Machine learning and computational methods to identify molecular and clinical markers for complex diseases – case studies in cancer and obesity

Abstract

In biomedical research, applied machine learning and bioinformatics are the essential disciplines heavily involved in translating data-driven findings into medical practice. This task is especially accomplished by developing computational tools and algorithms assisting in detection and clarification of underlying causes of the diseases. The continuous advancements in high-throughput technologies coupled with the recently promoted data sharing policies have contributed to presence of a massive wealth of data with remarkable potential to improve human health care. In concordance with this massive boost in data production, innovative data analysis tools and methods are required to meet the growing demand. The data analyzed by bioinformaticians and computational biology experts can be broadly divided into molecular and conventional clinical data categories. The aim of this thesis was to develop novel statistical and machine learning tools and to incorporate the existing state-of-the-art methods to analyze bio-clinical data with medical applications. The findings of the studies demonstrate the impact of computational approaches in clinical decision making by improving patients risk stratification and prediction of disease outcomes.

This thesis is comprised of five studies explaining method development for 1) genomic data, 2) conventional clinical data and 3) integration of genomic and clinical data. With genomic data, the main focus is detection of differentially expressed genes as the most common task in transcriptome profiling projects. In addition to reviewing available differential expression tools, a data-adaptive statistical method called Reproducibility Optimized Test Statistic (ROTS) is proposed for detecting differential expression in RNA-sequencing studies. In order to prove the efficacy of ROTS in real biomedical applications, the method is used to identify prognostic markers in clear cell renal cell carcinoma (ccRCC). In addition to previously known markers, novel genes with potential prognostic and therapeutic role in ccRCC are detected. For conventional clinical data, ensemble based predictive models are developed to provide clinical decision support in treatment of patients with metastatic castration resistant prostate cancer (mCRPC). The proposed predictive models cover treatment and survival stratification tasks for both trial-based and realworld patient cohorts. Finally, genomic and conventional clinical data are integrated to demonstrate the importance of inclusion of genomic data in predictive ability of clinical models. Again, utilizing ensemble-based learners, a novel model is proposed to predict adulthood obesity using both genetic and social-environmental factors.

Overall, the ultimate objective of this work is to demonstrate the importance of clinical bioinformatics and machine learning for bio-clinical marker discovery in complex disease with high heterogeneity. In case of cancer, the interpretability of clinical models strongly depends on predictive markers with high reproducibility supported by validation data. The discovery of these markers would increase chance of early detection and improve prognosis assessment and treatment choice.