Prognostic biomarker detection, machine learning bias correction, and differential coexpression module detection

In this thesis, we present three projects on prognosis biomarker detection, machine learning bias correction and identification of differential coexpression modules in complex diseases. In the first project, we aimed to identify fusion transcripts that are of predictive value on prostate cancer prognosis, an important task to avoid overtreatment to patients. We discovered eight fusion transcripts from 19 RNA-seq datasets and validated its predictive value on >200 patients from three sites (Pittsburgh, Stanford and Wisconsin). The constructed prediction model showed consistently high accuracy on predicting prostate cancer recurrence and aggressiveness in all three cohorts. In the second project, we consider a common practice to apply many (up to several hundred) machine learning classifiers to a dataset and report the best cross-validated accuracy. We demonstrated a downward bias using this approach and proposed an inverse power law (IPL) method to correct the bias. The method was compared with several existing methods using simulation and real datasets and showed superior performance. For the third study, we developed a computational algorithm (MetaDiffNetwork) to identify coexpressioin modules that are consistently differential across disease conditions in multiple transcriptomic studies. We demonstrated good performance of the algorithm using simulated data and applied it to combine eight major depressive disorder studies (cases vs. controls) and four breast cancer studies (ER+ vs. ER-). The identified modules were validated by existing knowledge of disease pathways. These modules can be used to help generate new hypotheses regarding suspected disease genes. In conclusion, the three areas of research covered in this thesis are critical bioinformatic elements for biomedical applications and can be used to help understand the underlying disease mechanism and ultimately improve patient treatment.