Gene signature for response prediction to immunotherapy and prognostic metarkers in metastatic Urothelial Carcinoma

Peter Langfelder^1,2,3En-Tni Lin¹Yi-Ta Tsai⁴Tai-Lung Cha⁵Grace S. Shieh^1,6,7,8*

• ¹Inst. of Statistical Science, Institute of Statistical Science, Academia Sinica, Taipei, Taiwan
• ²Center for Neurobehavioral Genetics, The Jane and Terry Semel Institute for Neuroscience and Human Behavior, University of California, Los Angeles, Los Angeles, United States
• ³Department of Psychiatry and Biobehavioral Sciences, David Geffen School of Medicine, University of California, Los Angeles, Los Angeles, United States
• ⁴Graduate Institute of Life Sciences, National Defense Medical Center, Taipei, Taiwan
• ⁵National Institute of Cancer Research, National Health Research Institutes (Taiwan), Miaoli, Taiwan
• ⁶Bioinformatics Program, Taiwan International Graduate Program, Academia Sinica, Taipei, Taiwan
• ⁷Data Science Degree Program, Academia Sinica and National Taiwan University, Taipei, Taiwan
• ⁸Genome and Systems Biology Degree Program, Academia Sinica and National Taiwan University, Taipei, Taiwan

To date, immune checkpoint inhibitors (ICIs) have emerged as a leading treatment for metastatic cancer, significantly improving patient survival while causing relatively few 2 side effects. However, the objective response rate for ICIs remains low approximately 30% in urothelial carcinoma (UC), underscoring the urgent need for predictive response biomarkers. Several state-of-the-art signatures have been revealed in top-tier journals, highlighting the importance of this field. As the number of genes (~20,000) far exceeds the sample sizes of typical training sets (generally ≤ 300), we first developed feature selection procedures to reduce the number of features to a few hundred. We then trained multiple machine learning classifiers using the selected genes and the IMvigor210 dataset, which includes RNA-seq and clinical data from ~298 patients with metastatic UC (mUC). Notably, our predictor LogitDA, using the identified 49-gene signature, achieved a prediction AUC of 0.75 in an independent dataset, PCD4989g(mUC). Moreover, our signature outperformed six state-of-the-art signatures, PD-L1 IHC, and five tumor microenvironment signatures, including IFN-γ, T-effector, and T-cell exhaustion signatures. When we integrated each of the six known signatures with our own, our signature still surpassed the integrated ones in terms of prediction AUC and accuracy in the PCD4989g(mUC) dataset. From our signature, we identified key prognostic biomarkers, with the top five markers LYRM1, RFC4, CENPI, SPAG5, and CACYBP (Benjamini-Hochberg adjusted P < 0.0025) in the IMvigor210 dataset. Finally, we performed pathway analyses using Reactome (MSigDB) and KEGG, to reveal some immune-related pathways enriched such as MHC class II antigen presentation.