A self-supervised workflow for particle picking in cryo-EM

Authors: Donal McSweeney donal.mcs15@gmail.com; Qun Liu qun.liu@gmail.com

Referecne

https://doi.org/10.1107/S2052252520007241

System requirement

1. Anaconda2 with cuda/GPU/tensorflow/keras support
2. Relion2 or Relion3.0x (not 3.1) compiled with GPU and MPI support

Quickstart

Download the Self-Supervised to a directory under the relion project structure as it uses relion for 2D class averaging. All commands should run from the root directory of Relion.

Give an alias name for the directory containing motion corrected micrographs

ln -s /path/to/motition-corrected/micrographs aligned

Do ctf correction, select ctf corrected micrographs for iterative training and pikcing, rename its star file as below

mv -f CtfFind/whatever_ctf.star CtfFind/micrographs_defocus_ctf.star

Pre-processing. Pikcked particles will be under local/aligned/. Ideally, the script reads in defocus values and use them to setup defocus-based picking.

1. Configure settings in Self-Supervised/config.ini
2. Run python Self-Supervised/build_preprocess.py. This will generate a bash script for initial (local) picking and 2D class averaging.
3. Run sh preprocess.sh By default, the workflow uses 4 GPUs for 2D class average. Modify preprocess.sh if you use a different number of GPUs.

Self-Supervised training/particle picking. Pikced particles will be under kpicker/aligned/

1. Configure settings in Self-Supervised/config.ini
2. Run python Self-Supervised/build_workflow.py. This will generate a bash file for iterative training/picking.
3. Run sh workflow.sh By default, the workflow uses 4 GPUs for 2D class average. Modify workflow.sh if you use a different number of GPUs.

Picking all micrigraphs: Create an empty directory of Kpicker/aligned if it does not exist, then run

python Self-Supervised/kpicking_cpu.py --input_dir 'aligned' --output_dir 'Kpicker/aligned' --coordinate_suffix '_kpicker' --threshold 0.9 --threads 10 --particle_size 260 --bin_size 4

using Localpicker for initial particle picking based on shapes

for mrcfile in (ls */*.mrc); do ; python -W ignore Self-Supervised/localpicker.py --mrc_file=${mrcfile} --particle_size=260 --bin_size=9 --threshold=0.0015 --max_sigma=10; done