cwalker4 / deepiv-gwas Goto Github PK

• download all files from GTex
• download all files from TCGA (breast cancer only)

• normalize the p matrix (p - np.mean(p, axis=0)) / np.std(p, axis=0)
• try large/original network on AWS
• vary learning rate alpha
• add dropout and/or L2 regularization

I added BibTex compatibility to deepiv_milestone.tex with commit e0751bf6. It's a Tex feature used by researchers to streamline the addition of citations and bibliographies in LaTeX documents, and it saves headaches when it comes to keeping track of citations.

• match TCGA data
• write R scripts to clean data
• put results in derived data
• match GTEx data
• write R scripts to clean GTEx data
• put results in derived data
• match TCGA and GTEx data
• put results in derived data

• get in touch with Quake
• Contact Bologna

Same as in #10

The 230 development pipeline will ultimately (I believe) lead to cleaner code and a nicer development process as we move forward with this project. I'll silo myself in a branch until I'm sure everything is working.

See https://www.gtexportal.org/home/tissueSummaryPage#cause as a start.

It's not like we have coefficients on our inputs that neatly describe their effect on the outcome; do we have some script that fixes all other genes at a baseline, then perturbs a single gene at a time to look at its effect on y?

Go to milestone/writeup/deepiv_milestone.tex and start filling out information.

• Add problem statement, biology background to Introduction (pull from proposal)

• Talk about our candidate dataset

• Talk about how we will be getting our dataset extremely soon

Also, a note on git commit etiquette: please add the issue number relevant to your commit in your commit message, e.g. git commit -m "#1 adding a beautiful and perfect introduction"

Talk to @billyferguson for sketch of what we want; also see project examples on the 230 website, other CS project websites (e.g. CS229, 224W...)

• sample from second stage DGP
• use decision tree or Lasso regression

Some preliminary reading:

https://www.quora.com/How-can-convolutional-neural-networks-be-used-for-non-image-data

https://www.quora.com/Are-convolutional-neural-networks-useful-only-for-image-recognition-or-applications-having-spatial-features

My best guess based on skimming those links: yes, if our input data is something like a full genome where entries that are close together in our vector are also physically close to one another on the chromosome.