Tianyi Liu's Senior Honors Thesis

---

Repository for the R project for my senior honors thesis. I will constantly add to or edit the files. Questions and concerns regarding the codes should be directed to me through this link. The project is set to be finished by mid-April, 2019.

Research Supervisor

• Dr. Naim Rashid, Department of Biostatistics, UNC Gillings School of Global Public Health

Committee Members

• Dr. Jen-jen Yeh, Departmemt of Surgery, UNC School of Medicine
• Dr. Michael Kosorok, Department of Biostatistics, UNC Gillings School of Global Public Health

Research Synopsis

Multiple studies have been published extracting gene expression information from cancer patients via RNA sequencing. In tandem with known patient subtypes, machine learning tools can build on such data to assist in predicting cancer subtype in new patients. It has been previously demonstrated that parametric approaches such as penalized logistic regression can perform well in predicting the tumor subtypes from trial-generated RNA-seq data. However, non-parametric models such as random forests and support vector machines may offer more robustness to issues such as non-linearity in variable effects and potential interactions between genes. In addition, cross-study variability in the effect of each gene may further impact the accuracy of prediction. I am currently using existing machine learn approaches, such as random forests and support vector machines, to examine the accuracy of several strategies for training such models to account for cross-study heterogeneity in predicting cancer subtype, including rank-based transformation of expression datasets and horizontal data integration prior to model training. I will also apply recently developed multi-task learning techniques to better account for between-study heterogeneity jointly across multiple datasets and also demonstrate the improvement in prediction accuracy over the prior approaches.

---

Notes

• Datasets used in this project were uploaded to the 'dataset_used' folder.
• If you wish to run this script locally, please do the following:
  • Make sure your R is at least at version 3.5.0 (or the RMTL package will complain).
  • Download the files from the repository.
  • Create a new R project on the folder in which you store the files.
  • Create two directories named "result_tables" and "models_and_predictions" using the following command line
    • mkdir result_tables models_and_predictions
  • Specify Datasets and their local directory in file_loading.R.
  • Run the R scripts whose name starts with ml (actual code for training and prediction).
• result_visualization.Rmd contains all figures that will appear in the final thesis. The work is still in progress.