The determination of the cancer prognosis is paramount for patients and medical personnel so that they can devise treatment strategies. Transcriptional-based signatures and subtypes derived from cancer biopsy material have been used in clinical practice for several cancer types to aid in setting the patient prognosis and forming treatment strategies. Other genomic features in cancer biopsies, such as copy number alterations (CNAs), have been underused in clinical practice, and yet they represent a complementary source of molecular information that can add detail to the prognosis, which is supported by recent work in breast, ovarian, and lung cancers. Here, through a systematic strategy, we explored the prognostic power of CNAs in 37 cancer types. In this analysis, we defined two modes of informative features, deep and soft, depending on the number of alleles gained or lost. These informative modes were grouped by amplifications or deletions to form four single-data prognostic models. Finally, the single-data models were summed or combined to generate four additional multidata prognostic models. First, we show that the modes of features are cancer-type dependent, where deep alterations generate better models. Nevertheless, some cancers require soft alterations to generate a feasible model due to the lack of significant deep alterations. Then, we show that the models generated by summing coefficients from amplifications and deletions appear to be more practical for many but not all cancer types. We show that the CNA-derived risk group is independent of other clinical factors. Furthermore, overall, we show that CNA-derived models can define clinically relevant risk groups in 33 of the 37 (90%) cancer types analyzed. Our study highlights the use of CNAs as biomarkers that are potentially clinically relevant to survival in cancer patients.
Keywords: TCGA; biomarkers; cancer prognosis; survival models.