Overview

REliable PIcking by Consensus (REPIC) is an ensemble learning approach to cryogenic-electron microscopy (cryo-EM) particle picking. It identifies particles common to multiple picked particle sets (i.e., consensus particles) using graph theory and integer linear programming (ILP). Picked particle sets may be found by a human specialist (manual), template matching, mathematical function (e.g., RELION's Laplacian-of-Gaussian auto-picking), or machine-learning method. A schematic representation of REPIC applied to the output of three CNN-based particle pickers is below:

REPIC expects particle sets to be in BOX file format (*.box) where each particle has coordinates, a detection box size (in pixels), and (optional) a score [0-1].

Software requirements

Required:

1. Python v3.8 interpreter (Miniconda installation recommended)
2. Python package dependencies described in setup.py
3. Windows users - Ubuntu terminal environment with Windows Subsystem for Linux (WSL) (v22.04.2 LTS tested)

Optional:

1. Gurobi ILP optimizer (v9.5.2 used) - requires free academic license **
2. REgularised LIkelihood OptimisatioN (RELION) - particle and density analyses (v3.13 used)
3. UCSF Chimera - map alignment and density visualization (v1.16 used)

** Required to reproduce manuscript results but if the Gurobi package is not found, REPIC will use the SciPy ILP optimizer

Installation guide

REPIC installation is expected to only take a few minutes.

⚠️ WARNING: Only the Docker installation of REPIC includes pickers (SPHIRE-crYOLO, DeepPicker, and Topaz). If installing REPIC using either Conda or pip, pickers will need to be separately installed (see docs/ for installation instructions).

Install using Docker (recommended)

1. Install Docker if the docker command is unavailable
2. Install and set up NVIDIA Container Toolkit for building and running GPU-accelerated containers
3. Build CUDA-supported image from Dockerfile in REPIC GitHub repo:

docker build -t repic_img https://github.com/ccameron/REPIC.git#main

4. Run container with GPU acceleration (example iter_pick.py command shown below):

docker run --gpus all -v <file_path>/REPIC/examples:/examples repic_img repic iter_pick /examples/10057/iter_config.json 4 100

Install using Conda

⚠️ WARNING: if Python package conflicts are encountered during the Conda installation of REPIC, please ensure Conda channels are properly set for Bioconda. See Bioconda Usage for more information

1. Install Miniconda if the conda command is unavailable
2. Create a separate Conda environment and install REPIC and Gurobi:

conda create -n repic -c bioconda -c gurobi repic gurobi

3. Activate REPIC Conda environment:

conda activate repic

4. Obtain a Gurobi license and set Gurobi key:

grbgetkey <gurobi_key>

5. Remove unused or temporary Conda files:

conda clean --all

Install from source using pip

1. Either download the package by clicking the "Clone or download" button, unzipping file in desired location, and renaming the directory "REPIC" OR using the following command line:

git clone https://github.com/ccameron/REPIC

2. Install Miniconda if the conda command is unavailable
3. Navigate to REPIC directory:

cd <install_path>/REPIC

4. Create a separate Conda environment and install Gurobi for REPIC:

conda create -n repic -c gurobi python=3.8 gurobi

5. Activate REPIC Conda environment:

conda activate repic

6. Install REPIC using pip:

pip install .

7. Obtain a Gurobi license and set Gurobi key:

grbgetkey <gurobi_key>

8. Remove unused or temporary Conda files:

conda clean --all

To check if REPIC was correctly installed, run the following command (after activating the REPIC Conda environment or using a REPIC container):

repic -h

A help menu should appear in the terminal.

Integration

Run using published Docker image (with Singularity/Apptainer)

A REPIC Docker image is published on both DockerHub and the GitHub container registery. Apptainer (formerly Singularity) can be used to run this image:

1. Install Apptainer if the apptainer command is unavailable
2. Install and set up NVIDIA Container Toolkit for building and running GPU-accelerated containers
3. Pull REPIC Docker image and convert to Singularity image format (SIF) (requires >8 Gb of memory and ~40 mins for conversion):

apptainer pull docker://cjfcameron/repic:main

If SIF file creation is taking a long time, increase the mksquashfs mem parameter in the Apptainer config file (apptainer.conf). See here for more information.

4. Run container with GPU acceleration (example iter_pick.py command shown below):

apptainer run --nv --bind <file_path>/REPIC/examples:/examples repic_main.sif repic iter_pick /examples/10057/iter_config.json 4 100

Run using Scipion plugin

REPIC is available as a Scipion plugin: https://github.com/scipion-em/scipion-em-repic

See here for information about installing plugins for Scipion.

