DIMACS REU 2013
Participants List

Name:	Scott Norton
Email:	scott.s.norton@hotmail.com
Office:	CoRE 434
Home Institution:	University of Connecticut - Map - Website
Project:	Statistical Methods for Finding New miRNA Motifs
Curriculum Vitae:	Link
A relevant song parody	Get Results Out Of You

---

About My Project

miRNA (microRNA) are a class of short (~22-nucleotide) RNA molecules found in the cytoplasm of eukaryotic cells. Their function in the cell is to inhibit gene expression after transcription into messenger RNA (mRNA) but before translation into protein. Within the miRNA is a short (6-8 nucleotide, sometimes split) sequence. This sequence, called the motif, is the site at which the miRNA binds directly to the mRNA at the downstream regulatory region, known as the 3-prime untranslated region (3' UTR). This study will help us identify miRNA motifs associated with different cellular processes such as metabolism, sleep/hibernation, and replication. When problems occur in miRNA actions, certain vital processes are either silenced or over-expressed. If these problems interfere with DNA maintenance or apoptosis (programmed cell death), they present an elevated risk for the development of cancer.

Gene expression levels can be measured using the relative concentrations of the end products - usually proteins. In a microarray, small tissue samples are exposed to fluoresent antibodies, one per row, each of which lights up when it binds to a unique target protein. The more of a protein there is in a sample, the stronger the total fluorescence is. This process can be repeated for however many tissue samples the experimenter desires. Since we are looking at the effects of miRNA, our microarray data compares the expression levels of certain proteins in normal mice with those in mice for which the miRNA processing pathway has been disabled.

In order to identify miRNA motifs which are associated with specific cellular processes, our goal is to develop a new statistical method for correlating RNA microarray expression data with specific miRNA motifs. Our study will be based on gene expression data and 3' UTR sequences obtained from mice.

---

Research Log

Week 1: Jun 3-7

In progress: coding a module for the Python programming language that performs a multiple linear regression from a vector-to-scalar mapping. Linear regressions will be the foundation for the motif predicting model. First version is a naive regression on the sum of terms against the residuals in each iteration.

Pesentation slides, June 7

Week 2: Jun 10-14

Results and Status

The multiple linear regression module has passed the test for small integer and random input - the output is sensible and reasonably close to an ideal fit.

Next steps

This week, I will be familiarizing myself with the primary software package, GEMMA, which correlates SNP (single nucleotide polymorphism) genotype and phenotype data. It is possible that SNPs will come into play when we analyze data from knockout mice - that is, mice where we negate, at conception, the functionality of a gene by inducing a single nucleotide mutation.

Week 3: Jun 17-21

Analysis

The source data come from two sources: the 3' UTR sequences originate from the University of California, Santa Cruz (UCSC) and are labeled with Reference Sequence Database (RefSeq) accession numbers, and the microarray data were obtained experimentally and are labeled with Ensembl accession numbers. The experimental procedure for obtaining the microarray data (phenotypes) was to breed two sets of mice, one normal and one with a mutation that prevents the miRNA from maturing. At various points in the mice's life cycles, the levels of protein products were measured. About 26,000 proteins were sampled.

Next steps

The first task for this week is to pair the two data sets and filter out any missing entries from the lookup table. After that, I need to filter out any data points for which the 3' UTR nucleotide sequence is a substring of another 3' UTR sequence. This ensures that the data do not have duplicates. From there, the kinship matrix must be computed. Gemma, our linear mixed model software, requires a kinship matrix to estimate the size of the random effect on variance in the data. This matrix will be generated by correlating the k-mer frequencies for each sequence, where \(k\leq6\).

Results

The filter removed around 5000 phenotypes for which no 3' UTR was available, and another 3000 where the 3' UTR sequence was an exact match to part of a larger sequence in the data set.

Week 4: Jun 24-28

Next steps

Gemma was originally designed to work with single nucleotide polymorphisms. Therefore, it will only recognize true/false values for the genotype data. Gemma will be run using this information combined with the gene expression data and the kinship matrix from last week. We will also compute two additional kinds of kinship matrices: the dimer count kinship matrix which is computed using the same algorithm as last week with \(k=2\), and the shared motifs kinship matrix which takes the dot product of two rows of the binary matrix normalized to fall in \([0, 1]\).

Results

The kinship matrices were successfully generated. Gemma was run using each matrix, and the output seemd to not depend heavily on the matrix used, indicating that the various measurements of relatedness we used are not significantly different from each other.

Friday 6.28: Happy Tau Day, and Happy Birthday to me!

Week 5: Jul 1-5

Next steps

The p-values from the Wald linear mixed model will be used for each motif to test the hypothesis that the presence or absense of the motif is positively correlated with the expression data. Using an \(\alpha\) of 0.05 and a Bonferroni-corrected \(\alpha\) of \(1.22 \cdot 10^{-5}\), the motifs such that (\p<\alpha\) will be compared to the results of a naive linear model taken on the presence/absence data.

Analysis

We are looking for a partial overlap between the two sets of significant motifs. If the sets have no motifs in common, or if they are close to identical, there may be a problem with one or more of our methods.

Week 6: Jul 8-12

Results

We have 21 motifs from the Gemma output and 26 from the naive model. Of these, 17 motifs are common to both sets. However, the correspondance between these motif candidates and documented miRNA motifs is low, indicating a possible problem with the output. Upon further investigation, we determined that the 3' UTR sequences used were in fact the reverse complement of what we desired. Recall that DNA is double-stranded with one, "sense" strand containing the literal code and the other, "antisense" strand acting as a template to transcribe the sense strand. While our analysis thus far was based on the sense strand, the actual miRNA motifs would be found in the antisense strand.

