Single-cell references can reverse, inflate and blur the associations inferred from bulk deconvolution
This study demonstrates that the choice of single-cell reference can systematically distort, reverse, or inflate associations in bulk deconvolution due to unaccounted cell-state shifts and donor sampling noise, necessitating references that resolve dynamic states and statistical models that account for reference uncertainty to ensure accurate inference.
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 This is an AI-generated explanation of a preprint that has not been peer-reviewed.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 the population of a bustling city by taking a single, blurry photograph of the entire skyline. You cannot see the individual people, only the general shapes of buildings. To learn who lives there, you might compare that blurry photo to a detailed map of the city's districts. If the map is accurate, you can estimate how many people live in each neighborhood. This is essentially what scientists do when they study human biology using a technique called deconvolution. They take a sample of tissue, which is a mixture of many different cell types, and try to figure out the proportion of each cell type present. To do this, they rely on a "reference," which is a detailed catalog of single cells, usually gathered from a small group of volunteers. This reference acts as the map. For years, researchers have treated this map as perfect and exact, assuming that any conclusions drawn about disease or aging from their estimates were solid. However, a new study suggests that the map itself might be flawed in ways that completely change the story it tells.
The researchers, led by Yunhao Jiang at the University of Edinburgh, set out to test whether the way this reference map is built could actually reverse or invent scientific conclusions. They used two large, real-world datasets where they knew the true answers. The first involved blood samples from 981 donors, where the actual mix of immune cells was known. The second involved pancreatic islet cells from 48 donors, where the relationship between cell types and blood sugar levels was known. In both cases, they simulated the process of taking a bulk sample and trying to guess the cell proportions using different versions of the reference map. They tested these maps against four different mathematical methods used to make the guesses, ensuring that any errors came from the map itself, not just the calculator.
The results were startling. In the blood study, the researchers looked at how the proportion of a specific immune cell, the CD8 T-cell, changed with age. The true data showed that these cells decreased slightly as people got older. Yet, when the scientists used 80 different reference maps, every single one of them flipped the conclusion, suggesting that these cells actually increased with age. The error was not a small mistake; it was a complete reversal of the truth. The cause was a subtle shift in the cells themselves. As people age, CD8 T-cells change their internal state, becoming more like "effector-memory" cells. The reference maps used in the study averaged all these cells together into one generic type, failing to notice this shift. Because the average profile did not match the aging cells, the math got confused and produced the wrong answer. Similarly, in the pancreas study, a weak trend where alpha-cells increased with blood sugar levels became a strong, statistically significant finding in many of the reference maps, simply because the error in the map grew larger as blood sugar levels rose.
Crucially, the study found that this problem is not fixed by changing the mathematical method used to do the counting. The researchers tested four different classes of algorithms, including regression models and Bayesian approaches. The reversal of the age trend and the inflation of the blood sugar trend happened across all of them. The only thing that changed was the direction of the error for the pancreas cells; some methods made the trend look stronger, while one specific algorithm made it look weaker. This proves that the source of the error is the reference map itself, not the tool used to read it. Even more surprisingly, the researchers found that a map with higher accuracy did not necessarily lead to the correct conclusion. One set of reference maps was more accurate in predicting cell proportions than another, yet only the less accurate set gave the correct direction for the age trend. This means that simply checking if a map is "accurate" is not enough to ensure it tells the truth about biological changes.
The study also identified two distinct types of errors hidden in these maps. The first is a systematic error, which happens when the map fails to capture the specific states of cells that change with age or disease. The only way to fix this is to build a reference that resolves these shifting states, breaking down broad cell types into their specific sub-states. The second type is a random error caused by the specific group of donors used to build the map. If a reference is built from a small number of people, the random noise from those individuals can be as large as the natural variation seen in the bulk samples. The researchers showed that this random noise can be measured and accounted for by using a statistical technique called a "donor bootstrap," which essentially simulates drawing different groups of people from the same pool to see how much the results would wiggle.
Finally, the team looked at how the reference maps should be constructed to avoid these pitfalls. They analyzed 226 public datasets and found that in 60% of them, the median donor contributed fewer than 10 cells for a given cell type. This is a critical threshold; below this number, the signal from a donor is mostly noise. The researchers proposed a new way to weigh the data from each donor, giving less influence to those with very few cells and more influence to those with many. This adjustment reduced the error significantly for one of the main methods used in the field. They also found that for a fixed amount of data, it is better to spread the cells across many different donors rather than collecting many cells from just a few people. This approach, which aligns with power analyses for single-cell studies, helps to average out the random noise of individual donors.
The implications of this work are profound for how biological data is interpreted. It suggests that conclusions drawn from deconvolution are not just about the biology of the tissue being studied, but are deeply entangled with the construction of the reference map. A conclusion about aging or disease could be an artifact of how the reference was built, rather than a real biological phenomenon. The researchers argue that future studies must use references that resolve the specific states of cells known to shift with the variables being studied. They also recommend that scientists test their findings with more than one mathematical method and report a wider range of uncertainty that accounts for the specific donors used in the reference. By acknowledging that the map can decide the answer, the scientific community can begin to build better maps and draw more reliable conclusions about the complex machinery of the human body.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.