Considerations for feature selection using gene pairs and applications in large-scale dataset integration, novel oncogene discovery, and interpretable cancer screening

BACKGROUND: Advancements in transcriptomic profiling have led to the emergence of new challenges regarding data integration and interpretability. Variability between measurement platforms makes it difficult to compare between cohorts, and large numbers of gene features have encouraged the use black box methods that are not easily translated into biologically and clinically meaningful findings. We propose that gene rankings and algorithms that rely on relative expression within gene pairs can address such obstacles. METHODS: We implemented an innovative process to evaluate the performance of five feature selection methods on simulated gene-pair data. Along with TSP, we consider other methods that retain more information in their score calculations, including the magnitude of gene expression change as well as within-class variation. Tree-based rule extraction was also applied to serum microRNA (miRNA) pairs in order to devise a noninvasive screening tool for pancreatic and ovarian cancer. RESULTS: Gene pair data were simulated using different types of signal and noise. Pairs were filtered using feature selection approaches, including top-scoring pairs (TSP), absolute differences between gene ranks, and Fisher scores. Methods that retain more information, such as the magnitude of expression change and within-class variance, yielded higher classification accuracy using a random forest model. We then demonstrate two powerful applications of gene pairs by first performing large-scale integration of 52 breast cancer datasets consisting of 10,350 patients. Not only did we confirm known oncogenes, but we also propose novel tumorigenic genes, such as BSDC1 and U2AF1, that could distinguish between tumor subtypes. Finally, circulating miRNA pairs were filtered and salient rules were extracted to build simplified tree ensemble learners (STELs) for four types of cancer. These accessible clinical frameworks detected pancreatic and ovarian cancer with 84.8 and 93.6% accuracy, respectively. CONCLUSION: Rank-based gene pair classification benefits from careful feature selection methods that preserve maximal information. Gene pairs enable dataset integration for greater statistical power and discovery of robust biomarkers as well as facilitate construction of user-friendly clinical screening tools.