Immunotherapy has reshaped what is possible in non-small cell lung cancer (NSCLC), but it has also created a new clinical problem: determining who is benefiting, who remains at meaningful risk of recurrence, and who may be receiving treatment without disease left to treat.
Imaging remains central to these decisions, but it is an imperfect window into microscopic disease and early treatment response. Circulating tumor DNA (ctDNA) and molecular residual disease (MRD) testing offer a different view one based on tumor-derived DNA released into the bloodstream.
The attraction is obvious. A molecular signal that falls during treatment could provide an earlier indication of response. Persistent or re-emerging ctDNA after curative-intent therapy could identify patients at particularly high risk of recurrence. And, eventually, a sufficiently reliable negative result could help identify patients who may not need additional therapy.
But in lung cancer, the biology and the technology have not yet made those decisions straightforward.
The promise and the limitation of the signal
Detecting ctDNA is fundamentally a signal-to-noise problem. Tumor-derived DNA represents only a small fraction of circulating cell-free DNA, and distinguishing genuine tumor variants from background signals requires increasingly sensitive and specific approaches.
Tumor-informed MRD assays address part of this problem by tracking patient-specific variants identified from tumor tissue. Newer platforms are pushing analytical sensitivity still further. Yet assay sensitivity alone cannot solve the biological problem: some lung cancers shed very little DNA into the circulation, and certain sites of disease, particularly isolated CNS recurrence, may remain difficult to detect through plasma.
A negative result therefore cannot currently be interpreted as proof that residual disease is absent.
You can also read Lung Cancer Immunotherapy: Real-World Survival and Costs in France on OncoDaily.
MRD is prognostic before it is predictive
This distinction is central to where the field stands today.
In NSCLC, postoperative ctDNA positivity is strongly associated with recurrence risk, and molecular response during systemic therapy has repeatedly correlated with clinical outcomes. What has not yet been established is whether changing treatment because of that signal consistently improves survival.
That makes MRD primarily a prognostic and monitoring biomarker, rather than a treatment-directing biomarker, in current lung cancer practice.
The difference matters. Knowing that a patient is at high risk is clinically valuable, but its value becomes much greater when there is an evidence-based intervention available because of that information.
Why lung cancer has been particularly difficult
Early-stage NSCLC creates an especially challenging setting for MRD-guided treatment.
After surgery, many patients are already cured, while others retain microscopic disease that conventional staging cannot identify. Adjuvant therapy therefore treats a biologically mixed population: some patients have residual cancer that may benefit from treatment, while others are exposed to therapy despite having no disease remaining.
MRD offers a potential way to separate those groups more precisely.
But lung cancer is also a relatively low-shedding tumor, making false-negative results an important concern. Tissue availability can further complicate tumor-informed testing because diagnostic lung biopsies are often small and already required for histology and molecular profiling.
These limitations help explain why ctDNA-guided strategies that appear conceptually straightforward can be much harder to execute in NSCLC than in some other malignancies.
The field is moving from detection toward action
The broader oncology landscape shows where this could lead.
Across several tumor types, ctDNA is moving beyond recurrence prediction toward patient selection, surveillance, treatment escalation, and treatment de-escalation. Lung cancer has not reached that point yet, but the direction is increasingly clear.
The critical question is no longer simply whether increasingly sensitive assays can detect residual disease. It is whether clinicians will have sufficiently effective and differentiated treatment options to act on what those assays reveal.
That distinction may define the next phase of MRD in immuno-oncology.
The technology for seeing residual disease is improving rapidly. The therapeutic options available after that signal appears are expanding as well, from new immunotherapy combinations and antibody-drug conjugates to increasingly precise targeted therapies.
The challenge now is connecting the two.
Experts Perspective: A Thoracic Oncologist’s View
To better understand the current role of ctDNA and MRD in NSCLC from the promise of earlier detection to the challenges of using molecular signals to guide treatment, OncoDaily IO spoke with Chris Warfield, a thoracic oncology specialist focused on lung cancer biology, detection, and clinical decision-making. He works at Natera.
Immunotherapy reshaped non-small cell lung cancer. After treatment, what’s the hardest question an oncologist is actually left with, and is it really about detecting disease?
It’s less about detection than people assume. A large share of patients don’t respond to our best drugs, and those drugs aren’t free to give: checkpoint inhibitors carry real, sometimes lasting toxicity, from endocrinopathies to pneumonitis to colitis. So the hard moment is deciding whether to keep exposing someone to a demanding therapy that may not be helping, or whether a patient in front of you may already be cured. For years the honest answer was that we couldn’t see well enough to know, and the instinct that we needed better sight was correct; seeing genuinely was the constraint. That has begun to change only recently. The signal is catching up, and we’re crossing into a different problem: less whether we can detect residual or responding disease than what we can do about it once we have. That crossing is new, the shape of the last year or two.
