The Immunotherapy Lottery

Imagine you walk into a casino. The house tells you roughly one in three players walk out with their lives. The rest go home empty. Now imagine this happens in real life, except in an oncologist’s office, and the chips on the table are years. That, stripped of its clinical veneer, is the current reality of cancer immunotherapy. In theory, it’s an elegant premise. Instead of poisoning the body with traditional chemotherapy, these drugs “unmask” the disease, training the patient’s own killer T-cells to recognize and destroy the tumor. But in practice, this cellular battle rarely goes as planned. When clinicians talk about a successful response, they mean significant, measurable tumor shrinkage or long-term durable survival. Responses have been detected in only twenty to forty percent of patients across most cancer types⁽¹⁾, and in some settings, the effective response rate sits even lower ⁽²⁾. We have spent the better part of a decade celebrating immunotherapy as a triumph in modern medicine, and it mostly is. But a treatment that fails for 60 percent of those who receive it is, by any standard, a coin flip in a lab coat.

The core issue is prediction. Specifically, we are not good at it. PD-L1, a protein that cancer cells use to hide from the immune system, has historically been the default marker doctors use to decide who gets immunotherapy. It works, kind of, sometimes, more in some tumors than others. The problem is that PD-L1 is highly dynamic and wildly inconsistent. A single needle biopsy captures only a frozen snapshot of a tumor; it can vary dramatically between the primary site and a metastasis, or change entirely after a round of radiation. More fundamentally, looking for PD-L1 expression only tells you if the tumor is actively trying to suppress an immune attack. If the tumor is “immunologically cold” (meaning killer T-cells never managed to infiltrate the tissue in the first place), blocking PD-L1 does absolutely nothing. You are releasing the brake on an engine that isn’t even running. Tumor mutational burden (TMB), a measure of how many genetic errors have accumulated in a tumor, got FDA approval as a selection tool for one major immunotherapy drug back in 2020. But the more carefully researchers look, the clearer it becomes that neither marker is sufficient on its own, relying on one instead of the other simply trades one blind spot for another⁽²⁾. TMB indicates how genetically unstable a tumor is. The theory is intuitive: the more mutations a tumor has, the weirder its cells will look to the body, making it easier for T-cells to spot them as foreign threats. TMB does not tell you whether the immune system in and around that tumor is actually capable of mounting a response, which turns out to be important. If the local tumor microenvironment is fiercely immunosuppressive, teeming with cells designed to shut down inflammation, even the most conspicuous, highly mutated tumor will stay unbothered.

Fumet and colleagues put it clearly in a 2020 review: TMB has biological plausibility as an immunotherapy marker, but the clinical evidence is messier than the theory would predict ⁽³⁾. Tumors with high TMB do not reliably respond better. The cutoffs are arbitrary. The assays vary between labs. It is not that TMB is useless; it is that it is being asked to carry more predictive weight than the underlying biology actually supports.

Jardim and colleagues traced a lot of this back to a more fundamental problem: TMB was developed under a specific set of assumptions about how the immune system recognizes cancer cells, and those assumptions hold true in some contexts and not at all in others ⁽⁴⁾. Prior treatment history, tumor type, and other biological factors all interact with TMB in ways that make a universal cutoff essentially meaningless. Wang and colleagues extended this analysis in 2024, showing the prediction problem is even worse when applied across different cancer types ⁽⁸⁾. They demonstrated that a “one-size-fits-all” TMB threshold is effectively a statistical mirage. Because every organ has a different biological baseline for mutation, driven by cell turnover rates, exposure to environmental carcinogens, and intrinsic DNA repair efficiencies, a TMB value considered “high” in lung cancer might be indistinguishable from a “low” score in a sarcoma. By forcing a universal number onto radically different biological landscapes, we aren’t just missing the mark but instead trying to measure temperature with a ruler. We keep reaching for simple molecular proxies for a biological problem that is not simple.

The Real Culprits Are Harder to Measure

What both PD-L1 and TMB miss is the actual immune landscape inside the tumor. Not just whether the immune system has something to recognize, but whether the immune cells that are supposed to do the killing are present, functional, and not actively being shut down. That shutting down is the part we have been underweighting, and it is carried out by specific cell populations that standard tests do not even look at.

Regulatory T cells are the most well-characterized of these cell populations. Think of them as the immune system’s peacekeepers, cells whose normal job is to prevent the immune system from attacking healthy tissue. In a tumor, they do that same job at exactly the wrong time, suppressing the killer cells that immunotherapy is trying to activate. Togashi and colleagues found in 2019 that high levels of regulatory T cells inside a tumor correlate with worse outcomes across multiple cancer types ⁽⁵⁾.

