Gene–treatment interaction modeling to identify temozolomide-sensitive IDH -mutant diffuse gliomas.

2085 Background: IDH-mutant diffuse gliomas are biologically heterogeneous. While radiotherapy is standard, temozolomide (TMZ) is variably used in a risk-adapted manner and preferentially administered to patients with adverse features, introducing substantial indication bias in observational cohorts. Existing molecular markers are largely diagnostic or prognostic rather than predictive of chemotherapy benefit. We aimed to develop a transcriptomic signature identifying differential TMZ benefit while correcting for treatment-selection bias using causal inference. Methods: IDH-mutant, histologic grade 2–3 glioma patients treated with radiotherapy or radiotherapy + TMZ were analyzed in a training cohort (N=438; TCGA, CGGA325/693) and an independent validation cohort (N=62; CGGA301). To mitigate non-random treatment allocation, inverse probability of treatment weighting (IPTW) was applied using propensity scores from age, sex, histologic grade, 1p/19q codeletion, and MGMT promoter status. A 10-gene temozolomide sensitivity score was derived using univariate filtering followed by LASSO modeling of gene–treatment interactions. Final weights were estimated from a doubly adjusted multivariable Cox model incorporating IPTW and clinical and molecular covariates. Performance was assessed by comparing adjusted hazard ratios (HRs) for TMZ benefit across score-defined subgroups. Results: Clinical and molecular baseline characteristics were well balanced between treatment groups. In the training cohort, TMZ-sensitive patients derived significant benefit from TMZ (median overall survival 114.0 vs 85.7 months; adjusted HR 0.58, 95% CI 0.35–0.96; P=0.035). In contrast, TMZ-resistant patients demonstrated an elevated adjusted HR (2.98, 95% CI 1.63–5.45; P<0.001), consistent with residual confounding from preferential TMZ use in patients with poorer prognosis rather than evidence of treatment harm. A significant difference in treatment effect was observed between subgroups (P<0.001). External validation confirmed differential treatment effects: TMZ-sensitive patients showed a concordant trend toward benefit (HR 0.51, 95% CI 0.13–1.94; P=0.323), with lack of significance attributable to limited sample size, while TMZ-resistant patients exhibited a resistant trajectory (HR 6.04, 95% CI 1.56–23.38; P=0.009), with the difference in treatment effect remaining significant (P=0.011). Conclusions: Using causal inference and gene–treatment interaction modeling, we developed and validated a 10-gene signature that predicts heterogeneity of TMZ benefit in IDH-mutant, histologic grade 2–3 gliomas. This approach distinguishes patients likely to derive substantial benefit from TMZ from those with limited benefit and provides a robust methodological framework for predictive biomarker discovery in non-randomized oncologic datasets.