TMS Model Balance Determines the Prognostic Direction of Tumor Mutational Burden in Microsatellite-Stable Cancers

Background: Tumor mutational burden (TMB) is widely used as a prognostic biomarker, but its prognostic direction in microsatellite-stable (MSS) cancers remains contradictory. We hypothesized that the prognostic direction of TMB is determined by the balance of the trinucleotide mutation spectrum (TMS) model—the degree to which protective and risk mutational pathways offset each other. Methods: We trained TMS models in four independent MSS cohorts (SYSUCC-COADREAD as discovery, N=665, DFS; TCGA-COAD, N=437, OS; TCGA-STAD, N=337, OS; TCGA-UCEC, N=400, OS). For each cohort, we stratified patients by TMS tertiles and, within the extreme strata, by stratum-specific median TMB. We tested the TMS×TMB interaction using Cox models and meta-analyzed the interaction coefficients. We quantified TMS model balance using |cor(TMB, TMS)|. Results: In balanced TMS models (SYSUCC: |cor|=0.05; TCGA-STAD: |cor|=0.43; TCGA-COAD: |cor|=0.16), TMB was protective in TMS-Low patients and harmful in TMS-High patients (SYSUCC: HR=0.61 vs 1.47, interaction LRT P=0.0014; TCGA-STAD: HR=0.84 vs 1.28, interaction LRT P=0.047; TCGA-COAD: HR=0.89 vs 1.15, P=0.71, limited by power). In the unbalanced model (TCGA-UCEC: |cor|=0.93), the interaction was not statistically significant (P=0.30), even after removing TMB-TMS collinearity (P=0.56). Meta-analysis showed a consistent interaction direction (pooled HR=1.81, 95% CI 1.29–2.55, P=0.0006; I²=0%). In 100 repeated 70:30 splits in SYSUCC, the primary HR was >1 in 99.0% of held-out test sets (median HR=3.67; C-index=0.643). Conclusions: The prognostic direction of TMB in MSS cancers is determined by the balance of the TMS model. In balanced models, TMB is protective in TMS-Low patients and harmful in TMS-High patients; the two effects cancel out in unstratified analysis. In unbalanced models, TMB systematically shifts the TMS score, masking the interaction. This framework provides a unified explanation for the contradictory TMB literature and offers a practical four-quadrant risk stratification approach using routine sequencing data.