You’ve described ctDNA detection as a signal-to-noise problem before it’s a sequencing one. What do you mean, and why does the depth of the assay matter so much?
A handful of tumor-derived molecules circulate against a large background of normal cell-free DNA, much of it shed by blood cells, and clonal hematopoiesis alone accounts for a big share of the variants you find in plasma. Pulling a true tumor signal out of that background is the whole game, which is why the field has moved toward counting molecules against patient-specific, tumor-confirmed variants while filtering out blood cell noise.
Depth falls on a ladder: a broad liquid biopsy profile reads to roughly one part in ten thousand; tumor-naïve MRD reaches somewhat deeper, though how far depends heavily on the platform and technology utilized; tumor-informed MRD reaches toward one part per million; and the newest whole-genome, tumor-informed assays operate at low single digit parts per million.
But depth is a balance, not a straight line. Tracking many confirmed variants lets you call a lower signal, yet each added target carries its own error. So past a point, data seems to show that reaching for more depth can cost specificity. Lower isn’t automatically better; the gain comes from the right confirmed targets, not simply more of them. And analytical sensitivity is necessary, not sufficient: a negative still depends on whether a patient’s disease was shedding into plasma the moment blood was drawn.
Where does the technology still fall short in lung specifically?
Three honest places. First, the tumor-informed approach needs adequate, quality tissue to build the personalized assay, and that build takes time; in lung, tissue is often the binding constraint, because biopsies are small and what little there is has usually been spent on diagnosis and biomarker testing.
Second, lung sheds less than many tumors, which sets a biological floor: at the single postsurgical landmark, sensitivity clusters around forty to sixty percent, so a negative there is nearly as likely to be a missed low shedder as a true all clear. It climbs into the three-quarters range, and higher with the most sensitive whole-genome assays, when patients are followed longitudinally rather than tested once. Third, some disease is effectively invisible to blood: brain-only relapse, sequestered behind the blood-brain barrier, is poorly detected in plasma, particularly with isolated CNS disease, a gap none of the available assays have closed, and one more reason imaging keeps its place. On the tissue problem, I’d be careful reading it as an argument for tissue-free testing. Where tissue and turnaround allow, tumor-informed assays currently carry the strongest evidence base for postoperative MRD in lung, so it’s worth improving how we acquire tissue rather than routing around it. But it’s context dependent, not a hierarchy: tumor-naïve, methylation-based approaches matter where tissue isn’t available or timing won’t wait, and each has trade-offs worth naming.
That’s why some of the most useful recent progress isn’t in the sequencing lab at all: cryobiopsy outperforms conventional bronchoscopic sampling on both diagnostic yield and the cellular quality molecular testing needs, and disposable, self-contained probes are lowering the barrier for centers without dedicated capital equipment.
There’s a persistent tension between MRD as a prognostic marker and MRD as something that should change treatment. How should the signal be read today?
As a monitoring input, not a mandate. In lung, MRD today is a prognostic marker, a strong one, not a predictive one, and that’s not a failure; it’s the ordinary adolescence of some biomarkers. HER2 and the multigene recurrence scores in early breast cancer both arrived as prognostic tools before they earned the right to direct therapy, and PD-L1 entered as an enrichment marker for immunotherapy. Resistance at the prognostic stage is normal and will hopefully prove to be temporary in this case. Read that way, the question an MRD result answers, at this point, is “how closely, and when, am I watching,” not “what drug does this change.” And a negative is absence of evidence, not evidence of absence: it runs alongside imaging, never instead of it.
Where is ctDNA already earning its place across the immunotherapy course?
In three settings, all observational. Before surgery, pathologic complete response remains the established surrogate, but ctDNA clearance during chemoimmunotherapy is already associated with better survival, a complementary molecular read on whether the therapy is working. That signal rests on broad ctDNA assays with real limits, though; in the pivotal neoadjuvant trial only about a quarter of patients were even evaluable for clearance, which is exactly the kind of readout more sensitive, tumor-informed MRD could sharpen.
On treatment, molecular response behaves as a dynamic biomarker: an early drop or clearance tracks with durable outcomes, and a rising signal can precede the scan and point toward resistance. And after curative-intent therapy, MRD stratifies who is at genuine risk of recurrence. But the value of all three depends on something outside the assay.
What’s the “something outside the assay”?
The menu. An early, accurate signal is only as useful as the move it unlocks. Where a differentiated next step exists, a targetable resistance mechanism, an evidence-backed escalation, a trial, or the still-investigational confidence to de-escalate and spare toxicity in a deep responder, the readout changes care. Where a patient is a non-responder with nothing better to switch to, an early signal too often just tells you sooner what you can’t yet fix. Naming that asymmetry honestly is the point.The one exception worth holding onto: identifying a durable responder has standalone value even without a new drug, because sparing an unnecessary, toxic therapy is itself a decision worth making. For a while the assumption has been that the assay is the ceiling. Right now, it isn’t. The menu is.
