Integration and comparison of different genomic data for outcome prediction in cancer

In cancer, large-scale technologies such as next-generation sequencing and microarrays have produced a wide number of genomic features such as DNA copy number alterations (CNA), mRNA expression (EXPR), microRNA expression (MIRNA), and DNA somatic mutations (MUT), among others. Several analyses of a specific type of these genomic data have generated many prognostic biomarkers in cancer. However, it is uncertain which of these data is more powerful and whether the best data-type is cancer-type dependent. Therefore, our purpose is to characterize the prognostic power of models obtained from different genomic data types, cancer types, and algorithms. For this, we compared the prognostic power using the concordance and prognostic index of models obtained from EXPR, MIRNA, CNA, MUT data and their integration for ovarian serous cystadenocarcinoma (OV), multiform glioblastoma (GBM), lung adenocarcinoma (LUAD), and breast cancer (BRCA) datasets from The Cancer Genome Atlas repository. We used three different algorithms for prognostic model selection based on constrained particle swarm optimization (CPSO), network feature selection (NFS), and least absolute shrinkage and selection operator (LASSO).

The integration of the four genomic data produced models having slightly higher performance than any single genomic data. From the genomic data types, we observed better prediction using EXPR closely followed by MIRNA and CNA depending on the cancer type and method. We observed higher concordance index in BRCA, followed by LUAD, OV, and GBM. We observed very similar results between LASSO and CPSO but smaller values in NFS. Importantly, we observed that model predictions highly concur between algorithms but are highly discordant between data types, which seems to be dependent on the censoring rate of the dataset.

Gene expression (mRNA) generated higher performances, which is marginally improved when other type of genomic data is considered. The level of concordance in prognosis generated from different genomic data types seems to be dependent on censoring rate.