Try MR-JIT as a new method for multivariate tissue prediction

[gaow]

This method is motivated by Mendelian randomization

https://www.nature.com/articles/s41588-020-0706-2

Could be a possible application candidate for Alzheimer's data-set. We can discuss its model assumptions when we are at that point.

[gaow]

@nickluo1313 Here is a link to a tutorial for MR-JIT

https://www.dropbox.com/sh/i9elg3m4wav4o5g/AAABdxZbVyBclbfa_1KKVftDa?dl=0

The software and other information are on github. But the tool is not particularly well engineered and written. This happens a lot in computational biology -- a seemingly good method published in a high impact journal but the tools are only usable to the authors. I assessed MR-JIT it does not seem that badly engineered but certainly there will be some rough edges.

What we need your input on is to read through the paper above (it's okay if you don't understand everything) and follow the dropbox tutorial above, and convert it to a Jupyter Notebook filling in enough details such that it is fool-proof to reproduce.

The big picture is that @hsun3163 is developing our own prediction pipeline with our method I outlined to you when we last met. We would like to at least try one additional new method like this one on our data. The new model assumption applied to our data will look nice in a paper and maybe it will also find us something new.

[nickluo1313]

Hi Dr.Wang, do you have a Jupyter notebook example that I can reference on? For example, the article summarized by our colleagues earlier? I am not sure what format would be good for a paper summary.

[gaow]

@nickluo1313 I'm not sure what you meant by this. Reading the paper is for your own understanding of the problem but not the expected outcome. If you look at the dropbox link you'll find a file MR-JTI_manual.docx which shows how to analyze the data with MR-JIT. You will see that document is very poorly written and hard to reproduce as is. The expected deliverable from you is a Jupyter Notebook that has same content as the tutorial ,but better formatted and organized, and actually contains code sections that runs the commands. For example:

https://github.com/statgenetics/statgen-courses/blob/master/notebooks/VAT.ipynb

[nickluo1313]

Thank you Dr.Wang. The statgen notebook is the example I was looking for.

[nickluo1313]

Meeting 03/29
ToDo;

• upload this note to github ticket
  • push original JTI.r and overwrite debuged JTI.r
  • Modify JTI.r code
    • Convert Line 82-170 into a function
    • Add 2 lines:
      • write inputs requirements into function into RDS, i.e: saveRDS(list(expr=expr, geno=geno), "some_file.rds")
      • uni%>%saveRDS("$[_output[1]]")
    • Check what happened in dosage raw file conversion
• Part ii MetaXcan
• For Gao:
  • Look into metaXcan + communicate with the authors to figure out the weight database
  • Review MR-JTI notebook you have and reproduce

[gaow]

Thanks @nickluo1313 this is very good.

Check what happened in dosage raw file conversion

Not sure what this is ... but we do want to check with @hsun3163 to see how this can be integrated to his pipeline. Maybe afer you are done with modifying the code let me know. I'll perhaps spend a day or two to iron out everything before we try a production run.