Cryo_READ is a computational tool using deep learning to automatically build full DNA/RNA atomic structure from cryo-EM map.
Copyright (C) 2022 Xiao Wang, Genki Terashi, Daisuke Kihara, and Purdue University.
License: GPL v3. (If you are interested in a different license, for example, for commercial use, please contact us.)
Contact: Daisuke Kihara (dkihara@purdue.edu)
For technical problems or questions, please reach to Xiao Wang (wang3702@purdue.edu).
Xiao Wang, Genki Terashi & Daisuke Kihara. De novo structure modeling for nucleic acids in cryo-EM maps using deep learning. bioArxiv, 2022.
@article{xiao2022CryoREAD,
title={De novo structure modeling for nucleic acids in cryo-EM maps using deep learning},
author={Xiao Wang, Genki Terashi, and Daisuke Kihara},
journal={bioArxiv},
year={2022}
}
Project website: https://kiharalab.org/emsuites
Detailed Instructions can be found https://kiharalab.org/emsuites/cryoread.php
DNA and RNA play fundamental roles in various cellular processes, where the three-dimensional (3D) structure provides critical information to understand molecular mechanisms of their functions. Although an increasing number of structures of nucleic acids and their complexes with proteins are determined by cryogenic electron microscopy (cryo-EM), structure modeling for DNA and RNA is still often challenging particularly when the map is determined at sub-atomic resolution. Moreover, computational methods are sparse for nucleic acid structure modeling.
Here, we developed a deep learning-based fully automated de novo DNA/RNA atomic structure modeling method, CryoREAD. CryoREAD identifies phosphate, sugar, and base positions in a cryo-EM map using deep learning, which are traced and modeled into a 3D structure. When tested on cryo-EM maps determined at 2.0 to 5.0 Å resolution, CryoREAD built substantially accurate models than existing methods. We have further applied the method on cryo-EM maps of biomolecular complexes in SARS-CoV-2.
- Structure Detection by deep neural network Cryo-READ networks;
- Tracing backbone according to detections;
- Fragment-based nucleotide assignment;
- Full atomic structure modeling.
Python 3 : https://www.python.org/downloads/
Phenix: https://phenix-online.org/documentation/install-setup-run.html
Coot: https://www2.mrc-lmb.cam.ac.uk/personal/pemsley/coot/
Pymol (for map visualization): https://pymol.org/2/
Chimera (for map visualization): https://www.cgl.ucsf.edu/chimera/download.html
1. Install git
git clone git@github.com:kiharalab/CryoREAD.git && cd CryoREAD
You have two options to install dependency on your computer:
3.1.1install pip.
pip3 install -r requirements.txt --user
If you encounter any errors, you can install each library one by one:
pip3 install biopython
pip3 install numpy
pip3 install numba
pip3 install scipy
pip3 install ortools
pip3 install mrcfile
pip3 install torch==1.6.0
3.2.1 install conda.
conda create -n CryoREAD python
conda activate CryoREAD
pip install -r requirements.txt
Each time when you want to run this software, simply activate the environment by
conda activate CryoREAD
conda deactivate(If you want to exit)
To verify phenix is correctly installed for final refinement step, please run
phenix.real_space_refine -h
If it can print out the help information of this function, then the refinemnt step of our program can be supported. If not, please always remove --refine command line in all the commands, then CryoREAD should output structure without refinement.
usage: main.py [-h] [-F F] [-M M] [-P P] --mode MODE [--contour CONTOUR] [--stride STRIDE] [--box_size BOX_SIZE] [--gpu GPU] [--batch_size BATCH_SIZE] [-f F] [-m M]
[-g G] [-k K] [-R R] [--rule_soft RULE_SOFT] [--frag_size FRAG_SIZE] [--frag_stride FRAG_STRIDE] [--top_select TOP_SELECT] [--resolution RESOLUTION]
[--num_workers NUM_WORKERS] [--prediction_only PREDICTION_ONLY] [--no_seqinfo NO_SEQINFO]
optional arguments:
-h, --help show this help message and exit
-F F Input map file path. (str)
-M M Pre-trained model path. (str) Default value: "best_model"
-P P Optional fasta sequence file path. (str)
--mode MODE Control Mode for program: 0: cryo_READ structure modeling. Required parameter. (Integer), Default value: 0
--contour CONTOUR Contour level for input map, suggested 0.5*[author_contour]. (Float), Default value: 0.0
--stride STRIDE Stride for scanning of deep learning model. (Integer), Default value: 16.
--box_size BOX_SIZE Input box size for deep learning model. (Integer), Default value: 64
--gpu GPU Specify the gpu we will use. (str), Default value: None.
--batch_size BATCH_SIZE
Batch size for inference of network. (Integer), Default value: 4.
-f F Filter for representative points, for LDPs, removing points' normalized density<=-f (Float), Default value: 0.05
-m M After meanshifting merge points distance<[float]. (Float), Default value: 2.0.
-g G Bandwidth of the Gaussian filter, (Float), Default value: 3.0.
