ResponseVerdict overturns computationally validated immunotherapy signatures, recovers their missed responders and outperforms every published response predictor

Cancer immunotherapy has revolutionized how doctors treat disease, offering a way to harness the body's own immune system to hunt down tumors. For a minority of patients, these treatments produce long-lasting remissions, but for many others, the drugs simply do not work. The central challenge for oncologists is knowing in advance which patients will benefit and which will not, sparing the latter from unnecessary side effects and costs while ensuring the former receive the life-saving care they need. To solve this, scientists have spent years trying to build "signatures"—specific lists of genes in a tumor that act as a warning light, flashing green for a likely success or red for a failure. The hope has been that by reading these genetic codes, doctors could predict the outcome before a single dose of medicine is given.

However, a new analysis suggests that the map many researchers have been following is fundamentally flawed. A team of scientists has examined thousands of these genetic signatures proposed in recent years and found that, when tested against real patients who received treatment, they perform no better than a random guess. The study reveals that the standard methods used to validate these predictions were misleading, causing researchers to mistake a patient's general health outlook for their specific reaction to the drug. By replacing these broken tools with a new method that relies strictly on measured outcomes from other patients, the researchers have created a system that consistently identifies more people who will actually respond to immunotherapy.

The problem began with how these genetic signatures were built. Most were developed using data from large databases of tumor samples that contained information about how long patients lived, but no record of whether they had ever received immunotherapy. Because the researchers did not have treatment data, they fitted their models to predict survival. In many cancers, a high level of immune cells in a tumor is actually a sign of a worse prognosis if the patient is not treated, because the immune system is struggling to control the disease. When these models were later repurposed to predict immunotherapy response, the direction was flipped: the high immune levels that signaled a poor future were reinterpreted as a sign that the immune system was ready to fight the drug. This created a paradox where the very features that suggested a patient would do poorly without treatment were used to predict they would do well with it.

To verify these predictions, the scientific community relied heavily on a computational tool called TIDE. This software was designed to estimate how well a tumor might respond to immunotherapy by looking at signs of immune exhaustion. The new study found that TIDE itself was misinterpreting the data. It treated the presence of exhausted immune cells—cells that are present but worn out—as a sign of failure, when in reality, these cells are often the ones that can be reawakened by the drug. Consequently, TIDE frequently labeled tumors that were actually rich in immune cells and likely to respond as "non-responders." Because so many new signatures were built to agree with TIDE, they inherited this error, leading to a cycle where incorrect predictions were validated by other incorrect tools.

The researchers tested this by gathering 1,566 different signatures from over 1,200 scientific papers and applying them to data from more than 2,000 patients who had actually received immunotherapy. The results were stark: the vast majority of these signatures had no ability to predict who would respond. Their accuracy was indistinguishable from flipping a coin. In fact, the study showed that if one were to pick a random set of genes, it would pass the same validation tests that these complex signatures passed, proving that the tests were not actually measuring response to treatment. The signatures were so poor that they performed worse than simple, older measures of immune activity that had been available for years.

In response to this failure, the team developed a new approach called ResponseVerdict. Instead of relying on computer simulations or agreement with other tools, this method looks directly at how patients in other groups actually responded to treatment. It uses those real-world outcomes to orient the prediction models, ensuring that the direction of the prediction matches the reality of the patient's experience. When tested against the old signatures and other established predictors, ResponseVerdict consistently identified more patients who would benefit from the therapy. In one specific test involving a prospective cohort of patients where the outcomes were not yet known, the new method correctly identified significantly more responders than the previous standard.

The impact of this correction extends beyond just better predictions; it changes how experiments and clinical trials are designed. The study found that because of the flawed signatures, many laboratory experiments were being performed in the wrong types of cells, chasing signals that were not actually driving the immune response. Furthermore, a current clinical trial for gastric cancer uses one of these flawed signatures to decide which patients are eligible. The researchers showed that this rule was inadvertently excluding patients who would have responded well to the treatment. By using the new method, the trial could have included more of the right patients, potentially changing the trial's outcome.

The researchers also applied their method to decide how to allocate drugs in randomized clinical trials, a scenario where the goal is not just to find responders, but to give the experimental drug to the people who will gain the most from it compared to the standard care. In a major trial for kidney cancer, the old method of predicting drug sensitivity led to patients receiving the wrong treatment, costing them months of time without disease progression. When the researchers used their new allocation rule, they identified a group of patients who gained a significant amount of time without their cancer progressing, while the rest of the group did not benefit as much. This suggests that the new method can not only predict who will respond, but also optimize who gets which treatment to maximize the overall benefit for a group of patients.

The study concludes that the field has been operating on a foundation of computational agreement rather than measured reality. By shifting the focus back to actual patient outcomes and using data that was available at the time the original papers were written, the researchers demonstrated that a better path was always possible. The new system does not require new data or futuristic technology; it simply requires a different way of asking the question. Instead of asking if a model agrees with a computer program, it asks if the model agrees with the patients. This shift has already shown that it can recover dozens of additional responders for every hundred patients treated, a gain that could translate into meaningful improvements in survival and quality of life for people facing cancer.