cryoSRPNT, a Simulator of Realistic Particles using Noise Terms, allows for generation of arbitrary, realistic cryo-EM datasets derived from known ground truth volumes. Such datasets are useful for developing and benchmarking cryo-EM reconstruction and analysis tools. cryoSRPNT has three scripts designed to be used sequentially:
project3d.py: takes as input a 3D volume and a pose sampling scheme, and produces corresponding noiseless 2D projections of the volume.acn.py: takes as input the 2D projections, global SNR, and CTF parameters per-particle, and produces corresponding noisy + CTF-corrupted particles.write_starfile.py: takes as input 2D projections (and optionally pose + ctf + additional parameters) to write a RELION 3.0 format star file
Both scripts require various components of cryoDRGN for internal calculations and data handling; additionally, the pose and ctf inputs are expected as cryoDRGN-formatted .pkl files. Following cryodrgn installation and cryodrgn conda env activation, clone this repo: git clone https://github.mit.edu/bmp/cryoSRPNT.git
See "Background and options" section below for more details regarding common options for each command. All options are visible via python $script.py --help. Notably, these scripts can additionally take a .txt file as input specifying a tilt series stage tilt scheme (simulating corresponding rotations+projections and tilt-dependent sample thickness weighting, respectively).
python project3d.py volume.mrc noiseless_projections.mrcs [--in-pose poses.pkl | --healpy-grid RESOLUTION | --so3-random NIMGS] --outpose sampled_poses.pklpython acn.py noiseless_projections.mrcs realistic_projections.mrcs [--ctf ctf.pkl | --dfu DEFOCUS_U --dfv DEFOCUSV ...] --snr1 1.4 --snr2 0.05python write_starfile.py realistic_projections.mrcs --Apix APIX [--ctf ctf.pkl --poses sampled_poses.pkl] -o realistic_projections.star
project3d.py offers three possible pose determination schemes: (1) user-provided poses, (2) uniform spherical projection sampling at specified resolutions using healpy, or (3) poses sampled from SO3 (therefore randomly sampled from a uniform distribution of spherical projections). Users may optionally save a pose.pkl file containing the rotations and translations applied when creating the stack.
acn.py follows the sequential (1) structural noise --> (2) CTF --> (3) shot noise cryo-EM image formation model described in Baxter et al, 2009. Both noise processes are modeled as the real-space addition of gaussian noise with independent standard deviations. CTF is modeled following standard cryo-EM procedures. Users may optionally save a ctf.pkl file containing the CTF parameters applied when creating the stack.
write_starfile.py creates a RELION 3.0 format star file minimally containing the paths to each particle in the specified stack (_rlnImageName). If pose.pkl is supplied, the star file will also contain rotation and translation parameters; if ctf.pkl is supplied the star file will also contain CTF parameters.
Both simulation scripts are built using a pytorch/cuda framework to leverage GPU acceleration. cryoSRPNT can simulate 1M particles at 100 x 100 px in ~ 5 minutes on a system equipped as follows: CPU: Intel Xeon Gold 6242R; GPU: Nvidia RTX 3090; RAM: 512 GB.
However, the (Nvidia) GPU is not required, and RAM is predominantly used to hold a single copy of the particle stack in memory with np.float32 precision (4 bytes per pixel). Therefore, the particle stack described above could be simulated on a system with ~48 GB RAM and an older / no GPU.