Additive-input encoding fails operator-level target-held-out prediction on current Perturb-seq screens

Imagine a world where scientists can turn off individual genes in a cell and watch what happens, all at the same time. This is the promise of a technology called Perturb-seq. By using a molecular tool to silence specific genes and then reading the cell's entire genetic response, researchers hope to build a map of how genes talk to one another. The ultimate goal is to understand the hidden rules of life: if you pull one thread, which other threads tighten, and which ones go slack? For years, the standard way to interpret these massive experiments has been to assume the cell's response is a simple sum of parts. The idea is that if you turn off Gene A, it adds a specific amount of change; if you turn off Gene B, it adds another specific amount. If you turn off both, the result should just be those two changes added together. This "additive" view is attractive because it is mathematically clean and easy to calculate, allowing scientists to infer a master control map of the cell's internal wiring.

However, a new study from researchers at Brigham Young University suggests that this simple addition rule might be fundamentally broken when applied to real biological data. The team, led by Kyler Fullmer, Dalton Kutzen, and Tommy W. Terooatea, took three of the largest and most respected gene-silencing experiments ever performed and tested whether the additive model could actually predict what would happen to a gene the researchers had never seen before. They did not just look at the data; they built a rigorous testing ground to see if the model could generalize. They asked a straightforward question: if a computer learns the rules of the game from a set of known genes, can it correctly guess the outcome for a brand-new gene it has never encountered?

The answer, they found, was a resounding no. When the researchers tried to use the additive model to predict the behavior of unseen genes, the model failed completely. In fact, the model performed no better than a person who simply guessed that nothing would change at all. To ensure this failure wasn't just a fluke or a result of messy data, the team created a perfect simulation. They generated fake data where the rules were known to be perfectly additive, exactly as the theory claims. When they ran their model on this fake data, it worked beautifully, recovering the hidden rules with high accuracy. This proved that their testing method was sound and that the computer was capable of learning. The problem, therefore, was not with the math or the noise in the data, but with the assumption that real cells behave like simple sums.

The researchers tested this on three different datasets: two large screens involving human cells grown in a dish, and one that tested different doses of gene silencing. In every single case, the model that assumed genes simply add their effects together could not predict the outcome of a new gene. The error rate was so high that the model's predictions were indistinguishable from random guessing. The team ruled out several possible reasons for this failure. They checked if the problem was caused by picking the "easiest" genes to study, and found that even when they randomly selected genes, the model still failed. They checked if the problem was due to the number of genes being studied or the specific way the genes were measured, and found that the failure persisted regardless of these factors. Even when they stripped away the details of how strong the gene silencing was and looked only at the direction of the change, the model still could not predict the outcome.

There was one small exception that offered a glimmer of hope, but it was modest. On one specific set of cells, the model could predict the outcome of new genes slightly better than just guessing the average result. However, even in this best-case scenario, the model only recovered about seven percent of the information that a perfect, well-behaved system would have provided. It was a tiny victory in a sea of failure. The researchers also tested more sophisticated ways of encoding the gene information, including methods that looked at how genes are connected in the cell's natural state or used machine learning to learn the gene's "fingerprint" from the data. None of these advanced methods could overcome the fundamental problem. They failed to predict new genes on two of the three cell types tested, and on the third, they only managed a small improvement that fell far short of what the theory promised.

The study concludes that the current way of interpreting these massive gene-silencing screens is likely flawed. The assumption that genes act like independent switches that simply add up their effects does not hold up when tested against unseen targets. The researchers suggest two main possibilities for why this happens. First, the way they translate a gene's identity into a mathematical input might be throwing away most of the important information, like trying to describe a complex landscape by only looking at a single, blurry photo. Second, the biological reality might be that genes do not act in a simple, linear way; instead, their effects might saturate or hit a threshold, meaning that turning off a gene does not produce a steady, predictable change.

This work does not mean that gene-silencing screens are useless, but it does mean that the maps we have built from them so far may not be reliable guides for predicting what happens when we target new genes. The researchers have released a new toolkit to help other scientists check their own models against these strict standards, ensuring that future discoveries are built on solid ground. The lesson is clear: biology is often more complex than our simplest equations allow, and until we find a way to capture that complexity, our predictions of how cells will react to new interventions will remain uncertain.