Video
Author(s):
Guru P. Sonpavde, MD, discusses the methodology and key results from an analysis of the Cancer Genome Atlas in muscle-invasive bladder cancer.
Guru P. Sonpavde, MD, medical director, Genitourinary (GU) Oncology, assistant director, Clinical Research Unit, Christopher K. Glanz Chair, Bladder Cancer Research, AdventHealth Cancer Institute, discusses the methodology and key results from an analysis of the Cancer Genome Atlas (TCGA) in muscle-invasive bladder cancer (MIBC).
The risk of recurrence in MIBC after radical cystectomy without neoadjuvant therapy is correlated with pathological stage. To better predict long-term patient survival outcomes after radical cystectomy, a prognostic model was designed, Sonpavde begins. The model utilized deep learning and neural network analysis of previous tumor and germline profiling. Deep learning analysis integrates interactions between clinical features rather than evaluating isolated genetic or tumor alterations, Sonpavde explains.
Evaluable patients were in TCGA database and had undergone radical cystectomy. The primary objective of this study was to predict 5-year survival by integrating genomic and transcriptomic data from germline and tumor tissues, Sonpavde says.
Patients were carefully selected, Sonpavde emphasizes. Patients were required to have survival data at 5-years post-radical cystectomy or have passed within 5 years of treatment. However, patients who passed away at or beyond 4 months of diagnosis were excluded from this population, as this outcome could be potentially attributed to post-operative complications rather than bladder cancer–related mortality, Sonpavde notes. Patients who had passed away were then propensity-matched against patients alive at 5 years according to distinct clinical features, Sonpavde continues. These included pathologic stage, adjuvant chemotherapy, gender, race, age and smoking status.
Results showed that germline alterations were a major prognostic factor for 5-year survival, followed by tumor genomic alterations, Sonpavde states. Interestingly, tumor transcriptomics had a comparatively low prognostic value, he details. Moreover, the deep learning model showed 97% accuracy on 38 training samples, and 94% accuracy across the 79 evaluation samples. This indicates that neural network analysis of germline and tumor profiling may be a valid and optimal tool to improve patient selection for adjuvant therapy and tailored surveillance in this population, Sonpavde concludes.