4. Conclusions
An intelligent computational framework was applied to DNA methylation profiling data from an Italian epidemiological cohort comprising breast cancer, B-cell lymphoma and control samples. DNA methylation data, extracted by Illumina’sInfinium Human Methylation 450K Bead Chip, was used to select from an initial set of ~480,000 genome-scale methylation measurements, subsets of CpG sites that correspond to epigenetic biomarkers and show pre-disposition to the particular cancer phenotypes studied here. The framework applied included an evolutionary feature selection scheme, a novel selection scheme based on the semantic exploration in the GO tree, and a suite of classifiers appropriately evaluated on the basis of available data. Results showed that the subsets of CpG sites, delivered by the feature selection schemes, can provide encouraging classification performance measurements obtained both on resampling and testing to an unknown set. The biologically-inspired methodology proposed here of selecting biomarkers yielded promising results both for the various classification tasks undertaken and the biological content delivered, thus comprising another strategy for the derivation of biomarkers from molecular data of various kinds.