Authors response to “Letter to the Editor”: “Tumor mutational burden and survival on immune checkpoint inhibition in >8000 patients across 24 cancer types”

In their Letter to the Editor, Bu et al¹ outline potential limitations and future directions for research of our recently published article.² We feel that reasoned dialog about the interpretation of results is essential for the healthy practice of science, and we want to ensure that such discussions are grounded in accurate information. Inaccuracies in the Letter regarding published research merit highlighting and discussion.

Lack of harmonization in TMB measurement across platforms

Tumor mutational burden (TMB) is a complex biomarker signature that varies from platform to platform.³ As of the time of this writing, there is only one assay (FoundationOne CDx) that has passed the analytical rigor of the US Food and Drug Administration (FDA) review process to support a companion diagnostic.⁴ The authors state:

The study relies on a single TMB assay (FoundationOne CDx), which limits the generalizability of the findings. While the FDA-approved assay ensures analytical rigor, the broader clinical adoption of TMB is hindered by inter-laboratory variability in panel size, bioinformatic pipelines, and germline variant filtering. For example, Nassar et al demonstrated that TMB algorithms relying on public germline databases (eg, gnomAD) underperform in non-European populations due to ancestral bias.² Although the authors note that FoundationOne [sic] uses a proprietary ancestry-balanced database, they do not provide comparative data on TMB performance across diverse genetic backgrounds.

These analyses had already been published and referenced in the discussion section of Gandara et al, reference 31, and do not alter the conclusions of the current paper:

Corroborating these findings, a recent study by Nassar et al reported that a TMB algorithm with database-derived algorithmic germline filtering had reduced predictive ability in patients of non-European genomic ancestries, likely due to under-representation of these ancestries in publicly available databases.³² In contrast, the FDA-approved TMB algorithm used in our study makes use of a genomic database for germline filtering that is much more representative of non-European ancestries compared with publicly available databases¹⁹ used for germline filtering by TMB assays evaluated by Nassar et al and has been found by Huang, Graf, and Oxnard to have similar stratification of outcomes on ICI in both European and non-European ancestry.³¹

The referenced study by Huang et al⁵ evaluated TMB stratification of outcomes on immune checkpoint inhibitor (ICI) for patients with non-small cell lung cancer of predominant European, African, and Asian ancestry, and observed better outcomes on ICI in TMB≥10 populations for all three, in contrast to the Nassar et al study,⁶ which compared TMB algorithms that did not support a companion diagnostic. While we agree with Bu et al that evaluating potential ancestral performance bias is a valuable component of biomarker development, omitting works that have addressed this topic is not conducive to healthy scientific discourse.

Retrospective design and immortal time bias

As non-interventional real-world data becomes more commonly used for filling clinical practice knowledge gaps and in drug development, awareness of the strengths and limitations of these types of data are essential. Particularly germane to any analysis of databases combining clinical and genomic data is immortal time. Bu et al brought up this topic in their Letter:

Despite risk-set adjustment for delayed cohort entry, residual immortal time bias may inflate survival estimates. Patients entering the database post-comprehensive genomic profiling report (median follow-up: 31.7 months) likely represent a survivor cohort with slower disease progression.

We acknowledge the relevance of immortal time bias. However, it was for this reason that our analyses specifically relied on relative estimates of outcomes, such as HRs, where the risk of an event in one group is evaluated relative to another group, which is less prone to potential immortal time bias. We also made considerate use of other best practices to account for immortal time, detailed in the Methods of Gandara et al, including risk set adjustment.⁷

Unexplored mechanisms behind microsatellite stable colorectal cancer exception

Bu et al state:

The MSS CRC subgroup uniquely fails to show a survival benefit for TMB≥10 (HR 1.02; 95% CI: 0.72 to 1.44). This anomaly contradicts the pan-tumor trend and suggests CRC-specific resistance mechanisms. The study does not explore potential confounders, such as *POLE/POLD1* mutations (significantly associated with ultra-mutated MSS CRC) or immunosuppressive features of the CRC tumor microenvironment (eg, Transforming Growth Factor-β (TGF-β) dominance4). A granular analysis stratifying MSS CRC by mutational signatures (eg, UV exposure) or immune cell composition (eg, T-reg infiltration) could clarify this discrepancy. Single-cell RNA sequencing or spatial profiling would be critical to dissect localized immune evasion in CRC.

Our study did show a survival benefit for high TMB in colorectal cancer (CRC), but we point out that it is associated with microsatellite instability (MSI) in this cancer subtype. Furthermore, the results of our study are consistent with prior evaluations of microsatellite stable (MSS) CRC^{8 9} observing minimal to no enrichment in outcomes among the MSS-only TMB≥10 group. In exploratory analyses provided in the Supplement, we evaluated stratification of overall survival (OS) by higher TMB levels in the MSS-only subgroup and observed a signal suggesting that TMB may have potential validity here. Nevertheless, given the exploratory nature of this analysis, we view this result as broadly supporting the validity of TMB with respect to biological gradient analyses. We agree with Bu et al that additional mechanistic evaluations of this subgroup might help shed light on important learnings about CRC and immunotherapy, but these types of analyses would require a considerable shift in the scope of our work.

Threshold ambiguity in ICI-chemotherapy combinations

It is not uncommon for retrospective analyses of outcome on ICI by TMB level in the literature to combine those receiving ICI and ICI+chemotherapy into one group⁸ or not specifying whether patients received ICI with chemotherapy.⁹ Our statistical analysis plan prespecified analyses of outcomes associated with anti-programmed death-ligand 1 (PD(L)1) monotherapy use, with an exploratory analysis of a separate cohort of patients with anti-PD(L)1 plus chemotherapy use.² Bu et al commented:

The exploratory analysis of ICI-chemotherapy combinations (n=4369) identifies TMB≥20 as the only predictive threshold (HR 0.65; p<0.001). However, this finding lacks biological rationale or clinical validation. The authors hypothesize that chemotherapy may dilute TMB-driven immunogenicity but omit mechanistic data (eg, neoantigen clonality, human leukocyte antigen (HLA) diversity) to support this.

To our knowledge, this is the first time that stratification of TMB levels has been assessed separately for ICI or ICI+chemotherapy cohorts in a pan-tumor setting. While we choose not to speculate regarding the biological rationale to support the hypothesis that chemotherapy may dilute TMB-driven immunogenicity, we also did not present this hypothesis in the article. In fact, nowhere in the published article did we make any biological hypotheses for the observed association.