Example data

Example SPHIRE-crYOLO, DeepPicker, and Topaz picked particle coordinate files for $\beta$-galactosidase (EMPIAR-10017) micrographs are found in examples/10017/. These files were generated by applying the pre-trained pickers to $\beta$-galactosidase micrographs, filtering false positive per author suggested thresholds, and then converting files to BOX format using coord_converter.py.

Example motion corrected T20S proteasome (EMPIAR-10057) micrographs and normative particles for iterative ensemble particle picking are freely available via Amazon Web Services (AWS). To download this data, please run get_examples.sh (see Quick start below for how to run this Bash script).

Installation instructions for SPHIRE-crYOLO, DeepPicker, and Topaz are found in docs/.

Example commands for fitting and running SPHIRE-crYOLO, DeepPicker, and Topaz models are found in repic/iterative_particle_picking/.

Parameters used for particle picking algorithms and RELION are found in supplemental_data_file_2.ods.

Quick start

Creating consensus particle sets

1. Calculate the particle overlap (Jaccard index [JI]) and enumerate cliques using get_cliques.py (expected run time: 1-3 mins):

repic get_cliques examples/10017/ examples/10017/clique_files/ 180

Note - REPIC will use the folder names found in the provided input directory (e.g., examples/10017/) to assign method labels (e.g., "crYOLO", "deepPicker", "topaz")

Correctly executing the above command will produce the following files for each micrograph in the output folder examples/10017/clique_files/:

• *_clique_coords.pick: pickled clique (x,y) coordinates (in BOX format)
• *_constraint_matrix.pickle: pickled constraint matrix file
• *_weight_vector.pickle: pickled clique weight vector file
• *_runtime.tsv: runtime tracking TSV file

2. Find optimal cliques using the ILP solver and create consensus particle BOX files using run_ilp.py (expected run time: <1 min):

repic run_ilp examples/10017/clique_files/ 180

Correctly executing the above command will produce a particle coordinate file (in BOX format) for each micrograph in the output directory examples/10017/clique_files/. The final column in these BOX files represents the clique weight for a consensus particle.

Particle picking by iterative ensemble learning

1. Download example data from AWS S3 bucket using get_examples.sh (expected run time: 1-5 mins):

bash $(pip show repic | grep -in "Location" | cut -f2 -d ' ')/repic/iterative_particle_picking/get_examples.sh examples/10057/data/ &> aws_download.log

2. Create a configuration file for iterative ensemble particle picking using iter_config.py (expected run time: <1 min):

repic iter_config examples/10057/ 176 224 <file_path>/gmodel_phosnet_202005_N63_c17.h5 <file_path>/DeepPicker-python 4 22

<file_path> must be replaced with the full file paths to the SPHIRE-crYOLO pre-trained model and DeepPicker directory, respectively. See picker installation instructions in docs/ for more information.

A configuration file iter_config.json will be created in the current working directory.

3. Pick particles by iterative ensemble learning using a Python script wrapper iter_pick.py of run.sh (expected run time: 20-30 min/iteration):

repic iter_pick ./iter_config.json 4 100

The final set of consensus particles for the testing set should be found in: examples/10057/iterative_particle_picking/round_4/train_100/clique_files/test/*.box

Command line details

Identifying consensus particle sets with REPIC

1. Calculating particle overlap (JI) and enumerate cliques using get_cliques.py:

usage: repic get_cliques [-h] [--multi_out] [--get_cc] in_dir out_dir box_size

positional arguments:
  in_dir       path to input directory containing subdirectories of particle bounding box coordinate files
  out_dir      path to output directory (WARNING - script will delete directory if it exists)
  box_size     particle bounding box size (in int[pixels])

optional arguments:
  -h, --help   show this help message and exit
  --multi_out  set output of cliques to be members sorted by picker name
  --get_cc     filters cliques for those in the largest Connected Component (CC)

2. Finding optimal cliques using ILP solver and creating consensus particle BOX files using run_ilp.py:

usage: repic run_ilp [-h] [--num_particles NUM_PARTICLES] in_dir box_size

positional arguments:
  in_dir                path to input directory containing get_cliques.py output
  box_size              particle bounding box size (in int[pixels])

optional arguments:
  -h, --help            show this help message and exit
  --num_particles NUM_PARTICLES
                        filter for the number of expected particles (int)

run_ilp.py will create a plot of the concensus particle distribution (particle_dist.png) with a recommended (REC) num_particles value (70% of consensus particles) in the in_dir.

