NEW YORK – A Columbia University-led team has started translating tumor transcriptomes into gene modules that show promise for predicting cancer prognoses.
"We find that biologically coherent sets of genes (gene modules) provide a rich source of biomarkers with substantial prognostic utility, superior to single-locus observations," senior and corresponding author Saeed Tavazoie, a biology, systems biology, biochemistry, molecular biophysics, and cancer dynamics researcher at Columbia, and his colleagues explained in Cell Genomics on Thursday.
"By utilizing DNA/RNA motif discovery in co-expressed genes," they added, "we have identified a large set of gene modules with significant prognostic value that promise to reveal novel biology with potentially significant contribution[s] to cancer progression."
For their study, the researchers first came up with a computational strategy designed to weigh the prognostic potential of a range of molecular features in tumors, in relation to other factors that are known to impact patient outcomes and survival such as tumor stage, histopathology, and patient age.
When they applied this approach to pan-cancer tumor features and survival patterns for 8,620 patients from 25 Cancer Genome Atlas (TCGA) cohorts, for example, the approach pointed to the apparent prognostic advantages of transcriptome features over other tumor alterations such as somatic gene mutation status or copy number variations, prompting further analyses on biologically informed expression modules.
With 199 gene modules centered on 22 tumor suppressor genes and 45 oncogenes, for example, the team established "module perturbation scores" (MPSs) related to dysregulation of those modules. In modules for 40 of the 67 cancer-related genes considered, those MPSs appeared to reflect not only gene expression changes, but also underlying somatic mutation and copy number changes.
Those module shifts, in turn, often coincided with patient outcomes, the researchers reported, with positive or negative MPSs for pancreatic cancer, stomach cancer, melanoma, and other cancers tracking more closely with patient survival than did mutation status for the cancer-associated genes at the heart of the same modules.
Based on those findings, the team went on to search for prognostic modules outside of cancer-related genes, putting together modules based on thousands of pathways, gene functions, or regulatory targets, systematically testing MPSs for these modules across TCGA samples — and flagging module perturbations linked to disease progression or survival within specific cancer types in TCGA.
"Our results suggest that MPSs capture at least some of the molecular complexity underlying cancer," the researchers suggested, adding that "prognostic power of individual cancer modules motivated us to develop high-order machine-learning approaches that generate models by combining multiple [prognostic cancer modules] with substantially improved survival prediction."
More broadly, the researchers argued that such findings point to the importance of performing transcriptomic analyses on broader sets of tumor biopsies beyond the research setting, adding that "cost-effective transcriptomic sequencing at reduced depths may be advantageous" to address sample storage and handling limitations.
"[W]e expect the predictive power of module-based approaches, which are less reliant on noisy single-gene measurements, capturing patterns of coordinated gene expression changes instead, to be minimally impacted by information loss even when samples are sequenced at shallow depths," they explained.