Final project of TRGN 510

Title

Building a classification model of NGS RNA-seq data with MLseq.

Author :Bo Cheng

Overview

1. In this project, I try to use a machine learning package, MLSeq from Bioconductor to find out the best model that predicts breast cancer subtype.

2. For my project, the machine learning process will use 5-fold cross-validation, and 10 repeats. I use 28 data sets from TCGA to train and test this model using 12 datasets.

3. The main goal of this project is to have a taste of machine learning in R. So I may do many copy and paste, but I will give my understanding and opinions in the R notebook.

4. The original package link: https://www.bioconductor.org/packages/devel/bioc/vignettes/MLSeq/inst/doc/MLSeq.pdf

5. The original Vignettes/sample link: https://www.bioconductor.org/packages/devel/bioc/vignettes/MLInterfaces/inst/doc/MLprac2_2.pdf

Data availability

All data are available from TCGA

The dataset list(Case ID)

1.The training data and testing data are from TCGA, 40 datasets in total. Training datasets : testing datasets = 7:3

2.When we download datasets from TCGA, we can add the htseq file into cart, and download them as a comprised file.

Milestone1

Progress

According to the Vignette, I've input the data and converted the data to be right data frames which are ready to do MLSeq. And the next step is to choose a model, do the Normalization and transformation, and use the normalized data to train model.

Milestone2

1. What I've done

1.I use 28 datasets to training all the model offered by MLSeq, and use 12 dataset to test the model.

2.The best model I have now is "voomNSC", which has about 75% accuracy.

The comparison between prediction and actual class:

\	Actual	Actual
Predicted	iia	i
i	3	5
iia	2	2

Below is the table of accuracy of the four models

Model	Accuracy
voomNSC	75%
plda	50%
plda2	46.43%
svm	53.57%
NSC	51.85
nblda	46.43%

3.The possible biomarkers I found is:

ensembl_gene_id_version	entrezgene_id	description	hgnc_symbol
ENSG00000124107.5	6590	secretory leukocyte peptidase inhibitor [Source:HGNC Symbol;Acc:HGNC:11092]	SLPI
ENSG00000198888.2	4535	mitochondrially encoded NADH:ubiquinone oxidoreductase core subunit 1 [Source:HGNC Symbol;Acc:HGNC:7455]	MT-ND1

*This is selected possible biomarkers are from inner_join of voomNSC,plda, and plda2 model, there is no possible biomarkers selected from NSC model. However, if NSC model has slected possible biomarkers, we should also do the inner_join.

2. What issues I found

1.In the vignette, I found the "voomDLDA" method, which has errors for now I can't solve, so I just skip this method.

2.Another thing is the model accuracy is low, maybe I need more dataset.

3.For the HTML, I still can't make the HTML, the issue is from my input data function, I input my data once a time and print the table, so the HTML will contain so much table, so HTML does not work well. The HTML file is more than 100Mb, so I did not put it up.

Repeatability

1.I use set.seed function, so you can generate the same results as mine.

2.I upload all the data so that you can reproduce my project.

Deliverable

1.I use R MarkDown to display an understandable codes.

2.I tried to render and upload the HTML file, but there is some issue as mentioned in issue part.