Particle picking by iterative ensemble learning

1. Create a configuration file for iterative ensemble particle picking using iter_config.py:

usage: repic iter_config [-h] [--cryolo_env CRYOLO_ENV] [--deep_env DEEP_ENV] [--topaz_env TOPAZ_ENV] [--out_file_path OUT_FILE_PATH]
                         data_dir box_size exp_particles cryolo_model deep_dir topaz_scale topaz_rad

positional arguments:
  data_dir              path to directory containing training data
  box_size              particle bounding box size (in int[pixels])
  exp_particles         number of expected particles (int)
  cryolo_model          path to LOWPASS SPHIRE-crYOLO model
  deep_dir              path to DeepPicker scripts
  topaz_scale           Topaz scale value (int)
  topaz_rad             Topaz particle radius size (in int[pixels])

optional arguments:
  -h, --help            show this help message and exit
  --cryolo_env CRYOLO_ENV
                        Conda environment name or prefix for SPHIRE-crYOLO installation (default:cryolo)
  --deep_env DEEP_ENV   Conda environment name or prefix for DeepPicker installation (default:deep)
  --topaz_env TOPAZ_ENV
                        Conda environment name or prefix for Topaz installation (default:topaz)
  --out_file_path OUT_FILE_PATH
                        path for created config file (default:./iter_config.json)

data_dir/ is expected to contain a three-column TSV file of CTFFIND4 defocus values: (1) micrograph filename, (2) defocus x, and (3) defocus y. If this file is not found, then all micrographs will be assigned the same defocus value. A defocus file can be built from the output of a RELION CTF refinement job using the following Bash script:

EMPIAR_ID=<complete>  # only integers - i.e., EMPIAR-10017 would be 10017
out=<install_path>/REPIC/examples/${EMPIAR_ID}/data/defocus_${EMPIAR_ID}.txt
rm -rf ${out}
for file in <relion_path>/relion/CtfFind/job00[0-9]/<mrc_pattern>; do
  grep '' /dev/null ${file} | tail -n 1 | awk -F ":| " '{print $1,$3,$4}' >> ${out}
done

<mrc_pattern> is dependent on the naming convention used for micrographs and will need to be set to your specific substring. For EMPIAR-10017 and -10057, the substrings are '*0.txt' and '*[0-9].txt', respectively.

<relion_path>/relion/CtfFind/job00[0-9]/*<mrc_suffix> should list all CTFFIND4 output files in RELION's CtfFind/.

2. Iteratively pick particles using a Python script wrapper iter_pick.py of run.sh:

usage: repic iter_pick [-h] [--semi_auto] [--score] [--out_file_path OUT_FILE_PATH] config_file num_iter train_size

positional arguments:
  config_file           path to REPIC config file
  num_iter              number of iterations (int)
  train_size            training subset percentage (int)

optional arguments:
  -h, --help            show this help message and exit
  --semi_auto           initialize training labels with known particles (semi-automatic)
  --score               evaluate picked particle sets
  --out_file_path OUT_FILE_PATH
                        path for picking log file (default:<data_dir>/iter_pick.log)

train_size references the output of build_subsets.py, which builds training subsets of sizes 1%, 25%, 50%, and 100% (i.e., 100% will use the entire training set). For more information on dataset handling please see "iterative ensemble particle picking with REPIC" in the Methods section of the REPIC manuscript.

Testing

The REPIC software has been tested on two computer systems:

1. Ubuntu 16.04.6 LTS (Xenial Xerus) running CUDA v10.1 with four Nvidia GP102 TITAN Xp
2. Ubuntu 16.04.7 LTS (Xenial Xerus) running CUDA v11.3 with four Nvidia GeForce GTX 1080

Citing REPIC

If REPIC was used in your analysis/study, please cite:

Cameron, C.J.F., Seager, S.J.H., Sigworth, F.J., Tagare, H.D., and Gerstein, M.B. REPIC - an ensemble learning methodology for cryo-EM particle picking. BioRxiv. DOI: 10.1101/2023.05.13.540636v1

Contact

Submitting a GitHub issue is preferred for all problems related to REPIC.

For other concerns, please email Christopher JF Cameron.

Releases

v0.2.1

• Scipion plugin created
• Docker/Singularity/Apptainer integrated

v0.2.0

• k-d tree algorithm integrated to reduce graph building runtime
• Approval to include DeepPicker with REPIC install/distribution added: https://github.com/ccameron/REPIC/blob/main/imgs/deeppicker_approval.png
• Various bug fixes

v0.1.0

• SciPy ILP optimizer integrated to remove Gurobi package requirement
• Read the Docs documentation created: https://repic.readthedocs.io/en/latest/
• Various bug fixes