MicroRNA-based signatures of early and late immune-related adverse events to anti-PD1 treatment.

Abstract

2661 Background: Immune checkpoint inhibitors (ICIs) have transformed cancer treatment but are associated with toxicity in the form of immune-related adverse events (irAEs). We previously reported a genetic signature predicting anti-PD1 irAEs in a retrospective analysis of a heavily pretreated melanoma cohort, which validated in a pan-cancer cohort. Here we investigate the applicability of that signature in a prospectively collected cohort of GU, breast, and NSCLC cancer patients. We evaluated clinical and genetic differences between the original training set and this cohort and leveraged the expanded dataset to develop novel models for timing-specific anti-PD1 toxicity. Methods: We analyzed clinical and genetic differences in the melanoma training cohort (n=58) and the new cohort of patients, all treated with single agent anti-PD1/PDL1 therapy (n=137). Clinical and genetic differences were assessed using Fisher’s exact test for categorical variables: pre-treatment, concurrent radiation, and SNP genotype across 165 loci. Kruskal-Wallis tests were used for age, toxicity timing, and severity of toxicity. Predictive genetic models were constructed using elastic net, random forest, and boosted tree algorithms and evaluated using leave-one-out cross-validation (LOOCV) metrics. Outcomes included cycle-specific toxicity (early ≤5 cycles, late ≥15 cycles). SNPs were pre-filtered for inclusion in genetic models using Fisher or Jonckheere-Terpstra p-values (<0.2) for relevance to outcomes. Results: The training and current cohort were different in their toxicity timing, with earlier toxicity onset in the new cohort (median 5 cycles vs. 20 cycles in melanoma, Kruskal p = 0.00018). Genetic analysis identified 12 significantly different SNP genotypes (Fisher p < 0.05) between cohorts, including mir146A rs2910164 (Fisher p = 0.0005). This SNP was associated with early toxicity overall and within data subsets. Refining our model to account for cycle-specific toxicity events significantly enhanced performance. For late toxicity (≥15 cycles), the refined model achieved a LOOCV AUC of 0.793 (genetics + clinical). A newly developed early toxicity model (≤5 cycles) using genetics alone demonstrated robust predictive accuracy with an AUC of 0.753. Conclusions: These findings emphasize the importance of defining clinical and genetic diversity in refining predictive models for anti-PD1/PDL1 outcomes. The development of an early toxicity model offers significant clinical utility. Next steps will be to use time to event (toxicity) versus cycle number. This study provides a foundation for the application of personalized genetic tools to predict the safety of ICIs for currently treated patient cohorts. New LOOCV performance metrics in expanded PD1 data (n=234). Outcome Covariates Sensitivity Specificity PPV NPV F1 AUC Late Toxicity (>= 15 cycles) SNPs + Clin 0.692 0.894 0.450 0.959 0.545 0.793 Early Toxicity (<=5 cycles) SNPs Only 0.600 0.907 0.486 0.939 0.537 0.753