Characterization of non-monotonic relationships between tumor mutational burden and clinical outcomes

Jordan Anaya, Julia Kung, Alexander S. Baras

Abstract

Potential clinical biomarkers are often assessed with Cox regressions or their ability to differentiate two groups of patients based on a single cutoff. However, both of these approaches assume a monotonic relationship between the potential biomarker and survival. Tumor mutational burden (TMB) is currently being studied as a predictive biomarker for immunotherapy, and a single cutoff is often used to divide patients. In this study we introduce a two-cutoff approach that allows splitting of patients when a non-monotonic relationship is present, and explore the use of neural networks to model more complex relationships of TMB to outcome data. Using real-world data we find that while in most cases the true relationship between TMB and survival appears monotonic, that is not always the case and researchers should be made aware of this possibility.

Introduction

In the field of oncology, Tumor Mutational Burden (TMB)—the total number of mutations per megabase of DNA—has emerged as a critical biomarker, particularly for predicting patient response to immune checkpoint inhibitors. However, the standard statistical frameworks used to evaluate TMB, such as Cox proportional hazards models or single-cutoff patient stratification (e.g., TMB-High vs. TMB-Low), rely on the fundamental assumption of monotonicity. This assumption posits that as TMB increases, the risk of death (hazard) must either strictly increase or strictly decrease.

Biological systems, however, often exhibit non-monotonicity. If a "Goldilocks zone" exists where moderate TMB levels confer a different risk profile than extreme values, traditional linear models will fail to capture this complexity, leading to suboptimal patient stratification and potentially flawed clinical conclusions.

Methodology

To address these limitations, this study explores two alternative modeling strategies designed to detect and accommodate non-monotonic relationships:

  1. Two-Cutoff Stratification: Unlike the standard single-cutoff approach, this method searches for two optimal values to group patients. This allows for the grouping of extreme TMB values (very low and very high) against a middle-range group, effectively capturing U-shaped or inverted U-shaped risk profiles.
  2. Fully Connected Neural Networks (FCN): The researchers implemented a deep learning model using TensorFlow to estimate partial hazards. Unlike the linear Cox model, the FCN can learn complex, non-linear transformations of the input data, allowing it to fit non-monotonic hazard functions.
Conceptual framework showing linear vs. non-monotonic hazard functions and the resulting stratification performance across different cutoff strategies.

The models were validated using simulated datasets with known "ground truth" relationships and then applied to large-scale real-world genomic and clinical datasets, including The Cancer Genome Atlas (TCGA), AACR Project GENIE BPC, and cohorts from Memorial Sloan Kettering (MSK).

Results and Discussion

Validation on Simulated Data

In simulations where the true relationship between TMB and hazard was non-monotonic, the FCN model demonstrated superior performance, accurately recovering the underlying curve. In contrast, the standard Cox model attempted to fit a straight line to the data, resulting in significant loss of predictive accuracy (C-index) and poor log-likelihood scores.

Comparison of model fits (True vs. Cox vs. FCN) on simulated linear (A, B) and non-monotonic (C, D) data, demonstrating the FCN's ability to capture complex hazard shapes.

Real-World Evidence: The Case of Melanoma

While TMB-survival relationships were found to be monotonic in the majority of cancer types studied, Skin Cutaneous Melanoma (SKCM) in the TCGA dataset served as a notable exception. The FCN identified a strong non-monotonic relationship where patients at the extremes of the TMB distribution had different risk profiles than those in the middle. Stratifying patients based on FCN-predicted risk yielded a much higher statistical significance (p=4.9E-05) compared to the standard median TMB split (p=3.7E-02).

Implications for Clinical Research

The study suggests that neural networks can serve as a valuable diagnostic tool. A significant performance gap between an FCN and a Cox model (measured by C-index or log-likelihood) can act as a "red flag," indicating that the underlying biological relationship is likely non-monotonic.

Conclusion

While monotonicity is the norm for TMB in most cancers, it is not a universal law. Researchers should be cautious of the "monotonicity trap" and consider more flexible modeling approaches, such as neural networks or multi-cutoff strategies, when evaluating biomarkers. Future work will focus on incorporating additional clinical variables and validating these flexible models in prospective clinical trial settings to improve personalized immunotherapy selection.

Summary updated

Abstract

ABSTRACT Potential clinical biomarkers are often assessed with Cox regressions or their ability to differentiate two groups of patients based on a single cutoff. However, both of these approaches assume a monotonic relationship between the potential biomarker and survival. Tumor mutational burden (TMB) is currently being studied as a predictive biomarker for immunotherapy, and a single cutoff is often used to divide patients. In this study we introduce a two-cutoff approach that allows splitting of patients when a non-monotonic relationship is present, and explore the use of neural networks to model more complex relationships of TMB to outcome data. Using real-world data we find that while in most cases the true relationship between TMB and survival appears monotonic, that is not always the case and researchers should be made aware of this possibility. Significance When a non-monotonic relationship to survival is present it is not possible to divide patients by a single value of a predictor. Neural networks allow for complex transformations and can be used to correctly split patients when a non-monotonic relationship is present.
BibTeX
@inproceedings{anaya2024characterization,
  author = {Jordan Anaya and Julia Kung and Alexander S. Baras},
  title = {{Characterization of non-monotonic relationships between tumor mutational burden and clinical outcomes}},
  booktitle = {Cancer},
  year = {2024},
  doi = {10.1158/2767-9764.crc-24-0061},
}
Figure