Key Ideas

PD-L1 remains the only biomarker in routine practical use for immunotherapy selection, but it's an imperfect, single-time-point measurement, and tumor mutational burden (TMB), despite genuine early promise, has essentially failed as a standalone predictive biomarker in NSCLC. The field's real goal should be a composite, predictive (not merely prognostic) biomarker that integrates PD-L1, TMB interpreted in proper context, ctDNA, and emerging data on tumor architecture and host immunity, and getting there will require validating these composite models in prospective trials.

Why We Need Biomarkers at All

Only a minority of patients meaningfully benefit from immunotherapy, and right now we largely rely on a single, binary metric (PD-L1) to try to select who that is. It's worth being precise about the distinction between a predictive biomarker (one that guides a treatment decision) and a prognostic one (one that simply estimates outcome, like cancer stage); a true predictive biomarker for immunotherapy response, in most contexts, remains an unmet goal, not something we've actually achieved yet.

PD-L1: Necessary, but Genuinely Limited

PD-L1 has real predictive value, and a 50% expression threshold is meaningfully used for immunotherapy monotherapy decisions, with reasonably well-worked-out assay standardization. But it's captured at a single point in time and space, and expression can be spatially and temporally unstable, meaning what gets sampled on a given biopsy may not represent the whole tumor. PD-L1 also appears less relevant in the context of chemo-immunotherapy specifically, and in squamous histology, though this remains genuinely debated even with substantial data already available. And PD-L1 shouldn't be interpreted in isolation from molecular profiling: driver mutations like EGFR and ALK identify patients who derive essentially no benefit from immunotherapy regardless of PD-L1 status, arguably the clearest, most reliable biomarker we currently have for excluding a treatment approach rather than selecting one.

Co-mutations like STK11, KEAP1, and SMARCA4 are less clearly actionable as biomarkers specifically for withholding immunotherapy: they're clearly associated with poor prognosis, but also with poor outcomes across essentially every treatment approach, not just immunotherapy, raising a real question of whether they should influence the immunotherapy decision at all, versus simply reflecting an overall harder-to-treat tumor regardless of what's offered.

TMB: A Real Idea That Hasn't Yet Worked in Practice

I'll disclose upfront that I'm a co-inventor on a patent related to TMB as an immunotherapy biomarker, which, if anything, should underscore that my skepticism about TMB isn't from a lack of investment in the concept. Early work (including some of my own, looking at hazard ratios across a range of possible TMB cutoffs) showed clearly that there's no single meaningful TMB threshold, it's a continuous variable, and any given cutoff's relevance appears to vary by cancer type rather than applying universally. TMB showed real promise in KEYNOTE-158 (enrichment for response in the TMB-high group), but when a TMB cutoff of 10 was formally tested as an integrated biomarker in CheckMate 227's TMB cohort, it ultimately did not show an overall survival benefit. Part of the problem is that TMB, as typically measured, is an imperfect, purely static snapshot: it only captures potential antigenicity (how many mutations exist that could theoretically be recognized), not whether those specific mutations are actually presented effectively by a given patient's HLA type, how clonal they are, or whether they successfully generate a real T-cell response. TMB from blood (rather than tumor tissue) carries all of these same limitations plus additional ones, including confounding from clonal hematopoiesis, an important caveat for anyone using blood-based TMB clinically.

What a Better Biomarker Would Need to Capture

A meaningful predictive biomarker for immunotherapy likely needs to integrate several additional layers beyond PD-L1 and TMB alone: mutational context (HLA genotype, clonality, and the quality of specific neoantigens, not simply mutation count), ctDNA (offering real-time tracking, potentially ahead of imaging, though requiring real caution in how it's validated and used), and emerging spatial and architectural data (tertiary lymphoid structures, B-cell responses, and the kind of tumor "neighborhood" analysis discussed earlier in this session).

There are now over 2,000 proposed biomarkers or models for immunotherapy response in the literature, most validated as single markers in individual datasets. Composite models show real promise: one example (the "I-score" model, developed with computational immunologist Diego Chowell, integrating TMB alongside copy number variation and HLA status) meaningfully improves prediction over any single metric alone. Even more strikingly, a separate model using nothing more than standard blood lab values (a basic CBC and metabolic panel) fed into a machine learning algorithm, validated across eight prospective clinical trials, was able to meaningfully predict response and survival with immunotherapy, a genuinely notable finding given how simple and widely available that underlying data already is.

Where This Needs to Go

The real challenge isn't a shortage of candidate biomarkers, it's figuring out how to combine them into a validated, prospectively tested composite model that can actually guide individual treatment decisions, rather than simply adding another single-marker study to an already crowded literature.

For Patients

Right now, doctors have only one biomarker (a protein called PD-L1) in routine, practical use to help decide who is likely to benefit from immunotherapy, and it's an imperfect one. A blood-based genetic test called tumor mutational burden (TMB) once seemed promising as an additional tool but has not held up well in lung cancer specifically when tested more rigorously. Researchers are increasingly trying to combine multiple types of information, genetic, blood-based, and even basic blood lab values, into more sophisticated computer models that may better predict who will actually benefit from immunotherapy, though this approach is still being developed and tested. Ask your care team what biomarker testing (beyond PD-L1 alone) has been done on your tumor, and whether that information changes your treatment options.

Key Takeaways

  • PD-L1 remains the only routinely used predictive biomarker for immunotherapy, but it's a single-time-point measurement with known spatial and temporal instability.

  • Driver mutations (EGFR, ALK) reliably identify patients unlikely to benefit from immunotherapy; co-mutations like STK11 and KEAP1 are more prognostic than clearly predictive, since they're associated with poor outcomes across most treatment approaches, not immunotherapy specifically.

  • TMB lacks a single meaningful cutoff (it's a continuous variable that varies by cancer type) and failed to show an overall survival benefit when formally tested as an integrated biomarker (CheckMate 227); blood-based TMB carries these same limitations plus confounding from clonal hematopoiesis.

  • Composite biomarker models, integrating TMB, copy number variation, HLA status, or even simple blood lab values into machine learning algorithms, show meaningfully better predictive performance than any single marker alone.

  • The central challenge for the field is prospectively validating these composite models so they can be used to genuinely guide individual treatment decisions, not just describe population-level trends.

References

  1. Samstein RM, et al. Tumor mutational burden predicts survival across multiple cancer types. Nat Genet. 2019.

  2. Borghaei H, et al. Nivolumab plus ipilimumab by tumor mutational burden in NSCLC (CheckMate 227).

  3. Chowell D, et al. Improved prediction of immune checkpoint blockade outcomes with a composite genomic biomarker (I-score).

  4. Machine learning prediction of immunotherapy outcomes using routine complete blood count and metabolic panel data across prospective trials.