NEW YORK (GenomeWeb) – Biomarkers predicting cancer drug response can be uncovered from omics data of patient samples as well as cell lines, according to a new study.
Investigators have previously had mixed results when trying to use cancer cell line profiling to test the function and therapeutic relevance of molecular alterations uncovered in tumor samples. According to a trio of University of California, San Francisco, researchers this, is in part due to differences between cell lines and the tumors they represent.
They developed a method they dubbed modular analysis of genomic networks in cancer (MAGNETIC) that they used to uncover gene modules that are preserved in both tumor samples and cancer cell lines. When they applied this approach to breast cancer samples, they uncovered 219 such gene modules, but, as they reported in Cell Systems, many of the strongest signals from the Cancer Genome Atlas (TCGA), as well as other modules mostly made up of gene expression and methylation data, were not among them.
The researchers argued that their approach could help overcome a key issue in pharmacogenomics.
"[W]e show that modules preserved in cell lines can act as accurate biomarkers that are more robust than standard approaches because they are more reflective of a tumor context," UCSF's Sourav Bandyopadhyay and his colleague wrote in their paper. "This work reveals an approach for the integrative analysis of molecular programs within human tumors and provides a powerful and clinically relevant way to connect tumor genotype to therapy."
The researchers pulled together various threads of omic data, including gene expression, DNA methylation, copy number alteration, exome sequencing, and reverse phase protein array data, for 941 breast cancer patients from TCGA for analysis with MAGNETIC.
They first developed a correlation network by comparing pairs of gene features across patients, as well as within and between analysis platform types, and then combined these gene linkages into a single network. When they applied a clustering algorithm to identify inter-connected modules within the network, they identified 219 modules.
These modules, the researchers said, could reflect tumor biology. They noted that one module largely contained genes with promoters that were enriched for histone H3 lysine 27 trimethylation, which is linked to the repression of genes in development and differentiation.
They also reasoned that modules shared across patient samples and cancer cell lines would reflect shared biology and thus be better poised to inform pharmacogenomics studies. They examined whether their patient-derived modules were present in a panel of 82 breast cancer cell lines.
Not all modules were preserved, they reported. In particular, modules uncovered in the patient samples that drew mostly on gene expression or methylation data were not reflected in cancer cell lines. They suggested that those modules might instead reflect differences in biology between human tumor samples and cell lines.
The researchers then investigated whether these preserved modules could be used to discern drug sensitivities. Within a panel of 82 breast cancer cell lines, they uncovered 271 module-drug relationships, linking 74 drugs and 99 modules.
Slightly more than a third of these interactions could be predicted by PAM50 profiling, though the researchers reported that the addition of these modules improved that prediction. Meanwhile, for the other interactions that couldn't be predicted by PAM50 subtype, these modules could improve response prediction for three-quarters of the drugs. This suggested that the modules could be biomarkers, and complement subtype information.
They also argued that their approach could improve drug response predictions because of its reliance on patient samples. They compared gene- and module-based biomarkers on a set of 39 breast cancer patient-derived xenograft models. Of the 13 drugs common to the cell line and PDX studies, the response to three could be predicted in the cancer cell lines. For two of these three drugs, modules were better at predicting drug response than gene-based biomarkers, they reported.