We ported SSRI to Relion 3.0! Check out https://github.com/alncat/ssri!!!
Our latest version can do CTF refinement with 3D refinement simultaneously!!! checkout CTF refinement
You can compile this program with fftw in external directory!
First, modify the CMakeLists.txt file, replace the cuda architecture related variables to make it compatible with your GPU! and set the MPI_LIBRARIES, MPI_INCLUDE_PATH to the mpi related paths on your system! Note that the default setting of cuda architecture is sm_61. To execute this program on Nvidia GPUs with other CUDA architectures, you should change all sm_61 and compute_61 variables in /root/gpu/ssri_remote/CMakeLists.txt to the corresponding architecture of your GPUs.
We recommend using cmake >= 3.14, gnu8 and cuda 10.1 !
mkdir build & cd build
FFTW_LIB=/work/jpma/luo/ssri/ssri_remote/external/fftw/lib FFTW_INCLUDE=/work/jpma/luo/ssri/ssri_remote/external/fftw/include/ CC=`which mpicc` CXX=`which mpic++` MPI_HOME=/opt/ohpc/pub/apps/openmpi/4.0.2_cuda10.1/ cmake -DCMAKE_INSTALL_PREFIX=/work/jpma/luo/ssri/ssri_remote/build/bin/ -DCMAKE_BUILD_TYPE=relwithdebinfo -DGUI=OFF ..
Remember to replace the environment variables with your settings!
Finally,
make
OPUS-SSRI(Sparsity and Smoothness Regularized Imaging) is a stand-alone computer program for Maximum A Posteriori refinement of (multiple) 3D reconstructions in cryo-electron microscopy. It is developed in the research group of Jianpeng Ma in Baylor College of Medicine. This implementation is based on the software RELION.
OPUS-SSRI can be obtained as a docker image by executing command below.
docker pull alncat/opus-ssri:first
Another important step is setting up gpu support for docker. You can follow the instruction at https://www.tensorflow.org/install/docker . We can then create a docker container from the docker image via
sudo docker run --name ssri-test --gpus all -it -v /data:/data alncat/opus-ssri:first bash
After entering the docker containter, the OPUS-SSRI program is located at /relion-luo/build/bin. You can perform the 3D refinement using the program in that directory.
OPUS-SSRI introduces several more options to the 3D refinement program of RELION. OPUS-SSRI solves the 3D reconstruction problem in the maximization step with penalty function of the form,
or in text format,
α|x_j|/(|x_j^i|+ϵ) + β‖∇x_j‖_2/(‖∇x_j^i‖_2+ϵ') + γ‖x_j-x_j^i ‖^2.
The new options are:
| Option | Function |
|---|---|
| --tv | toggle on the OPUS-SSRI 3D refinement protocol |
| --tv_eps | the |
| --tv_epsp | the |
| --tv_alpha | propotional to the |
| --tv_beta | propotional to the |
| --tv_weight | propotional to the |
| --tv_iters | the number of iterations of the optimization algorithm |
| --tv_lr | propotional to the learning rate of the optimization algorithm |
The command for running OPUS-SSRI is shown below,
mpiexec --allow-run-as-root -n 3 /relion-luo/build/bin/relion_refine_mpi --o /output-folder --i particle-stack --ini_high 40 \
--dont_combine_weights_via_disc --pool 4 --ctf --ctf_corrected_ref --iter 25 --particle_diameter 256 \
--flatten_solvent --zero_mask --oversampling 1 --healpix_order 2 --offset_range 5 --offset_step 2 \
--norm --scale --j 8 --gpu 0,1,2,3 --tv_alpha 1.0 --tv_beta 2.0 --tv_eps 0.01 --tv_epsp 0.01 \
--tv_weight 0.1 --tv_lr 0.5 --tv_iters 150 --ref initial-map --free_gpu_memory 256 --auto_refine \
--split_random_halves --low_resol_join_halves 50 --tv --adaptive_fraction 0.94 --preread_images --sym C4
You should replace output-folder, particle-stack and initial-map with your own path and files, respectively. You should also set the particle_diameter and sym according to the parameters of your system. For very large datasets that cannot fit in the memory, you should remove the preread_images option. tv_eps and tv_epsp are two adjustable parameters. You can set them around the level of density as the threshold for creating mask. You can also experiment with different tv_alpha, tv_beta and tv_weight.
We added the capacity to perform per-particle CTF refinement in a new version. You can use this program to perform consensus 3D refinement with per-particle CTF refinement. The program can be obtained as a docker image by
docker pull alncat/ssri-torch:ctf1
To run the CTF refinement, you need to create a docker container using the downloaded image by
sudo docker run --name ssri-ctf --gpus all -it -v /data:/data alncat/ssri-torch:ctf1 bash
--name is the name of container, -v argument is used to bind directory in host to the directory in container, the first option before colon refers to the directory in host, the second option refers to the directory in the container.
The docker container can be accessed later via
sudo docker start ssri-ctf && sudo docker exec -it ssri-ctf bash
The compiled program with CTF refinement capability is located at /root/gpu/ssri_remote/build/bin in your docker container. And the source code can be found at /root/gpu/ssri_remote/src .
Note that the program is compiled against Nvidia GPU with compute capacity sm_61. To execute this program on Nvidia GPUs with other CUDA architectures, you should recompile this program by changing all sm_61 and compute_61 variables in /root/gpu/ssri_remote/CMakeLists.txt to the corresponding architecture of your GPUs. You can then recompile the program by first loading the compilation toolset,
scl enable devtoolset-7 bash
then executing
cd build && make
The options for controlling the CTF refinement are listed below,
| Option | Function |
|---|---|
| --ctf_order | Default is 2, start CTF refinement when the healpix sampling order is larger than this value |
| --refine_ctf_angle | Add this option to your command to refine the defoucs angle |
| --ctf_defocus_dev | Default is 1.0, the restraint strength of the deviation of defocus values from previous refinement |
| --ctf_defocus_iso | Default is 0.5, the restraint strength of the difference between two defocus depths |
An exmaple of the command for running OPUS-SSRI with CTF refinement is shown below,
mpiexec --allow-run-as-root -n 3 /root/gpu/ssri/remote/build/bin/relion_refine_mpi --o /output-folder --i particle-stack --ini_high 40 \
--dont_combine_weights_via_disc --pool 4 --ctf --ctf_corrected_ref --iter 25 --particle_diameter 256 \
--flatten_solvent --zero_mask --oversampling 1 --healpix_order 2 --offset_range 5 --offset_step 2 \
--norm --scale --j 8 --gpu 0,1,2,3 --tv_alpha 1.0 --tv_beta 2.0 --tv_eps 0.01 --tv_epsp 0.01 \
--tv_weight 0.1 --tv_lr 0.5 --tv_iters 150 --ref initial-map --free_gpu_memory 256 --auto_refine \
--split_random_halves --low_resol_join_halves 50 --tv --adaptive_fraction 0.94 --preread_images --sym C4 --sigma2_fudge 0.5 --ctf_defocus_iso 0.5
You should replace output-folder, particle-stack and initial-map with your own path and files, respectively. You should also set the particle_diameter and sym according to the parameters of your system. For very large datasets that cannot fit in the memory, you should remove the preread_images option.