scEPS integrates genetic and single-cell disease atlas data to provide granular mechanistic insights into complex human diseases

Bridging the Gap Between Genetic Risk and Cellular Action

Why do some people carry high-risk genetic variants for Alzheimer's but remain cognitively healthy, while others progress rapidly? Understanding the biological link between a person's DNA and their actual disease symptoms is one of the most difficult challenges in modern medicine. Researchers want to combine Genome-Wide Association Studies (GWAS)—large-scale studies that identify genetic variants associated with diseases—with single-cell data, which provides a high-resolution map of how individual cells behave.

Currently, scientists struggle to connect these two worlds. Most existing methods look for "enrichment," essentially checking if disease-associated genes happen to be active in certain cell types. However, these methods often fail to model the actual disease itself. This creates a "chicken and egg" problem. It is difficult to tell if a change in a cell is a cause of the disease or merely a downstream response to the damage already done.

A new study introduces scEPS (single-cell Expression exPlainability Statistics), a method designed to solve this by explicitly modeling the disease phenotype (the observable physical characteristics or symptoms of a disease). Instead of just looking for active genes, scEPS asks whether the expression of disease-linked genes actually explains the variation in disease severity across different patients. This allows researchers to distinguish between the early genetic drivers of a disease and the later, symptomatic biological responses.

The limitation of enrichment-based mapping

Existing integrative approaches typically treat the connection between genetics and cells as a simple overlap problem. They ask: "Are the genes we flagged in the GWAS study highly expressed in this specific lung cell?" While useful, this approach is indirect. It doesn't account for how much a specific gene's activity actually contributes to the clinical reality of the patient.

As noted in the paper, these methods often struggle to identify the specific cell populations that translate genetic signals into physical disease. Furthermore, because they don't model the disease outcome directly, they are highly susceptible to reverse causation (when a disease causes a change in a cell, rather than the cell change causing the disease). In a diseased state, cells undergo massive transcriptional shifts. An enrichment method might flag a cell type simply because it is reacting to the disease, not because it is driving it. This makes it difficult to pinpoint the true mechanistic "neighborhoods"—small, functionally similar groups of cells—where the disease actually takes hold.

Modeling disease as an outcome of expression

To move beyond simple overlap, the authors developed scEPS to treat the disease phenotype as a mathematical outcome of gene expression. The core logic is built on a comparison of explanatory power. The method calculates a $d$ statistic. This represents the difference in the variance of a disease explained by GWAS-prioritized genes versus the variance explained by a set of randomly selected control genes. These control genes match the GWAS genes in mean expression.

The architecture follows a structured pipeline as illustrated in :

1. Gene Prioritization: The method uses MAGMA to identify genes with strong associations in GWAS summary statistics.
2. Neighborhood Definition: Rather than looking at isolated cells, scEPS defines "cell neighborhoods" using a k-nearest-neighbor (k-NN) graph. Think of this like grouping houses in a neighborhood based on their similarity rather than just looking at individual properties. This balances the granularity of the data with enough donor information to be statistically meaningful.
3. Linear Modeling: For every neighborhood, the authors model the disease phenotype as a linear function of the pseudo-bulk expression (the combined expression of all cells in a group) of these genes.
4. Variance Comparison: The $d$ statistic is then derived by subtracting the variance explained by control genes from the variance explained by GWAS genes.

Crucially, the authors apply this to Polygenic Risk Scores (PRSs) in healthy donors. A PRS represents a person's inherent genetic predisposition. Using it allows scEPS to capture the genetic covariance between expression and disease. This helps to bypass the noise of symptomatic responses.

Outperforming current benchmarks

The study demonstrates that scEPS is significantly more sensitive at detecting disease-associated cell populations than existing tools. When tested across eight different neurological and respiratory disorders, the authors report that scEPS identified 1.77× more significant associations than a modified CNA approach. It also identified 5.13× more associations than the scDRS method. These increases mean the tool can uncover many more relevant cell types that previous methods missed.

The power of this approach is most evident in its ability to distinguish between different stages of disease. In the analysis of Alzheimer's disease, the authors found that clinical diagnoses and AD PRSs implicated different cell populations. For instance, microglia-PVM subtypes showed high associations with clinical dementia. Meanwhile, different neuronal pathways were more prominent in the PRS analysis [Figure 3b].

Furthermore, the method provided more granular biological insights. Through Gene Set Enrichment Analysis (GSEA)—a method to see if specific biological pathways are overrepresented in a dataset—the authors found that scEPS could pinpoint specific pathways. Examples include RNA metabolism in neurons for AD risk or collagen degradation in the lungs for fibrosis [, Figure 6].

Technical trade-offs and blind spots

Despite its performance gains, scEPS is not a universal panacea. The method carries specific requirements and assumptions that practitioners must consider. First, the statistical power is heavily dependent on the diversity of the dataset. The authors recommend using single-cell data containing at least 20 donors with varying disease phenotypes.

There are also inherent modeling trade-offs. The method assumes a linear relationship between gene expression and the disease phenotype. It also assumes that the effect sizes of different genes are independent. If a disease is driven by complex, non-linear gene-gene interactions, scEPS might underestimate the true biological complexity. Additionally, the authors note that the method does not currently account for the correlation in MAGMA statistics caused by linkage disequilibrium (the tendency for certain genetic variants to be inherited together). This could potentially bias results toward genes located near very strong GWAS loci.

The verdict: A new lens for mechanistic discovery

scEPS is a significant step forward for anyone attempting to move from "what genes are active" to "how do these genes drive disease." By shifting the focus from mere enrichment to the actual explanation of phenotypic variance, it provides a much clearer view of the transition from genetic risk to clinical symptom.

The tool is ready for research environments, particularly for those looking to differentiate between the drivers of early-stage disease and the markers of late-stage pathology. Code for the implementation is reportedly available; see the paper for the canonical link. For researchers aiming to build a complete picture of disease, the authors suggest that scEPS should not replace existing methods like CNA or scDRS. Instead, it should be integrated with them to capture the full spectrum of cellular involvement.

Figures from the paper