Date of Award


Document Type

Open Access Dissertation

Degree Name

Doctor of Philosophy in Statistics (PhD)

Administrative Home Department

Department of Mathematical Sciences

Advisor 1

Kui Zhang

Committee Member 1

Qiuying Sha

Committee Member 2

Xiao Zhang

Committee Member 3

Hairong Wei


This dissertation contains three Chapters. The following is a concise description of each Chapters.

In Chapter 1, we introduced the Random Forest, a machine learning method, to foresee whether a virus is capable of infecting humans. The Covid pandemic informs us the importance of predicting the ability of a zoonotic virus that can infect humans from its genomic sequence. We used the -mer with and as features of a virus to predict if it can affect humans. We further employed the Boruta algorithm to select the important features, then fed those important features into the Random Forest method to train the model and make predictions. After utilizing a dataset that is independent of the training dataset in the test procedure, the results show that the accuracy of the training step is almost the same as an existing model, however, the accuracy in the testing step is substantially improved. Moreover, the time consumption of our method is much less than the existing model.

In Chapter 2, we developed a new application of Long Short-Term Memory (LSTM) deep learning method for the human leukocyte antigens (HLA) allele imputation and implemented it in a software package, called LSTM*HLA. Methods for HLA allele imputation utilize single nucleotide polymorphisms (SNPs) around HLA loci and their relationship with HLA alleles to predict HLA alleles. That is the similar fundamental scheme as Bidirectional LSTM. We organized several consecutive SNPs together as an element of inputs for each cell of the LSTM algorithm and made a final imputation for HLA alleles by averaging results from different sets of hyperparameters. We evaluated and compared the performance of our method with two commonly used methods for HLA imputation with seven real data sets: CookHLA as the representative of conventional approaches and Deep*HLA as the representative of machine learning methods. We find that our method not only performs well when the reference samples and the target samples are from the same ethnic group, but also achieves high accuracy when they are from distinctive ethnicities. Moreover, because deep learning methods hold the nature that is less dependent on Linkage Disequilibrium, LSTM*HLA could enhance the accuracy of low-frequent HLA alleles which has great influence in the fields of clinical research and personal care.

In Chapter 3, we investigated how two factors, the sample size and the choice of reference samples, can affect the accuracy of HLA imputation since these two factors are important factors that need to be carefully considered in real studies. As our results show, greater than 50 individuals is highly recommended for a reference panel to achieve a high imputation accuracy. For the choice of reference panels, the reference panel with the same ethnicity as target samples is strongly suggested, expanding the reference panel with multiple similar ethnic groups may also improve the accuracy, however, augmenting the reference panel with unrelated ethnic groups would decrease the imputation accuracy.