Tumor-associated macrophages add another layer. Macrophages are immune cells that patrol the body and flag threats. Inside a tumor, a subset of them gets reprogrammed into something closer to the tumor’s bodyguards, secreting chemicals that suppress immune activity and actively remodel the environment around the tumor to protect it. Mantovani and colleagues laid this out in a 2022 review ⁽⁶⁾. These cells are not passive bystanders. They are working against the treatment. Myeloid-derived suppressor cells, another population of immune cells that tumors recruit to block attacks against them, pile on top of that. Veglia and colleagues described in 2021 how these cells disable killer T cells through multiple overlapping mechanisms ⁽⁷⁾. They don’t just hit a single off-switch; they wage a multi-front war. MDSCs mass-produce enzymes that starve T-cells of critical amino acids like L-arginine, rendering them too weak to divide. Simultaneously, they flood the local tissue with reactive oxygen species that chemically warp T-Cell receptors, effectively blinding the immune cells right as they arrive. To seal the perimeter, they secrete suppressive cytokines that recruit even more regulatory T-cells to the frontline. This creates a total systemic shutdown.

None of these populations show up in standard tumor tests. A tumor can have high mutational burden, strong expression of the PD-L1 protein, and still be sitting in an immune environment so thoroughly suppressed that no immunotherapy drug is going to matter. And we will not know that until after the patient does not respond.

The Tools Problem

Part of why we are still here, still debating the same markers a decade after they were proposed, is that mapping the immune environment inside a tumor at the resolution needed to actually understand it is hard. Single-cell RNA sequencing, a technology that can read the activity of tens of thousands of individual cells at once, is powerful, but standard analysis pipelines treat each cell as an isolated data point. They do not model how cells talk to each other. They do not account for where cells sit physically inside the tissue. And they have a well-documented tendency to lose rare cell populations entirely because those populations get statistically absorbed into larger groups during analysis.

The immunosuppressive cells we actually need to measure are often present at very low frequencies below the threshold where standard methods reliably detect them. This means we are building response prediction models on a picture of the tumor immune environment that systematically leaves out the cells most likely to explain why treatment failed. Newer approaches using graph-based deep learning, which model cells as a connected network rather than isolated points and can factor in where cells are physically located relative to each other, are starting to recover what conventional methods have missed. The clinical application of these approaches is still early, but the direction is the right one.

What Actually Needs to Change

The biomarker conversation in immunotherapy has been stuck in a loop. A new marker gets proposed, shows promise in early studies, gets approved or widely adopted, and then real-world use reveals it is not as predictive as hoped ⁽¹,³,⁴,⁸⁾. The cycle keeps repeating because the underlying problem has not changed: we are measuring features of the tumor itself when the relevant biology is in the immune cells surrounding it.

That is a harder measurement problem. The immune environment inside a tumor is dynamic, varies across different regions of the same tumor, and involves rare cell populations that current standard tests do not capture. Solving the prediction problem in immunotherapy is not going to come from a better version of TMB or a more sensitive PD-L1 test. It is going to require mapping the immune environment comprehensively enough to detect the suppressive networks that immunotherapy is trying to overcome, at a resolution that can actually guide clinical decisions.

The sixty percent of patients who do not respond are not failing because of bad luck. They are failing because of biology we currently do not have reliable ways to measure, suppressive networks operating below the threshold of every test we have standardized, in cell populations we have spent a decade not looking at. The biomarker we need already exists inside the tumor. We haven’t built the tools yet to find it.

References

1. Yu D, Yuan C, Zhang H, Chu W. The association of efficacy with PD-1/PD-L1 inhibition and tumor mutational burden in advanced nonsmall cell lung cancer: A PRISMA-guided literature review and meta-analysis. Medicine (Baltimore). 2022;101(29):e29676. doi:10.1097/MD.0000000000029676.
2. Dong A, Zhao Y, Li Z, Hu H. PD-L1 versus tumor mutation burden: Which is the better immunotherapy biomarker in advanced non-small cell lung cancer? J Gene Med. 2021;23(2):e3294. doi:10.1002/jgm.3294.
3. Fumet JD, Truntzer C, Yarchoan M, Ghiringhelli F. Tumour mutational burden as a biomarker for immunotherapy: Current data and emerging concepts. Eur J Cancer. 2020;131:40–50. doi:10.1016/j.ejca.2020.02.038.
4. Jardim DL, Goodman A, de Melo Gagliato D, Kurzrock R. The Challenges of Tumor Mutational Burden as an Immunotherapy Biomarker. Cancer Cell. 2021;39(2):154–173. doi:10.1016/j.ccell.2020.10.001.
5. Togashi Y, Shitara K, Nishikawa H. Regulatory T cells in cancer immunosuppression implications for anticancer therapy. Nat Rev Clin Oncol. 2019;16(6):356–371. doi:10.1038/s41571-019-0175-7.
6. Mantovani A, Allavena P, Marchesi F, Garlanda C. Macrophages as tools and targets in cancer therapy. Nat Rev Drug Discov. 2022;21(11):799–820. doi:10.1038/s41573-022-00520-5.
7. Veglia F, Sanseviero E, Gabrilovich DI. Myeloid-derived suppressor cells in the era of increasing myeloid cell diversity. Nat Rev Immunol. 2021;21(8):485–498. doi:10.1038/s41577-020-00490-y.
8. Wang X, Lamberti G, Di Federico A, et al. Tumor mutational burden for the prediction of PD-(L)1 blockade efficacy in cancer: challenges and opportunities. Ann Oncol. 2024;35(6):508–522. doi:10.1016/j.annonc.2024.03.007.