Omitting tumor proliferation from prognostic gene expression models: a preregistered cross-cancer test, a transcriptome-wide measurement, and a survey of how often the omission is made
This preregistered study demonstrates that omitting tumor proliferation from prognostic gene expression models in hepatocellular carcinoma and other cancers systematically biases nearly all gene estimates, yet the author's specific predictions regarding the consistency and magnitude of this bias across cohorts were largely refuted by the data.
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 This is an AI-generated explanation of a preprint that has not been peer-reviewed.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 world of cancer research, scientists often look for patterns in the genetic instructions inside tumor cells to predict how a patient's disease will behave. They search for specific genes that act like warning signs, hoping to find a list of markers that can tell a doctor if a tumor is likely to grow fast or spread. To do this, researchers build statistical models that weigh the activity of thousands of genes against the survival times of patients. However, these models must be careful to account for other factors that influence survival, such as a patient's age or the size of the tumor. If a researcher forgets to include a major factor that affects both the genes and the survival time, the model can produce a misleading result. It might look like a specific gene is the cause of a poor outcome when, in reality, that gene is just a passenger on a bus driven by a different, unmeasured force. One of the most powerful of these forces is tumor proliferation, which is simply a measure of how quickly the cancer cells are dividing and multiplying. Because rapidly dividing cells often have different genetic activity than slow-growing ones, ignoring this speed can distort the entire picture of what the genes are actually doing.
A researcher named Soon Woo Nam set out to investigate how often this specific mistake happens and what it costs when it does. He began by looking at fifty recent scientific papers that focused on liver cancer. He read through them to see if the author had included a measure of cell division speed in their models. The result was stark: not a single one of the forty-one papers he could fully read had adjusted for tumor proliferation. In fact, more than half of those studies mentioned that their high-risk groups showed signs of rapid cell division, yet they treated that speed as a biological curiosity rather than a variable that needed to be controlled for in their math. This suggested that a vast number of published conclusions about which genes predict survival might be built on a shaky foundation, missing a crucial piece of the puzzle.
To understand exactly what was lost by leaving this factor out, Nam turned to a massive public database containing genetic and survival data from hundreds of liver cancer patients. He ran a comprehensive test on nearly every gene that was active in those tumors. For each gene, he built two different models to predict survival. The first model was standard, accounting for age, sex, and the stage of the cancer. The second model was identical, except it also included a score representing how fast the tumors were dividing. He then compared the results. When the speed of cell division was added to the equation, the estimated importance of almost every single gene shifted. The size of this shift was not random; it depended directly on how closely that specific gene's activity was linked to the speed of division. If a gene was strongly tied to rapid growth, its predicted impact on survival changed dramatically once the growth speed was accounted for. In the liver cancer data, this adjustment was so powerful that it caused the number of genes that appeared to be statistically significant to collapse from hundreds down to just one.
Nam then asked a deeper question: does this pattern hold true for all types of cancer, or is it unique to the liver? Before looking at the data, he registered a formal prediction with a public science registry, stating that he expected this downward shift to happen in at least ninety percent of different cancer types. He tested this prediction across ten different cancer datasets, ranging from lung and kidney to colon and ovarian cancers. The result was a refutation of his own hypothesis. The pattern did not hold everywhere. In half of the cancer types he tested, the direction of the shift was opposite to what he had expected. In some cancers, adding the cell division score made the gene estimates go up; in others, it made them go down. The direction of the change was not a fixed rule of biology but was determined by how strongly cell division itself acted as a predictor of survival in that specific cancer. In cancers where rapid division was a strong sign of a poor outcome, the genes linked to division appeared less dangerous once the division speed was accounted for. In cancers where division did not strongly predict the outcome, the shift was minimal or reversed.
This discovery reveals that the bias introduced by ignoring tumor proliferation is not a simple error that always pushes results in one direction. Instead, it is a fluid distortion that depends entirely on the specific biology of the cancer being studied. The study also highlighted a practical problem in how science is reported. Because most papers do not state which factors they adjusted for, a reader cannot know if a reported gene is a true driver of the disease or merely a reflection of unmeasured cell division. The researcher suggests a simple fix: scientists should report the strength of the cell division signal in their own specific group of patients and show how their gene estimates change when that signal is included. This would not necessarily solve the problem of which genes are truly causal, but it would make the dependence visible, allowing doctors and other researchers to see the full context of the numbers they are reading. The work serves as a reminder that in the complex landscape of cancer genetics, the answer to a question about a single gene cannot be separated from the behavior of the tumor as a whole.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.