Albert Kuo @ DIMACS

DIMACS 2015

Albert Kuo

office:	CoRE 434
school:	The University of Chicago
e-mail:	albertkuo@uchicago.edu
project:	Development of machine learning algorithms in biology

The Project

DIMACS (the Center for Discrete Mathematics and Theoretical Computer Science) was founded as one of 24 Science and Technology Centers funded by the NSF. It hosts a 9-week REU program, which I attended in the summer of 2015.

I worked with Dr. Kevin Chen and explored epigenomic data for predicting enhancers -- sequences in the DNA that regulate gene expression. The machine learning algorithms I used included hidden Markov models (HMMs), support vector machines (SVMs), and logistic regression. I combined various datasets with these models to better understand the relationships between enhancer marks and their role in enhancer prediction.

Weekly Log

Week 1

I met with my mentor to discuss the goals and structure of my project. To prepare for my work this summer, I reviewed linear algebra topics such as singular value decomposition and read about hidden Markov models, principal components analysis, and the expectation-maximization (EM) algorithm. I was also given papers on tensor decomposition methods to read. On Thursday, we had a lab meeting with other professors and their groups to understand the lab work behind some of the data we are using.

Week 2

I read up on molecular biology, in particular on the relatively recent discovery of enhancer RNAs (eRNAs). I conducted exploratory analysis in R on four cell lines from the ENCODE project using two datasets. One dataset consisted of enhancer predictions using chromatin marks, i.e. whether a section of a gene is likely to be an enhancer based on histone modifications. The other dataset consisted of eRNA expression in TPM (transcripts per million). The two datasets were compared using various plots.

Week 3

I conducted Wilcoxon tests to compare the eRNA expression level of weak versus strong enhancers. I also ran Spectacle, which is based on ChromHMM, on one cell line (GM12878) with the additional input of eRNA expression.

Week 4

To further analyze the relationship between histone modifications and eRNA expression in terms of predicting enhancers, I examined the degree to which each chromatin state overlapped with intervals containing the regulatory protein, p300, using different plots. I also looked at the distance to the closest transcription start site (TSS) for the sections in each chromatin state by calculating some statistics and plotting the data. I did my data munging in Python using the 'pandas' package.

Week 5

I read papers on enhancer prediction based on various statistical models. I began implementing a support vector machine (SVM) in Python using packages from scikit-learn to predict enhancers.

Week 6

I ran SVM to predict enhancers using histone marks and eRNA and examined the feature weights. After discussing the results with my mentor, I changed how I selected the negative set from a random sampling to an interval-based selection. I also began reading about multi-view/Naive Bayes classification. I spent the rest of the week preparing for my presentation.

Week 7

I ran Spectacle with the additional input of p300 and experimented with different number of states. I read up on feature selection and feature ranking for SVM and L1 and L2 regularization for logistic regression. A particular concern was that correlation between features would lead to misleading feature weights. I ran SVM with principal components and logistic regression for enhancer prediction.

Week 8

I tuned the penalty parameters for SVM and logistic regression using cross-validation. I finished writing my report for DIMACS. I ran logistic regression with the additional feature from sequence data (transcription factor binding sites conserved in the human/mouse/rat alignment).

Week 9

I added two more datasets, Enhancer DRM and DNase, as features and replicated my results on 3 other cell types from the ENCODE project. In addition, I looked at two-way interaction factors in logistic regression. Lastly, I compiled all my results from the summer into a report for my mentor.

Links

My Work

Mid-Program Presentation Slides

Final Report

Github Repository

Additional Information

- Program Info: DIMACS REU

- My mentor: Dr. Kevin Chen