Predictive Versus Prognostic Cancer Biomarkers Explained

Two biomarkers can both be called "positive" and mean completely different things for a treatment decision. A prognostic biomarker forecasts where a cancer is likely headed no matter which treatment is given. A predictive biomarker forecasts whether one specific drug will bend that path. The short version: prognostic markers describe the disease, predictive markers describe a drug's likely payoff, and only the second kind is a reason to pick one therapy over another. The words look interchangeable, but the evidence each one requires is not, and mixing them up is one of the most common ways a biomarker claim gets oversold.

Key points#

• A prognostic marker is linked to outcome regardless of treatment; a predictive marker is linked to benefit from a particular therapy.
• Proving a marker is prognostic needs an association with outcome. Proving it is predictive needs a significant treatment-by-biomarker interaction, ideally from a randomized trial.
• A single-arm study showing that marker-positive patients respond cannot, on its own, establish that a marker is predictive.
• PD-L1 and TMB are both used to help select immunotherapy patients, yet across most cancers they are only weakly correlated and behave as largely separate markers.
• "Prognostic" and "predictive" describe the evidence behind a marker, not a fixed property of the molecule.

One molecule, two very different questions#

The cleanest way to keep these terms apart is to notice that they answer different questions. A prognostic biomarker answers: how is this cancer likely to behave, whether we treat it or not? A predictive biomarker answers: will this particular drug help this particular cancer? Karla Ballman laid out these working definitions in the Journal of Clinical Oncology in 2015. In her framing, a prognostic marker informs about a likely outcome such as recurrence, progression, or death in the absence of therapy or with standard therapy, while a predictive marker is tied to response, or the lack of it, to a specific treatment.

The trouble is that one molecule can be prognostic, predictive, both, or neither, and which label fits depends entirely on the data behind it. A marker associated with shorter survival tells you the road ahead looks harder. It does not, by itself, tell you that any given drug will smooth it out. Study abstracts and promotional summaries often slide from the first claim to the second as if they were the same statement. The appraisal habit worth building is to pause and ask which of the two claims the evidence in front of you actually supports.

Why the interaction test carries the weight#

This is where a genuine predictive marker parts ways with a hopeful one. To call a marker prognostic, you need to show it tracks with outcome across treatments. To call it predictive, you need something more demanding: evidence that the size of the treatment effect itself changes with the marker. Statistically, that is a significant treatment-by-biomarker interaction, and the design that can measure it is a randomized trial in which marker-positive and marker-negative patients are each split between the drug and a comparator.

Ballman is explicit that this comparison is what earns the predictive label. Without a control arm, "marker-positive patients did well" cannot be separated from "marker-positive patients did well because of the drug." A single-arm study reporting that biomarker-high patients responded more often is equally consistent with a purely prognostic marker, a purely predictive one, or some blend of the two. Only the interaction, tested against a comparator, can tell them apart.

A quick sanity check is to picture what the marker does across arms. If it sorts patients into better and worse outcomes even when everyone gets the same treatment, that is prognostic information. If the gap between treated and untreated patients grows or shrinks depending on the marker, that is predictive information. A marker can carry both signals at the same time, which is exactly why the labels describe the evidence rather than the molecule.

Two worked examples from immunotherapy#

Immune checkpoint inhibitors give two instructive cases that show the distinction in practice.

PD-L1: a predictive marker that only shifts the odds#

PD-L1, measured by immunohistochemistry on tumor or immune cells, is widely used to help select patients for checkpoint inhibitors, and it is generally treated as predictive because higher expression tends to track with a greater chance of benefit. It is also imperfect in ways that matter at the bedside. Expression sits on a continuum, the cut points differ between assays and between drugs, and plenty of patients with low or absent PD-L1 respond while some with high expression do not. The marker moves the probability; it does not settle the outcome.

TMB: a cleaner look at the regulatory line#

Tumor mutational burden, the number of mutations per megabase of coding DNA, offers a sharper illustration of how the predictive-versus-prognostic question plays out in approvals. In 2020 the FDA granted accelerated, tissue-agnostic approval to pembrolizumab for adult and pediatric patients with unresectable or metastatic TMB-high solid tumors, defined as at least 10 mutations per megabase by an approved companion diagnostic. As summarized in Clinical Cancer Research, the approval rested on a prospectively planned analysis inside the single-arm KEYNOTE-158 trial, where the TMB-high subset showed an overall response rate near 29 percent.

Read that carefully. A response rate in a single-arm study documents that TMB-high tumors respond; the missing randomized comparator is precisely why the predictive-versus-prognostic debate around TMB stays open. Accelerated approval is a conditional pathway that asks for confirmatory evidence, not a final verdict. The 10-per-megabase threshold is also a pragmatic line drawn across a continuous variable, and its performance shifts across tumor types and sequencing platforms.

Why two markers can disagree#

Because PD-L1 and TMB are so often discussed together, it is tempting to assume they measure the same underlying thing. They do not. Yarchoan and colleagues reported in JCI Insight that across most cancers the two are only weakly correlated and act as largely independent markers, with a model combining them predicting response better than either alone. A separate analysis in Cancers found the relationship between PD-L1 and TMB to be inconsistent and tumor-type dependent. The practical takeaway: two predictive markers can point in opposite directions in the same patient, and neither is a promise.

A short checklist for any biomarker claim#

When a biomarker turns up in a headline, a report, or a sales deck, a handful of questions do most of the work:

• Is the claim about outcome, or about benefit from a specific drug?
• Did the evidence come from a single-arm study or a randomized comparison?
• Was a treatment-by-biomarker interaction actually tested and found significant, or was the marker merely associated with response among treated patients?
• Is the cutoff a natural boundary or a chosen point on a continuous scale, and does it hold up across assays and platforms?

These questions rarely feature in promotional summaries, which is a good reason to keep them handy. The deeper point is that "prognostic" and "predictive" are statements about the strength and design of the evidence, not fixed properties stamped onto a molecule. The same marker can be prognostic in one setting, predictive in another, and unproven in a third. Holding that line steady is a large part of reading oncology evidence honestly.