OncoDaily IO covers the whole immuno-oncology landscape. What can lung learn from tumors that are further ahead?
Quite a lot, because the pattern repeats. In muscle-invasive bladder cancer, ctDNA became the first test to select patients for adjuvant checkpoint immunotherapy by detecting residual disease itself, reviving a drug that had failed in an unselected trial and delivering both disease-free and overall survival benefit, now reflected in a category 1 recommendation. In stage II colon cancer, a ctDNA-guided approach has safely reduced adjuvant chemotherapy without compromising recurrence-free survival.
In Merkel cell carcinoma, another immunotherapy-driven disease, ctDNA surveillance carries a positive guideline recommendation. In diffuse large B-cell lymphoma, molecular response is woven into how response itself is judged, now with the first inclusion of ctDNA in lymphoma guidelines, and it’s used around cellular and bispecific immunotherapies. And in HPV-driven head and neck cancer, viral ctDNA has become a widely used surveillance tool, flagging recurrence months before imaging, though its power to change outcomes, as opposed to detect them, is still being established.
The through-line: where the therapeutic menu is deep enough to offer a differentiated move, monitoring has already crossed from prognostic to decision guiding. That’s what lung is reaching toward, and, tellingly, guidelines don’t yet recommend ctDNA MRD in NSCLC at all.
Why has lung been harder, and why have the adjuvant immunotherapy trials become such a talking point?
The setting is everything, and precision matters here. Given before surgery, neoadjuvant or perioperatively, immunotherapy has delivered, with overall survival benefits now on the board. The disappointment sits in the adjuvant-only setting: there the benefit has been inconsistent across trials and has not tracked cleanly with PD-L1 the way it does in metastatic disease. One trial’s gains sat mostly in PD-L1-high patients; another showed a disease-free benefit overall but, unexpectedly, not in its PD-L1-high subgroup; a third missed its endpoint outright. Much of this is still disease-free survival, not mature overall survival.
The likeliest reason, increasingly argued in the literature, is the population. Adjuvant-only trials select on stage alone, casting a net over a group in which many are already cured by surgery; at the earlier stages, well over half never recur. Treating everyone dilutes any benefit and exposes cured patients to real, sometimes lasting toxicity for nothing. You cannot show a benefit in patients who have no disease to treat.
Tellingly, the settings that have worked best are the ones where response can actually be read, pathologic response, ctDNA clearance during neoadjuvant therapy, while the blind, stage selected adjuvant setting is where the signal thins. Lung has tried the bladder playbook: the MERMAID-1 and MERMAID-2 trials selected adjuvant immunotherapy by MRD, and both were discontinued in 2023, with the sponsor citing shifting priorities and a changing early-stage landscape rather than a futility readout. The deeper problem holds regardless, though: when lung sheds so little, MRD-positive patients are scarce, and an MRD-selected trial is hard to fill in the first place.And there’s a wrinkle worth naming: unlike bladder, lung’s MRD-negative patients still appeared to benefit from adjuvant immunotherapy. The likeliest explanation isn’t that selection is wrong but that the negatives weren’t truly negative: a meaningful share of “undetectable” patients went on to relapse, which is what you’d expect when the assay’s floor sits above the disease. The growing view is that deeper, next generation sensitivity should narrow that gap. It’s a hypothesis that is now beginning to be tested, bounded by the reality that some disease sheds too little for any assay to see. You can’t treat disease that isn’t there, and you can’t spare a patient until you’re sure it isn’t.
So where does that leave us, and where’s the optimism?
Both halves of this are young, and if anything the seeing has only edged slightly ahead of the doing in the last year or so. That small, recent lead explains the hesitation you see among thoracic oncologists asked to build tumor-informed monitoring into practice: the data increasingly show a benefit, but a signal you can’t always act on is a hard thing to sit with in a patient’s chart. That reluctance isn’t a failure to keep up; it’s a rational response to a gap that only recently opened.
What makes this the hopeful moment is that the gap is closing from both directions. The options are widening beyond more of the same. New classes with genuinely different mechanisms are arriving: PD-1/VEGF bispecifics, antibody drug conjugates, and next generation targeted agents. A rising signal needs somewhere to point, and these give it real options, even if much of this still has to prove it extends survival. And the same tool that reads the signal is helping build that menu faster: regulators now recognize ctDNA for patient selection and enrichment, which lets a new drug prove itself in fewer patients and less time. The first ctDNA-selected immunotherapy approval grew directly out of it.
Neither curve is finished, but they’re climbing together now, and pulling each other up. For a brief moment, we can see a little further than we can act. That’s not a gap to wait out; it’s the reason to lean in. I believe the same instrument that shows us residual disease is the one that will help build the menu to meet it.