Several Initial Issues (installation + execution)

[kulmsc]

Apologies that this is not the standard bug report, but after some labor I have finally gotten the code of this repository to run on my own machine and I thought it could be beneficial to report the steps I had to take to accomplish this task.

1. TeraPCA would not compile on a new M1 Mac (or at least it would not after 6 hours of trying). It could be great if the PCs were computed from a program that did not require the whole "make" process.
2. Python 3.6 was required, with numpy between 1.18 and 1.23. Additionally, torchvision and ray were required yet are not listed. A yml that makes a conda environment would really help this process.
3. The input file (specifically for access point 3) must be comma delimited. The "Population IDs" column must be labeled "ancs" and a numeric data type. There are some other confusing statements in the ReadMe (pheno is described twice).
4. I received many errors along the lines of "RuntimeError: mat1 and mat2 shapes cannot be multiplied (32x3 and 8x128)". This appears to be caused by a disconnect in NN layer dimensions. The 8 in this example corresponds to the num_covs + num_pcs +1, I don't know where the 3 comes. This appears to severely limit the possible features of the repo.
5. The results_summary function errored out because of a dictionary key issue. While not a bug, it would be great if there was a more direct way to save the PRSs themselves. I suspect many users want to analyze the PRSs rather than only rely on FairPRS calculated accuracy statistics.

Thank you for making a promising PRS tool. I hope these issues can help further development.

[aritra90]

Thank you @kulmsc for giving FairPRS a try and letting us know about the issues. We are still working on the final version of the code, so your feedback really helps.

1. Sorry that you ran into an issue with Intel MKL on the new MAC M1. I don't know about the issues with Intel MKL on the M1 but we've seen some issues about compatibility. However, TeraPCA works on MAC OS (haven't checked on the M1) and we have tested it before quite robustly. TeraPCA needs Intel MKL library which is available in the Intel oneAPI HPC toolkit which is available here: https://www.intel.com/content/www/us/en/developer/tools/oneapi/hpc-toolkit-download.html?operatingsystem=mac&distributions=webdownload&options=offline
   Once you have downloaded and installed it, you have to set the PATH environments with source . ~/intel/oneapi/setvars.sh
   Then, you can run the make in TeraPCA and install it. We can try to make it as a conda package in the future but with the specialized Intel MKL features that we use, we need to do the installation from scratch.

2. We will fix this for ease of use.

3. Thanks, we will fix the README too.

4. This is most likely a specific case, and we will make it more generic and explain the input in the comments.

5. This is most likely a bug and we will resolve it soon.

Again, thanks a lot and we will update this thread once we have resolved the issues.

[DiegoMac17]

Thank you @kulmsc for letting us know about the issues, hopefully with the update FairPRS can work well for you.

2. I completely understand, it can be very tricky to find the right versions of the packages to work with one another. I have added a yml file for an easy conda environment creation.
3. Thank you for catching these typos. I have corrected the ReadMe to better describe the input file specifications.
4. If it is possible for you to share your file I can try to debug what might be happening there. What might be causing the error is the formatting of the input file. The format that fairPRS accepts for real data in terms of columns is EID, PRS, Population IDs, Phenotype, Covariates. The model takes columns 3:(num_covs + num_pcs + 3) as input.
5. Could you describe more what the error is? So that I can try replicate the error. I was able to run the code fine with our data. I agree with your point, and I have added an extra specific output file just for the PRSs. Alternatively, all results, data, model configurations per each iteration can be found in the dictionary that framework outputs in the results directory.