🧬 genomics

cit_gsa: gene set analysis for scRNA-seq with covariate adjustment

The paper introduces cit_gsa, a novel one-step self-contained gene set differential expression analysis method for single-cell RNA-seq that supports complex multi-condition comparisons and covariate adjustment, demonstrating superior performance and improved biological interpretation compared to existing two-step approaches.

Original authors: Fallet, S., Godot, V., Neuvial, P., Hejblum, B. P.

Published 2026-10-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Fallet, S., Godot, V., Neuvial, P., Hejblum, B. P.

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

Imagine trying to understand a bustling city by listening to a single microphone placed in the town square. You would hear the average hum of traffic and conversation, but you would miss the specific arguments in the alleyways, the quiet conversations in the parks, or the unique rhythm of a single street corner. This is the difference between studying genes in a bulk sample, where thousands of cells are mashed together to find an average, and studying them one by one. Single-cell technology allows scientists to listen to each cell individually, revealing a world of diversity hidden within a single drop of blood or tissue. However, this new level of detail brings a new problem: the data is incredibly noisy, filled with gaps where a gene's signal simply vanished, and the cells are not independent actors but part of a complex, interconnected community. To make sense of this, researchers often group genes into teams based on what they do together, hoping that looking at the team's performance is clearer than trying to track every single player.

The challenge has been finding the right way to listen to these gene teams. Most existing methods for comparing groups of cells are like a simple two-sided debate; they can only compare Group A against Group B. They also struggle to account for other factors that might be influencing the results, such as the age of the cells or their specific type. If a researcher wants to know if a gene team is active in a disease state, but the disease also changes the types of cells present, old methods might mistake the cell type change for a disease signal. Furthermore, many current tools rely on a two-step process: first, they check every single gene to see if it changed, and then they try to guess if the teams changed based on those individual results. This approach often misses subtle but important shifts in how genes behave together.

In a new study, a team of researchers introduced a method called cit gsa to solve these problems. Instead of a two-step process, this method looks at the entire gene team in a single, unified step. It is designed to handle complex scenarios where more than two conditions are being compared, and crucially, it can adjust for other variables that might be clouding the view. Think of it as a way to listen to a choir while simultaneously filtering out the sound of the wind or the traffic outside, ensuring that what you hear is truly the choir's performance. The researchers tested this new tool against older, established methods using simulated data where they knew exactly what the answer should be. They created digital models of cells with specific patterns of gene activity, including scenarios where the genes didn't just get louder or quieter, but changed their entire pattern of behavior.

The results showed that the new method was significantly better at finding the truth in these complex situations. While older methods could detect simple changes, like a gene team getting generally more active, they often failed when the genes changed their behavior in more subtle, complex ways. The new approach, however, caught these intricate shifts. In the simulations, it proved more powerful than the standard competitive methods, which try to rank gene teams against the rest of the genome, and also outperformed other self-contained methods when the changes were not just about average levels but about the shape of the data itself. This suggests that by looking at the whole picture at once and accounting for the background noise, the new tool can reveal biological signals that were previously invisible.

To see if this worked in the real world, the researchers applied the method to a dataset of human immune cells known as dendritic cells. These cells act as the body's scouts, capturing information about invaders and presenting it to the immune system. The team wanted to find out which groups of genes were linked to the activation of these cells, specifically looking at a marker called CD83. When they analyzed the data without adjusting for the different subtypes of dendritic cells, they found that hundreds of gene teams seemed to be associated with activation. It looked like a massive, widespread response. However, when they used the new method to adjust for the fact that different subtypes of cells naturally have different gene profiles, the picture changed dramatically. The number of significant gene teams dropped from hundreds to just a dozen.

This shift revealed a critical insight: most of the apparent connections between gene teams and cell activation were actually just reflections of the different types of cells present, rather than a direct response to activation. By filtering out the influence of cell type, the new method isolated the true biological signal. The few gene teams that remained significant were likely the ones genuinely driving the activation process, free from the confusion of cell diversity. This demonstrates that the ability to adjust for covariates is not just a statistical technicality, but a necessary step to avoid drawing false conclusions. The researchers conclude that this approach offers a more accurate way to interpret single-cell data, allowing scientists to distinguish between what is truly happening in the biology and what is merely an artifact of the sample's composition. While the method currently treats each cell as an independent observation, the authors note that future work will aim to account for cells that come from the same biological sample, further refining the tool for even more complex studies.

