Advanced kernel learning and its applications in healthcare
Healthcare plays such an important role for human well beings. Modern healthcare technologies generate extensive electronic health and medical records, and provide an unprecedented source of information that can potentially help to improve cancer treatment and develop personalized medicine. Such improvement requires sophisticated computational methods to uncover the important hidden information, machine learning in particular, and then the discovered information helps to diagnose and predict diseases, and possibly to cure cancers and save people's lives eventually.
Kernel learning has shown great performance on data that has no obvious fixed-dimensional vector space representation. Selecting a suitable kernel for a kernel-based machine learning task is crucial but challenging. This dissertation presents multiple kernel learning for SVM and fuzzy c-means, and then applies it successfully to Magnetic Resonance Imaging (MRI) segmentation. Moreover, a new machine learning based microRNA target prediction approach named "CLIPSeed" is developed, which is a powerful, friendly, and flexible tool for predicting microRNA targets given mRNAs containing seed matches to microRNAs. Two versions of algorithms are provided based on whether the conservation features are considered or not, also user can specify FDR according to the practical needs. Given a microRNA and mRNA pair, CLIPSeed predicts whether this mRNA is a target of microRNA and outputs the binding sites information if it is a target. Five-fold cross validation on PAR-CLIP ALL-AGO (AGO1-4) data shows that CLIPSeed achieves >2 fold increase in prediction performance measuring by precision. Furthermore, testing on independent CLIP data confirms good performance of CLIPSeed algorithm. Finally, human genome-wide testing using expression data by Gene Set Enrichment Analysis (GSEA) verifies that the predictions of CLIPSeed are significant and the predicted miRNAs targets are down-regulated at mRNA level. In addition, this dissertation analyzes the internal mechanism of miRNA-mRNA interaction based on machine learning approaches. CLIPSeed algorithm not only reveals common recognized important features for being a true target, but also discovers some new factors contributing to be a true target.
The presented multiple kernel learning approach improves MRI segmentation accuracy and leads to more accurate tumor detection and better image interpretation. CLIPSeed, the new machine learning based miRNA target prediction algorithm, makes significant improvement in performance and highly reduces the false positive rate comparing to the existing approaches. Because of its high precision prediction, CLIPSeed provides biologists a more powerful tool for further study of cancer treatment and personalized medicine. Therefore, CLIPSeed meets biologists' needs more and serves healthcare better. Meanwhile, the new discovered contributing factors of being a true microRNA target supply meaningful clues for curing cancers.