Artificial Intelligence-Based Multimodal Prediction of Postoperative Adjuvant Immunotherapy Benefit in Urothelial Carcinoma: Results From the Phase III, Multicenter, Randomized, IMvigor010 Trial

While circulating tumor DNA (ctDNA) testing has demonstrated utility in identifying muscle-invasive urothelial carcinoma (MIUC) patients likely to benefit from adjuvant immunotherapy, the prognostic value of transcriptome data from surgical specimens remains underexplored. Using transcriptomic and ctDNA data from the IMvigor010 trial, we developed an artificial intelligence (AI)-driven biomarker to predict immunotherapy response in urothelial carcinoma, termed UAIscore. Patients with high UAIscore had significantly better outcomes in the atezolizumab arm versus the observation arm. Notably, the predictive performance of the UAIscore consistently outperformed that of ctDNA, tTMB, and PD-L1, highlighting its value as an independent biomarker. Moreover, combining ctDNA, tTMB, and PD-L1 with the UAIscore further improved predictive accuracy, underscoring the importance of integrating multi-modality biomarkers. Further analysis of molecular subtypes revealed that the luminal subtype tends to be sensitive to adjuvant immunotherapy, as it may exhibit the highest level of immune infiltration and the lowest degree of hypoxia. Remarkably, we elucidated the role of the NF-κB and TNF-α pathways in mediating immunotherapy resistance within the immune-enriched tumor microenvironment. These findings stratify patients likely to respond to adjuvant immunotherapy, concurrently providing a mechanistic rationale for combination therapies to augment immunotherapy efficacy in urothelial carcinoma.