Technical Summary of "cit gsa: gene set analysis for scRNA-seq with covariate adjustment"

Problem Statement
Single-cell RNA sequencing (scRNA-seq) enables high-resolution gene expression analysis but presents challenges including high dimensionality, dropout events, and complex dependencies between cells and genes. While Differential Expression Analysis (DEA) at the gene level is common, analyzing gene sets offers greater statistical power and biological interpretability. However, existing single-cell gene set analysis methods face significant limitations:

Two-step approaches: Many methods rely on a gene-level analysis followed by an enrichment step, which may propagate errors or lose information.
Limited experimental designs: Most methods are restricted to two-group comparisons and struggle with more than two conditions or continuous variables.
Lack of covariate adjustment: Existing tools often fail to adjust for confounding covariates (e.g., cell type), which can lead to spurious associations driven by biological heterogeneity rather than the factor of interest.
Distributional assumptions: Some methods rely on specific count distributions (e.g., negative binomial) that may not capture complex expression patterns like changes in modality.

Methodology: cit gsa
The authors introduce cit gsa, a novel, self-contained, one-step gene set DEA method designed specifically for scRNA-seq data.

Statistical Framework: The method extends the citcdf approach (Gauthier et al., 2021) from single genes to gene sets. It tests the null hypothesis of conditional independence between a vector of gene expressions ( $Y$ ) within a set and a factor of interest ( $X$ ), given a set of covariates ( $Z$ ). Formally, $H_0: Y_g \perp X | Z$ for all genes $g$ in the set.
Conditional Distribution Testing: Instead of modeling mean expression, cit gsa tests the equality of conditional cumulative distribution functions (cdfs): $H_0: F_{Y_g|X,Z} = F_{Y_g|Z}$ . This allows the detection of differences in the entire distribution of expression, including changes in modality, not just the mean.
Implementation:
- The method discretizes the expression space using a sequence of ordered thresholds ( $\omega_j$ ).
- It estimates conditional expectations using a sequence of logistic regressions: $h(E[1\{Y_{ig} \leq \omega_j\} | X_i, Z_i]) = \beta_{0jg} + X_i\beta^X_{jg} + Z_i\beta^Z_{jg}$ .
- The null hypothesis is tested by aggregating the coefficients associated with the factor of interest ( $\beta^X$ ) using a score test statistic ( $\hat{S}_n$ ).
- The resulting p-value is derived by comparing the observed statistic to a mixture of $\chi^2_1$ distributions, accounting for the covariance structure between genes within the set.
Flexibility: The method supports multivariate factors of interest and covariates, handling both discrete and continuous variables without strong parametric assumptions.

Key Contributions

One-Step Self-Contained Analysis: Unlike competitive or two-step methods, cit gsa performs gene set testing in a single step, directly testing the joint distribution of the gene set.
Covariate Adjustment: It explicitly models and adjusts for covariates (e.g., cell subpopulations), ensuring that detected associations are not confounded by known biological variables.
Complex Scenario Handling: The method is designed to handle more than two experimental conditions and continuous variables, moving beyond simple pairwise comparisons.
Robustness to Distributional Changes: By relying on conditional cdfs rather than mean regression, the method can detect complex differential patterns, such as differences in the modality of gene expression distributions.

Results

Simulation Studies:
- Comparison: cit gsa was benchmarked against globaltest (a self-contained regression-based method), GSEA, and the hypergeometric test (competitive methods).
- Scenarios: Simulations covered "Difference in Mean" (DMean) and "Difference in Modality" (DMod) scenarios with varying gene correlations.
- Findings: In the DMean scenario, cit gsa performed comparably to globaltest. However, in the DMod scenario (where the number of modes in the distribution changes), cit gsa demonstrated superior power. Competitive methods generally showed lower power, consistent with the theoretical restrictiveness of self-contained null hypotheses.
Real-Data Application (Dendritic Cells):
- Dataset: Analysis of 539 single dendritic cells (DCs) from Villani et al. (2017) to identify gene sets associated with the activation marker CD83.
- Covariate: DC subpopulation was included as a covariate.
- Findings: Without covariate adjustment, 237 out of 259 Blood Transcription Modules (BTMs) were significantly associated with CD83. After adjusting for DC subpopulations, only 12 gene sets remained significant. This indicates that many apparent associations were driven by differences between DC subsets rather than CD83 expression itself, highlighting the necessity of covariate adjustment for accurate biological interpretation.

Significance and Claims
The paper positions cit gsa as a tool that improves the accuracy and interpretability of gene set analysis in scRNA-seq by addressing the limitations of current methods regarding experimental design complexity and confounding factors. The authors claim that:

Covariate adjustment is critical for distinguishing true biological signals from artifacts caused by cellular heterogeneity.
Self-contained methods are statistically more interpretable and powerful for detecting gene set associations than competitive methods.
The ability to detect non-linear distributional changes (modality) makes the method particularly suitable for the complex nature of single-cell data.
While currently focused on scRNA-seq, the non-parametric nature of the approach allows for potential application to bulk RNA-seq data.

The authors acknowledge a current limitation: the method does not yet account for correlations between cells originating from the same biological sample (multi-sample correlation), noting that extending the framework to handle this is a direction for future work.

Technical Summary of "cit gsa: gene set analysis for scRNA-seq with covariate adjustment"

More like this