Run Seurat (a R Package) in a notebook interface on a server without root

tl;dr: Dependencies hurts when you try to set up R on server

After a while of playing around, I’ll say the best way to use R with a Notebook-style interface on a server where you are no superuser would be using Anaconda, and then run R inside Anaconda to get whatever package you need. It is designed to run for a normal user, so there’s no need for superuser permission, and dependency issue is also taken care of most of the time.
If you get the permission on server and are comfortable with R Studio, R Studio Server might be a even better option. That is because even if the packages are in Conda R, Conda Forge, or Bioconda, sometimes they are out of date, which might break dependencies, and conda skeleton and conda build unfortunately come with a lot of dependency issue as well. Besides, the visualization of environment and integration of documentation is better in R Studio for me.
What worked for me starts from creating a new environment in Anaconda and install the r-essentials, which should provide a platform ready for jupyter notebook.

conda create --name R
conda install -c r r-essentials

One thing to be noted is that sometimes the packages on Conda are not up-to-date, and that would create some version issue for the R packages, so you might check on different channels to see if there’s something you need with version number in mind. It might be simpler to just use install.packages() in the R environment directly though.
During my package installation, one error message popped out: ImportError: /[Conda path]/env/libgfortran.so.4, and it seemed to be libgfortran not installed on my system. I did conda install libgfortran gfortran libgcc gcc, hoping to get the libraries ready, but it did not fix the error. It turned out I needed to export $LD_LIBRARY_PATH=[Conda lib path for the virtual environment], or R seemed not to know where to find the libraries installed by Conda. After that, I am finally good to go with Jupyter Notebook running R.
The takehome message for me is that it could be complicated to install required libraries to user directory, and the dependencies for these are often too long to be manually managed. I found out I needed cmake to install something, but when I tried to install cmake from source, it warned a compiler supporting C++11 was not found, and then I tried to get new version of gcc but the configure file of cmake still could not find that copy of gcc… I spent almost 3 days in this maze of dependencies, and though I felt I learned a lot, I was not able to fix things this way.
Anyway, I guess that why we need package management. You could find other unsuccessful attempts below if you are interersted.

Attempt 1: Setting up Rkernal with IPython Notebook in Anaconda

Anaconda keeps an independent copy of R when it is installed with conda install -c r r-essentials. I wanted to install the latest version instead of the conda version of Seurat, so I googled for how to install packages from CRAN in Conda and find this article. According to the thread, I tried to do something like install.packages("Seurat", lib = [conda R path]) from my user copy of R.
There is actually a great reason against this approach — install.packages() decides whether to install dependencies based on the installed ones, so if you run install.packages() from another copy of R, the dependency would be a mess.
I tried to use the Conda copy of R, and do install.packages("Seurat"), and found out…

ERROR: dependencies ‘igraph’, ‘diffusionMap’ are not available for package ‘Seurat’
* removing ‘[My Conda path]/R/library/Seurat’

It turned out igraph was not compiled successfully and seems to be a result of system library dependency issue according to this thread, and I would need libssl-dev, libcurl4-openssl-dev, and libssh2-1-dev, and I gave up here because I saw a labyrinth of dependencies here.

Attempt 2: Install Rstudio Server without root

If I could do it, I would be able to use the user copy of R, which I am okay with, and I like R Studio quite a lot. Long story short. Installation is possible, but execution is not. Basically you could follow the official instruction, and finish with make install prefix=[somewhere you can write]. Nonetheless, starting a server unsurprisingly requires you to be root.
One additional hilarious thing here is that we did not have cmake on that server, and when I tried to install a copy, it failed because no C++ compiler supporting C++11 is available (though which g++ showed a nice one.) Until I am done through all these, I noticed that there was already a copy of rstudio-server on the machine I used.

Attempt 3: Make Jupyter Notebook use my user copy of R

I entered R console, and tried devtools::install_github("IRkernel/IRkernel"). Guess what, it failed with an error message of: Installation failed: An unknown option was passed in to libcurl.
Alas. A bit troubleshooting indicated that I might have some version issue for libcurl, but I don’t know which.
Alternatively, I tried to use install.packages("JuniperKernel"), which should serve the same function as IRkernel, but it also has its dependency issue: it requires gdtools, which won’t compile without cairo. cairo requires libpng and pixman. I tried to install from source of libpng and pixman with ./configure && make && make install prefix=$HOME/.local/bin, butpixman kept complaining about unable to find png.h. I was not capable of fixing this and turned to other options, but I am still curious how could I fix this.

Attempt 4: Fine, I’ll just go with the Conda package

Now it seems that my hope to use Seurat on CRAN is not that wise, and my dataset is in fact compatible with Seurat v2.2.0 on bioconda. Perhaps this would be a much more efficient work-around.
Since it is in the bioconda repository, I need to set the channel first.

conda config --add channels defaults
conda config --add channels conda-forge
conda config --add channels bioconda

and then conda install -c bioconda r-seurat.
What I got was

UnsatisfiableError: The following specifications were found to be in conflict:
- r-reprex
- r-seurat

Checking dependency with conda info r-seurat showed that it requires ver 3.4.1 of r-base, while r-reprex need >3.4.3. This prevented me from down-versioning r-base.
Instead, I built Seurat from CRAN with conda skeleton cran Seurat. Interestingly, it saved the recipe in [pwd]/r-seurat instead of [pwd]/Seurat. With conda build r-seurat, guess what I got?

Error : object ‘map_dfr’ is not exported by 'namespace:purrr'
ERROR: lazy loading failed for package ‘Seurat’
subprocess.CalledProcessError: Command '['/bin/bash', '-e', '[pwd]/anaconda/conda-bld/r-seurat_1525165846607/work/conda_build.sh']' returned non-zero exit status 1.

It turned out Seurat required purrr::map_dfr(), which was introduced in newer version than what is on bioconda. A new version of purrr from Conda-forge fixed this, but then there was compilaton error mentioned earlier in this blog.