Incorporating cell states for mapping eQTLs in single-cell studies

Genes are the instruction manuals for building and running a living organism, but they do not operate in a vacuum. Their activity is constantly adjusted by tiny variations in the DNA code itself, known as genetic variants. Scientists call the specific locations where these variants influence gene activity "expression quantitative trait loci," or eQTLs. For decades, researchers have mapped these relationships by studying groups of cells mixed together, like taking a smoothie of a tissue and measuring the average flavor. This approach has revealed how genes are regulated across different organs and broad cell types. However, the human body is not a smoothie; it is a complex ecosystem where individual cells within the same tissue can be in vastly different states of activity, maturity, or response to their environment. A cell might be resting, dividing, or fighting an infection, and its genetic regulation can change depending on that specific state.

The challenge for modern science is that these subtle, state-dependent changes are often lost when cells are blended together. Furthermore, the data collected from individual cells is incredibly sparse and noisy, making it difficult to apply standard statistical tools without running into errors. Researchers needed a way to look at the genetic regulation of genes not just across cell types, but across the continuous spectrum of cell states, while also handling the messy reality of single-cell data. Without a method that could navigate this complexity, many important genetic signals remained hidden, obscured by the averaging process or dismissed as statistical noise.

To solve this, a team of statisticians and biologists from Tsinghua University, the Capital University of Economics and Business, and Yale University developed a new framework called scarf-QTL. This method is designed to map how genetic variants influence gene expression as cells move through different states, such as during their development or in response to stimuli. Instead of grouping cells into rigid, pre-defined buckets, the researchers treated cell state as a continuous coordinate, similar to a timeline or a spectrum. They modeled gene expression as a smooth function moving along this coordinate, allowing them to capture how a genetic effect might be strong in one part of the spectrum and weak in another.

The core innovation of scarf-QTL lies in how it handles the mathematics of association. Traditional methods often assume that gene expression follows a perfect bell curve, an assumption that frequently fails with single-cell data, which is full of zeros and irregular spikes. When these assumptions are wrong, standard tests can produce false alarms or miss real signals. The new method uses a "retrospective" approach. Instead of asking how gene expression changes given a specific genotype, it asks how the genotype is distributed given the observed gene expression. This subtle shift in perspective makes the test robust against the messy, non-normal distributions typical of single-cell data. It allows the researchers to use a computationally efficient model while still maintaining strict control over false positives, ensuring that the signals they find are real.

The researchers tested their method extensively using computer simulations before applying it to real data. They created artificial datasets that mimicked various biological scenarios, including cells with constant genetic effects, effects that increased linearly, and complex patterns like peaks or cycles. In these simulations, the new method successfully controlled the rate of false alarms, even when the data was generated from distributions that did not follow standard statistical rules. In contrast, older methods that relied on strict assumptions about data distribution began to produce too many false positives. When it came to finding real signals, the new method proved more powerful than the standard "pseudobulk" approach, which averages cells together. It was particularly effective at detecting genetic effects that were concentrated in specific cell states or that changed direction, patterns that the averaging method often smoothed out until they disappeared.

The team then applied scarf-QTL to the OneK1K dataset, a massive collection of single-cell RNA sequencing data from 982 healthy individuals, containing over 1.2 million blood cells across 14 different immune cell types. In this real-world test, the method identified 30% more genes with significant genetic regulation than the previous pseudobulk analysis. This increase was not just a matter of finding more noise; the new method uncovered specific genes that were missed entirely by the older approach. For example, in dendritic cells, a type of immune cell crucial for fighting infection, the method revealed that genetic regulation often varies depending on the specific subset of the cell. One gene, known to be a marker for a specific subset of dendritic cells, showed a strong genetic effect only in cells that had progressed to a certain state. The older method, which treated all dendritic cells as a single group, failed to detect this association because the effect was diluted when averaged across the entire population.

By clustering the patterns of these newly discovered genetic effects, the researchers found distinct groups of genes that behaved differently as cells changed state. Some genes showed increasing genetic influence as cells moved toward a conventional dendritic cell state, while others showed decreasing influence as cells moved toward a plasmacytoid state. These patterns aligned with known biological functions, such as antigen presentation, confirming that the method was capturing genuine biological regulation rather than statistical artifacts. The ability to see these continuous, state-dependent changes suggests that many genetic variants act not as simple on/off switches, but as dials that turn up or down depending on the cell's current condition.

The computational efficiency of the method was also a key finding. Despite the complexity of modeling millions of individual cells, the algorithm scaled well, taking roughly the same amount of time as the simpler pseudobulk methods when analyzing typical datasets. This efficiency is achieved by using mathematical shortcuts that reduce the dimensionality of the problem without losing the nuance of the single-cell data. This means the method can be applied to even larger and more diverse datasets as they become available, without requiring prohibitive computing power.

The study concludes that while the new method has limitations, such as the difficulty of directly testing whether an effect varies across states or the need for careful choices in defining the cell-state coordinate, it represents a significant step forward. It provides a rigorous, scalable way to uncover the hidden layers of genetic regulation that exist within the dynamic landscape of single cells. By moving beyond the static view of cell types and embracing the continuous nature of cell states, researchers can now map the genetic architecture of the immune system with a level of detail that was previously out of reach. The findings suggest that the genetic regulation of our cells is far more fluid and context-dependent than previously understood, opening new avenues for understanding how genetic variation influences health and disease at the most fundamental level.