Subtype prediction in pediatric acute myeloid leukemia: Classification using differential network rank conservation revisited

Background: One of the most important application spectrums of transcriptomic data is cancer phenotype classification. Many characteristics of transcriptomic data, such as redundant features and technical artifacts, make over-fitting commonplace. Promising classification results often fail to generalize across datasets with different sources, platforms, or preprocessing. Recently a novel differential network rank conservation (DIRAC) algorithm to characterize cancer phenotypes using transcriptomic data. DIRAC is a member of a family of algorithms that have shown useful for disease classification based on the relative expression of genes. Combining the robustness of this family's simple decision rules with known biological relationships, this systems approach identifies interpretable, yet highly discriminate networks. While DIRAC has been briefly employed for several classification problems in the original paper, the potentials of DIRAC in cancer phenotype classification, and especially robustness against artifacts in transcriptomic data have not been fully characterized yet. Results: In this study we thoroughly investigate the potentials of DIRAC by applying it to multiple datasets, and examine the variations in classification performances when datasets are (i) treated and untreated for batch effect; (ii) preprocessed with different techniques. We also propose the first DIRAC-based classifier to integrate multiple networks. We show that the DIRAC-based classifier is very robust in the examined scenarios. To our surprise, the trained DIRAC-based classifier even translated well to a dataset with different biological characteristics in the presence of substantial batch effects that, as shown here, plagued the standard expression value based classifier. In addition, the DIRAC-based classifier, because of the integrated biological information, also suggests pathways to target in specific subtypes, which may enhance the establishment of personalized therapy in diseases such as pediatric AML. In order to better comprehend the prediction power of the DIRAC-based classifier in general, we also performed classifications using publicly available datasets from breast and lung cancer. Furthermore, multiple well-known classification algorithms were utilized to create an ideal test bed for comparing the DIRAC-based classifier with the standard gene expression value based classifier. We observed that the DIRAC-based classifier greatly outperforms its rival. Conclusions: Based on our experiments with multiple datasets, we propose that DIRAC is a promising solution to the lack of generalizability in classification efforts that uses transcriptomic data. We believe that superior performances presented in this study may motivate other to initiate a new aline of research to explore the untapped power of DIRAC in a broad range of cancer types.