This is the source code for paper entitled "Full-length Ribosome Density Prediction by a Multi-Input and Multi-Output Model".
Tian T, Li S, Lang P, Zhao D, Zeng J (2021) Full-length ribosome density prediction by a multi-input and multi-output model. PLOS Computational Biology 17(3): e1008842. https://doi.org/10.1371/journal.pcbi.1008842
To reproduce the results in our paper or analyze new ribo-seq dataset using RiboMIMO, you need to follow the steps below:
RiboMIMO is implemented in Python 3, and the requirements are listed as follows:
python 3.7.4
pytorch 1.5.0
sklearn 0.21.3
cuda 10.2
Processed ribosome profiling data is needed for training, and the input file should be in fasta format. Each sample contains three lines, including:
- gene name or notation start with ">";
- list of codons seperated with tab;
- list of counts for ribosome footprints at ribosomal A site.
For example,
>YKL096W-A 93
ATG CAA TTC TCT ACT GTC GCT TCC GTT GCT TTC GTC GCT TTG GCT AAC TTT GTT GCC GCT GAA TCC GCT GCC GCC ATT TCT CAA ATC ACT GAC GGT CAA ATC CAA GCT ACT ACC ACT GCT ACC ACC GAA GCT ACC ACC ACT GCT GCC CCA TCT TCC ACC GTT GAA ACT GTT TCT CCA TCC AGC ACC GAA ACT ATC TCT CAA CAA ACT GAA AAT GGT GCT GCT AAG GCC GCT GTC GGT ATG GGT GCC GGT GCT CTA GCT GCT GCT GCT ATG TTG TTA TAA
0 1995 399 175 224 377 195 219 283 230 965 1188 425 482 305 89 677 531 395 813 190 171 216 339 982 236 150 485 214 1083 1282 2416 1415 859 912 1058 959 532 621 1524 304 158 191 169 199 95 149 235 122 76 50 153 28 82 252 141 213 344 599 1847 343 189 139 84 95 195 475 464 968 656 621 1271 723 422 224 1325 333 59 652 1388 76 236 168 374 686 420 378 592 369 0 0 0 0
The already processed datasets including Subtelny14 (yeast, GSM1289257), Mohammad16 (ecoli, GSE72899), Mohammad19-1 (ecoli, GSM3358140) and Mohammad19-2 (ecoli, GSM3358142) can be found in
data/[DATASET].txt
You can also use customed datset following the same fasta format, name it [DATASET].txt, and put it in the data/ folder.
The RiboMIMO model can be trained and cross-validated with command:
cd src/
python -u runRiboMIMO.py --add_nt --add_aa --dataset [DATASET] --gpu [GPU ID]
For more information, please run help:
python -u runRiboMIMO.py --help
The program will create a working directory and the best models will be stored in the following [MODEL FOLDER]:
results/RiboMIMO_[DATASET]*/
To analyze the trained RiboMIMO model, we use saliency maps to account for the codon contributions, termed codon impact score (CIS).
CIS can be calculated using the trained RiboMIMO model with command:
cd src/
python -u getSaliency.py [GPU ID] [MODEL FOLDER]
The program will create a working directory and the CIS results will be stored in the following [CIS FOLDER]:
results/CIS_[DATASET]*/
Note that the folder name of [CIS FOLDER] follows the same pattern with [MODEL FOLDER].
The code for analyzing the trained models and CIS is provided in jupyter notebook src/analyses.ipynb.
The program will create a working directory and the analysis results will be stored in the following [ANALYSIS FOLDER]:
results/Analysis_[DATASET]*/
The predicted ribosome densities at each codon position of genes in the test sets are provided in the prediction/.
The results are stored in dict format, and can be loaded using the following python command:
import pickle
# load dict
predict_dict = pickle.load(open("./prediction/RiboMIMO_yeast_Subtelny14_pred", "rb"))
# get label and prediction results for gene YOR098C
label, pred = temp['YOR098C']
If you have any questions, please feel free to contact me.