DESeq2, SCEPTRE, Mixscale) often aggregate data across these guides to estimate a target-level perturbation effect. However, this aggregation is vulnerable to off-target effects. A guide may successfully perturb its intended target while simultaneously altering gene expression through off-target binding. When such off-target effects are present, conventional pooling or simple averaging of guide effects yields biased target-level estimates. These biases persist even with efficient target perturbation and can distort biological conclusions. Furthermore, existing filtering strategies that remove "suspicious" guides based on sequence or profile diagnostics do not guarantee unbiased estimates, as they may alter the weighting of remaining guides or fail to account for outcome-specific deviations.

Methodology: The PETRO Framework
The authors present PETRO (Perturbation Effect estimation Robust to Off-target bias), a causal inference framework designed to recover the on-target perturbation effect under a "majority rule" assumption.

Model Formulation: For a target-outcome pair, the effect of guide $j$ ( $\gamma_j$ ) is decomposed into a shared on-target effect ( $\Gamma$ ) and a guide-specific off-target effect ( $\delta_j$ ): $\gamma_j = \Gamma + \delta_j$ . A guide is considered "valid" if $\delta_j = 0$ .
Identification Assumption: PETRO assumes the majority rule: more than half of the guides for a given target are valid ( $\delta_j = 0$ ) for that specific outcome. Under this assumption, the on-target effect $\Gamma$ is identified as the unique value shared by the strict majority of guides, equivalent to the population median of the guide effects. This identification holds even if the specific identity of the valid guides is unknown and the invalid minority carries arbitrary off-target effects.
Estimation Strategy: PETRO estimates $\Gamma$ using robust M-estimation with Tukey-bisquare weights. It downweights guides whose estimated effects deviate significantly from the consensus, effectively reducing the influence of the invalid minority without requiring their pre-identification or removal.
Proportional Extension: To estimate the regulatory effect ( $\alpha$ )—the change in outcome per unit change in target gene expression—PETRO employs a proportional model: $\gamma_j = \alpha\beta_j + \delta_j$ , where $\beta_j$ is the guide's effect on target expression. This allows for variation in perturbation strength across guides while maintaining robustness to off-target effects.
Inference: The framework supports both parametric (Gaussian-based) and permutation-based inference. The authors emphasize that while robust estimation addresses bias, permutation tests are necessary to ensure calibrated false discovery rates, particularly in sparse Perturb-seq data where parametric approximations may inflate Type I error.

Key Contributions and Results

Reduction of Sequence-Associated Bias: In analyses of CRISPRa and CRISPRi datasets (RPE-1, Hs27, A375, K562), PETRO demonstrated that sequence-associated off-target bias (measured via PAM-proximal seed matching) persists after conventional pooling and can even increase after post-hoc guide filtering due to weight renormalization. PETRO largely eliminated this bias without using sequence annotations, reducing residual bias to near zero in high-risk strata where pooling remained biased.
Suppression of Promiscuous Guide Effects: Using seven independently flagged "promiscuous" guides in K562 cells, the study showed that PETRO reduced the transmission of these guide-specific off-target effects into target-level estimates by approximately 90% compared to pooling. This reduction was achieved without using the external flags during the fitting process. PETRO outperformed the Canon method (a Bayesian instrumental variable approach) in minimizing the transmission of these flagged effects.
Simulation Performance: In simulations with one invalid guide among four, PETRO matched the power of an oracle estimator (which knows the valid guides) in the absence of off-target effects. When off-target effects were present, pooling remained biased as cell numbers increased, whereas PETRO's bias approached zero.
Calibration of Inference: The authors constructed empirical null distributions from control cells. They found that while parametric tests showed modest marginal miscalibration, they led to substantial inflation of familywise error rates (FWER) under multiple testing. Permutation inference within PETRO successfully controlled both marginal and familywise error rates close to nominal levels.
Biological Reinterpretation (ZNF219 Case Study): In a reanalysis of a cortical Perturb-seq study, the authors found that pooled estimates for the target ZNF219 masked conflicting signals among its three guides. While pooling suggested ZNF219 represses neuronal differentiation, the individual guides supported opposing lineage and transcriptional effects. PETRO, by requiring majority support, retained few of the pathway enrichments derived from the pooled ZNF219 profile, whereas it preserved most pathway calls for concordant targets (ARX, NR2E1). This highlights how pooling can create spurious biological interpretations when guides disagree.
Selective Correction: Unlike global filtering, PETRO selectively downweights guides based on outcome-specific agreement. The authors demonstrated that guides globally retained by published filters could still carry large, outcome-specific deviations (e.g., FTL activation by a specific PRDM5 guide), which PETRO successfully attenuated.

Significance and Claims
The paper claims that PETRO enables robust target-level Perturb-seq analysis by leveraging replicate-guide agreement under a majority rule. The authors assert that this approach:

Recovers the on-target perturbation effect even when a minority of guides carry arbitrary off-target effects.
Avoids the pitfalls of global guide removal, which can increase bias through renormalization or discard valid guides for specific outcomes.
Provides a distinct advantage over existing instrumental variable methods (like Canon) by not requiring a proportional model assumption and by allowing arbitrary off-target effects in the invalid minority.
Demonstrates that pooled biological interpretations can be highly sensitive to the combination method, as illustrated by the ZNF219 case where pooling conflated conflicting guide phenotypes.

The authors conclude that while guide filtering and robust estimation are complementary, PETRO offers a principled way to identify the on-target effect without prior knowledge of which guides are valid, provided the majority rule holds. They note limitations, including the inability to verify the majority rule from observed data alone and the potential for failure if half or more of the guides are invalid.

Technical Summary: PETRO for Robust Perturb-seq Analysis

More like this