CURE AI as a predictor of TIGIT as a therapeutic target in pediatric cancers.

Abstract

e22000 Background: Clinical trials are not able to study individual treatment effects as methods to do so have not yet matured into practical use. For decades, we have studied groups of patients enrolled on a trial, comparing groups of patients as large cohorts and losing signal for important, complex features that define individuals. Foundation models take a different approach. Using the CURE AI foundation model generated from clinical and multi-omics data from hundreds of thousands of patients, complex, non-linear biological patterns can be found in new datasets that can provide insights into disease biology. We previously identified a complex signature predictive of immunotherapy response relative to chemotherapy response by analysis of a set of non-small cell lung cancer clinical trials (Weiss et al., AI in Precision Oncology, 2025). The immunotherapy benefit signature, applied to other cancer types such as colorectal cancer, identified known immunotherapy biomarkers such as MSI status as predictors of immunotherapy response. In the current study, we theorized that we could apply predictors of immunotherapy response/nonresponse from adult clinical trials to pediatric cancer patients. Methods: CURE AI was utilized to define the clinicogenomic features of therapeutic benefit by comparing tiragolumab/atezolizumab (anti-TIGIT/anti-PD-L1) versus atezolizumab (anti-PD-L1) on the Phase II CITYSCAPE first-line metastatic non-small lung cancer trial. CURE AI was fine-tuned on CITYSCAPE patient clinicogenomic features using the methodology described in the above cited publication. The model defined an omics-based benefit score that predicts benefit to TIGIT/PD-L1 combination therapy over PD-L1 monotherapy that is valid in held-out adult non-small cell lung cancer datasets. From this prediction, we then asked whether we could use this score to stratify pediatric patients from the St. Jude Cloud Genomics Platform by predicted benefit to anti-TIGIT therapy based on their baseline tumor RNA-sequencing. Results: Pediatric solid tumors were classified as having more benefit to TIGIT/PD-L1 combination therapy or as having more benefit to PD-L1 monotherapy. Multiple pediatric tumors including neuroblastoma, Wilms tumor, osteosarcoma, and rhabdomyosarcoma had major proportions of patients who are predicted to benefit more from TIGIT-containing combination therapy than PD-L1 monotherapy. When comparing patients who are predicted to benefit more from TIGIT-containing combination therapy than PD-L1 monotherapy, specific molecular signatures that are predictive of TIGIT-specific benefit were defined. Conclusions: As an indication expansion strategy, our approach to connect completed adult clinical trials to pediatric cancer treatment prediction could improve pediatric cancer trial design strategies.