TMBocelot: an omnibus statistical control model optimizing the TMB thresholds with systematic measurement errors

Xin Lai, Shaoliang Wang, Xuanping Zhang, Xiaoyan Zhu, Yuqian Liu, Zhili Chang, Xiaonan Wang, Yang Shao, Jiayin Wang, Yixuan WangFrontiers in Immunology 2025 · DOI

Abstract

Tumor mutation burden (TMB), defined as the number of somatic mutations of tumor DNA, is a well-recognized immunotherapy biomarker endorsed by regulatory agencies and pivotal in stratifying patients for clinical decision-making. However, measurement errors can compromise the accuracy of TMB assessments and the reliability of clinical outcomes, introducing bias into statistical inferences and adversely affecting TMB thresholds through cumulative and magnified effects. Given the unavoidable errors with current technologies, it is essential to adopt modeling methods to determine the optimal TMB-positive threshold. Therefore, we proposed a universal framework, TMBocelot, which accounts for pairwise measurement errors in clinical data to stabilize the determination of hierarchical thresholds. TMBocelot utilizes a Bayesian approach based on the stationarity principle of Markov chains to implement an enhanced error control mechanism, utilizing moderately informative priors. Simulations and retrospective data from 438 patients reveal that TMBocelot outperforms conventional methods in terms of accuracy, consistency of parameter estimations, and threshold determination. TMBocelot enables precise and reliable delineation of TMB-positive thresholds, facilitating the implementation of immunotherapy. The source code for TMBocelot is publicly available at https://github.com/YixuanWang1120/TMBocelot.

Introduction

Tumor Mutation Burden (TMB)—the count of mutations per megabase of DNA—has emerged as a cornerstone biomarker for predicting patient response to immune checkpoint inhibitors (ICIs). While regulatory bodies like the FDA have integrated TMB thresholds into clinical guidelines, the practical application of these thresholds is often hampered by "noise."

Measurement errors in TMB are pervasive, stemming from variations in sequencing platforms, tumor purity, and bioinformatic pipelines. Furthermore, clinical response labels (e.g., "responder" vs. "non-responder") are also subject to misclassification due to the limitations of medical imaging and subjective clinical assessment. When these "pairwise errors" are ignored, statistical models treat observed data as the ground truth, leading to biased parameter estimates and suboptimal patient stratification. TMBocelot was developed to address these systematic errors through a robust Bayesian framework.

Methodology

TMBocelot is an omnibus statistical framework designed to joint-model binary tumor responses (via logistic regression) and continuous time-to-event data (via Cox proportional hazards). The framework's primary innovation is its suite of error-control modules that denoise clinical data before threshold determination.

The TMBocelot framework overview, detailing the error-correction logic, data acquisition, and simulation modeling.

As illustrated in the framework overview, TMBocelot employs three specialized modules:

  • Bayes-ECM (Error Correction Method): This module addresses TMB measurement noise. It utilizes Dirichlet Process and Gaussian Mixture Models (GMM) to estimate error distributions without requiring a separate "gold standard" validation set.
  • Bayes-MCM (Misclassification Correction Method): This module targets errors in clinical response labels, using informative Beta priors to account for the probability of misclassification.
  • Bayes-PECM (Pairwise Error Correction Method): The most comprehensive version of the model, PECM simultaneously corrects for both TMB noise and response misclassification.

The model employs Markov Chain Monte Carlo (MCMC) sampling for parameter estimation. The final optimal TMB threshold is identified using the "minimal p-value principle" integrated with the TMBcat tool.

Validation and Results

The researchers validated TMBocelot using both synthetic simulations and real-world clinical data. The simulation study involved 500 iterations with cohorts of 200–500 individuals, demonstrating that TMBocelot could accurately recover "true" parameters from artificially noisy datasets where traditional "Naive" methods failed.

For clinical validation, the team analyzed a retrospective cohort of 438 patients with non-small cell lung cancer (NSCLC) across four major studies (Miao et al., 2018; Hellmann et al., 2018; Rizvi et al., 2018; Fang et al., 2019). Key findings included:

  1. Superior Stratification: TMBocelot identified thresholds that created significantly wider gaps in survival and response rates between "High" and "Low" TMB groups. In one specific cohort, the Naive method yielded a non-significant response disparity (p = 0.118), whereas TMBocelot’s threshold achieved high statistical significance (p < 0.001).
  2. Increased Predictive Power: ROC analysis indicated that using the model’s posterior TMB estimates improved the Area Under the Curve (AUC) for predicting patient response from approximately 0.81 to 0.90.
  3. Robustness: By using GMMs, the model proved capable of handling asymmetric or non-normal error distributions, which are common in real-world clinical settings but often break simpler statistical models.

Discussion and Conclusion

TMBocelot demonstrates that statistical denoising is as critical as the physical sequencing process in precision oncology. By accounting for the inherent noise in both biomarkers and clinical outcomes, the model provides a more reliable foundation for immunotherapy decision-making.

However, the model's complexity introduces certain trade-offs. The Bayesian MCMC approach requires more computational power than standard regressions, and the model's performance is optimized when "moderately informative priors" (existing clinical knowledge about error rates) are available. While currently validated in NSCLC, future work will focus on extending this framework to other cancer types and refining the computational efficiency for broader clinical adoption.

Summary updated

Abstract

Tumor mutation burden (TMB), defined as the number of somatic mutations of tumor DNA, is a well-recognized immunotherapy biomarker endorsed by regulatory agencies and pivotal in stratifying patients for clinical decision-making. However, measurement errors can compromise the accuracy of TMB assessments and the reliability of clinical outcomes, introducing bias into statistical inferences and adversely affecting TMB thresholds through cumulative and magnified effects. Given the unavoidable errors with current technologies, it is essential to adopt modeling methods to determine the optimal TMB-positive threshold. Therefore, we proposed a universal framework, TMBocelot, which accounts for pairwise measurement errors in clinical data to stabilize the determination of hierarchical thresholds. TMBocelot utilizes a Bayesian approach based on the stationarity principle of Markov chains to implement an enhanced error control mechanism, utilizing moderately informative priors. Simulations and retrospective data from 438 patients reveal that TMBocelot outperforms conventional methods in terms of accuracy, consistency of parameter estimations, and threshold determination. TMBocelot enables precise and reliable delineation of TMB-positive thresholds, facilitating the implementation of immunotherapy. The source code for TMBocelot is publicly available at https://github.com/YixuanWang1120/TMBocelot.
BibTeX
@inproceedings{lai2025tmbocelot,
  author = {Xin Lai and Shaoliang Wang and Xuanping Zhang and Xiaoyan Zhu and Yuqian Liu and Zhili Chang and Xiaonan Wang and Yang Shao and Jiayin Wang and Yixuan Wang},
  title = {{TMBocelot: an omnibus statistical control model optimizing the TMB thresholds with systematic measurement errors}},
  booktitle = {Frontiers in Immunology},
  year = {2025},
  doi = {10.3389/fimmu.2024.1514295},
}
Figure