Bladder cancer gene expression prediction with explainable algorithms

Abstract

In this study, we aimed to classify bladder cancer patients using tumoral and non-tumoral gene expression data. In this way, we aimed to determine which genes are effective on tumoral and normal tissues. In addition, for this purpose, we planned to perform this classification using interpretable methods (The aim of this study was to classify bladder cancer patients using gene expression data from tumoral and non-tumoral tissues. By doing so, we wanted to determine which genes were effective on both tumoral and normal tissues. Moreover, for this purpose, we planned to use interpretable methods for this classification.). Analyses using permutation feature importance (PFI), SHapley Additive exPlanations (SHAP), local interpretable model-agnostic explanations (LIME), and Anchor methods on data from Gene Expression Omnibus (GEO) and Curated Microarray Database we did (We performed analyses using permutation feature importance (PFI), SHapley Additive exPlanations (SHAP), local interpretable model-agnostic explanations (LIME), and Anchor methods on data from Gene Expression Omnibus (GEO) and Curated Microarray Database.). These are eXplainable methods used to determine the importance of genes in classification. According to the results of our study, the most important genes were determined as LINC00161, ACACB, and CBARP according to the PFI method, HSPA6, STON2, and RFC2 according to the SHAP method, PRUNE2 and ABCC13 according to the LIME method, and TMEM74, KLHL10, and GAMT according to the Anchor method. This study shows that genes involved in other cancer types are also effective in bladder cancer. In addition, it has been observed that using explainable methods in cancer data can support prognosis and treatment in the clinic.