Next steps

We expect a significiant improvement in the correspondance between the linear model's results and published results from redoing the analysis based on the antisense strand. This includes reconstructing the kinship matrices for Gemma.

Week 7: Jul 15-19

Results

Using the shared sixmers kinship matrix, the output from Gemma was nearly identical to the output from simply correlating the presence and absense of each sixmer with the observed phenotype. In addition, there is a stronger association between the most likely miRNA motifs and the published miRNA sequences compared to the observations from using the sense strand. These motif candidates are far richer in adenine and thymine, which is the case for the majority of known miRNA motifs. However, the size of the random effect on the Gemma output is around \(10^{-2}\) times the size of the marker effect, indicating that the kinship matrix has little influence on the output. This is not surprising considering how similar the output is to that of the simple model, which was computed with no consideration for random effects.

Gemma returned much better results with the dimer counts correlation kinship matrix. Not only were they significantly different from the output from using the shared sixmers kinship, but there are about 40% more matches with the list of mature miRs obtained from miRBase.org. In addition, the ratio of the random effect coefficient to the marker effect coefficient is on the order of \(10^0\). So far, this is our best model.

Presentation slides, July 19

Week 8: Jul 22-26

Exit interview with Kevin. Our current task is to organize the directory containing all the files used and generated during the semester. As part of this, I am transcoding, consolidating, and optimizing the codes to simplify future investigation.

---

References & Links

Here are the papers I have read for my project:

1. Lee RC, Feinbaum RL, Ambros V (December 1993). "The C. elegans heterochronic gene lin-4 encodes small RNAs with antisense complementarity to lin-14". Cell 75 (5): 843-54.
2. Bussemaker, H. J., Li, H., & Siggia, E. D. (2001). Regulatory element detection using correlation with expression. Nature, 167-171.
3. Ben-Moshe, N. B., Avraham, R., Kedmi, M., Zeisel, A., Yitzhaky, A., Yarden, Y., & Domany, E. (2012, August 29). "Context Specific MicroRNA Analysis (CoSMic): Identification of Functional MicroRNAs and their mRNA targets - Supplementary Data." Retrieved from Nucleic Acids Research: Oxford Journals: http://nar.oxfordjournals.org/content/suppl/2012/08/29/gks841.DC1/
   nar-00762-n-2012-File009.pdf
4. Karolchik D, Hinrichs AS, Furey TS, Roskin KM, Sugnet CW, Haussler D, Kent WJ. The UCSC Table Browser data retrieval tool. Nucleic Acids Res. 2004 Jan 1;32(Database issue):D493-6.
5. Paul Flicek, Ikhlak Ahmed, M. Ridwan Amode, Daniel Barrell, Kathryn Beal, Simon Brent, Denise Carvalho-Silva, Peter Clapham, Guy Coates, Susan Fairley, Stephen Fitzgerald, Laurent Gil, Carlos Garcia-Girón, Leo Gordon, Thibaut Hourlier, Sarah Hunt, Thomas Juettemann, Andreas Kähäri, Stephen Keenan, Monika Komorowska, Eugene Kulesha, Ian Longden, Thomas Maurel, William McLaren, Mattieu Muffato, Rishi Nag, Bert Overduin, Miguel Pignatelli, Bethan Pritchard, Emily Pritchard, Harpreet Singh Riat, Graham R. S. Ritchie, Magali Ruffier, Michael Schuster, Daniel Sheppard, Daniel Sobral, Kieron Taylor, Anja Thormann, Stephen Trevanion, Simon White, Steven P. Wilder, Bronwen L. Aken, Ewan Birney, Fiona Cunningham, Ian Dunham, Jennifer Harrow, Javier Herrero, Tim J. P. Hubbard, Nathan Johnson, Rhoda Kinsella, Anne Parker, Giulietta Spudich, Andy Yates, Amonida Zadissa and Stephen M. J. Searle. Ensembl 2013 Nucleic Acids Research 2013 41 Database issue:D48-D55
6. Chen, Uing, Huang, Zhouqing, Wang, Liansheng, Wang, Yue, Wu, Feizhen, Meng, Shu, & Wang, Chanqian. "MicroRNA-125a-5p partly regulates the inflammatory response, lipid uptake, and ORP9 expression in oxLDL-stimulated monocyte/macrophages". Cardiovascular Research, Oxford Journals. 12 June 2009, Vol 83.1 pg 131-139, at http://cardiovascres.oxfordjournals.org/content/83/1/131
7. Xiang Zhou and Matthew Stephens. Genome-wide efficient mixed-model analysis (GEMMA) for association studies. Nature Genetics, 44: 821-824, 2012
8. Xiang Zhou, Peter Carbonetto and Matthew Stephens. Polygenic modeling with Bayesian sparse linear mixed models. PLOS Genetics. in press. http://arxiv.org/abs/1209.1341
9. Bartel, David P (January 2009). "MicroRNAs: Target Recognition and Regulatory Functions". Cell 136 (2): 215-33.
10. Bartel, David P (January 2004). "MicroRNAs: genomics, biogenesis, mechanism, and function". Cell 116 (2): 281-97.

Here is my mentor's website, and the REU website:

• The Chen Group
• The REU Website