-k K Always keep edges where d<k parameter. (Float), Default value: 0.5
-R R Maximum length of local edges. (Float), Default value: 10.0.
--rule_soft RULE_SOFT
Use strict/soft rules to assemble collected fragments in DP step. (Integer), Default value: 0 (strict rules)
--frag_size FRAG_SIZE
Fragment size for sequence split.(Integer), Default value: 20
--frag_stride FRAG_STRIDE
Frag stride step. (Integer), Default value: 2
--top_select TOP_SELECT
Select top fragment candidate here. (Integer), Default value: 20
--resolution RESOLUTION
resolution of maps, used for final structure refinement. (Float), Default value: 2.5
--num_workers NUM_WORKERS
number of workers to fetch data for GPU inference. (Integer), Default value: 4
--prediction_only PREDICTION_ONLY
Optional input. Only run the deep learning prediction step. (True/False) Default value: False
--no_seqinfo NO_SEQINFO
Optional input. Build structures when no sequence information is available. (True/False) Default value: False
python3 main.py --mode=0 -F=[Map_Path] -M=[Model_Path] --contour=[half_contour_level] --gpu=[GPU_ID] --batch_size=[batch_size] --prediction_only
[Map_Path] is the path of the experimental cryo-EM map, [Model_Path] is the path of our pre-trained deep learning model, [half_contour_level] is 0.5* contour_level (suggested by author) to remove outside regions to save processing time, [GPU_ID] specifies the gpu used for inference, [batch_size] is the number of examples per batch in the inference (we used 8 with a 24GB GPU).
The predicted probability maps are saved in [Predict_Result/(map_name)/2nd_stage_detection] with mrc format. It will include 8 mrc files corresponding to 8 different classes.
Example Command:
python3 main.py --mode=0 -F=example/21051.mrc -M=best_model --contour=0.3 --gpu=0 --batch_size=4 --prediction_only
python3 main.py --mode=0 -F=[Map_Path] -M=[Model_Path] --contour=[half_contour_level] --gpu=[GPU_ID] --batch_size=[batch_size] --resolution=[Map_Resolution] --no_seqinfo --refine
[Map_Path] is the path of the experimental cryo-EM map, [Model_Path] is the path of our pre-trained deep learning model, [half_contour_level] is 0.5* contour_level (suggested by author) to remove outside regions to save processing time, [GPU_ID] specifies the gpu used for inference, [batch_size] is the number of examples per batch in the inference (we used 8 with a 24GB GPU), [Map_Resolution] is the resolution of the deposited maps.
"--refine" should be removed if you can not successfully install Phenix correctly.
The automatically build atomic structure is saved in [Predict_Result/(map-name)/Output/Refine_cycle[k].pdb] in pdb format, here default k is 3. However, it may fail if your dependencies are not properly installed, then you may only find Refine_cycle1.pdb or Refine_cycle2.pdb.
Example Command:
python3 main.py --mode=0 -F=example/21051.mrc -M=best_model --contour=0.3 --gpu=0 --batch_size=4 --resolution=3.7 --no_seqinfo --refine
python3 main.py --mode=0 -F=[Map_Path] -M=[Model_Path] -P=[Fasta_Path] --contour=[half_contour_level] --gpu=[GPU_ID] --batch_size=[batch_size] --rule_soft=[assignment_rule] --resolution=[Map_Resolution] --refine
[Map_Path] is the path of the experimental cryo-EM map, [Model_Path] is the path of our pre-trained deep learning model, [Fasta_Path] is the path of the input fasta file about sequence information, [half_contour_level] is 0.5* contour_level (suggested by author) to remove outside regions to save processing time, [GPU_ID] specifies the gpu used for inference, [batch_size] is the number of examples per batch in the inference (we used 8 with a 24GB GPU), [rule_soft] specifies the assignment rule, default is 0 to use the strict assignment assembling rule, [Map_Resolution] is the resolution of the deposited maps.
"--refine" should be removed if you can not successfully install Phenix correctly.
Example Command:
python3 main.py --mode=0 -F=example/21051.mrc -M=best_model -P=example/21051.fasta --contour=0.3 --gpu=0 --batch_size=4 --rule_soft=0 --resolution=3.7 --refine
The automatically build atomic structure is saved in The automatically build atomic structure is saved in [Predict_Result/(map-name)/Output/Refine_cycle[k].pdb] in pdb format, here default k is 3. However, it may fail if your dependencies are not properly installed, then you may only find Refine_cycle1.pdb or Refine_cycle2.pdb. Modeled structures without considering sequence information are also saved as [Predict_Result/(map-name)/Output/CryoREAD_noseq.pdb] (without refinement). Meanwhile, structures only considering the sequence information without connecting gap regions are saved in [Predict_Result/(map-name)/Output/CryoREAD_seqonly.pdb] (without refinement) for reference.
Cryo-EM map with mrc format. (Optional) Sequence information with fasta format. Our example input can be found here
1 *.mrc: a mrc file saved our detected probabilites by our deep learning model.
2 *.pdb: a PDB file that stores the atomic DNA/RNA structure by our method.
Our example output can be found here. All the intermediate results are also kept here.