Peptide retention time prediction using deep learning
- The performance of models trained using more than 100,000 peptides: Figure 1a; Supplementary Figure 1.
- The performance of models trained using 700~10,000 peptides with transfer learning strategy: Figure 1b; Supplementary Figure 3; Supplementary Figure 8-10;
- The performance of models trained using label free, iTRAQ or TMT data: Figure 1a; Figure 1b; Supplementary Figure 3; Supplementary Figure 8-10.
- The performance of models trained using immunopeptidomics data: Supplementary Figure 6;
- The performance of models trained using public iRT data: Figure 1a; Supplementary Figure 1.
$ git clone https://github.com/bzhanglab/AutoRTAutoRT is a python3 package. TensorFlow (>=2.6) is supported. Its dependencies can be installed via
$ pip install -r requirements.txtAutoRT has been tested on both Linux and Windows systems. It supports training and prediction on both CPU and GPU, but GPU is recommended for model training. Multiple GPUs are also supported.
AutoRT can also be used through docker:
docker pull proteomics/autortAutoRT supports training models from scratch as well as transfer learning. We recommend to train models using GPU.
$ python autort.py train -h
optional arguments:
-h, --help show this help message and exit
-i INPUT, --input INPUT
Input data for training
-o OUT_DIR, --out_dir OUT_DIR
Output directory
-e EPOCHS, --epochs EPOCHS
The number of epochs, default is 40.
-b BATCH_SIZE, --batch_size BATCH_SIZE
Batch size for training, default is 64.
-r2 MAX_RT, --max_rt MAX_RT
The maximum retention time. If the value is 0 (default), the maximum retention time will be automatically infered from the input training data.
-u UNIT, --unit UNIT The unit of retention time in training data, s: second, m: minute (default).
-sm SCALE_METHOD, --scale_method SCALE_METHOD
Scaling method for RT tranformation: min_max (default), mean_std and single_factor. This is used in training. Default is 'min_max'. The default method works well in most of cases. This should not be
changed unless users know well about the meaning of these methods.
-sf SCALE_FACTOR, --scale_factor SCALE_FACTOR
This is only useful when 'single_factor' is set for '-sm'.
-m MODEL_FILE, --model_file MODEL_FILE
Trained model file. Only useful when perform transfer learning and RT prediction.
-r, --add_reverse Add reversed peptide in peptide encoding. This parameter will be removed in a future version.
-n EARLY_STOP_PATIENCE, --early_stop_patience EARLY_STOP_PATIENCE
Number of epochs with no improvement after which training will be stopped.
-rlr, --add_ReduceLROnPlateau
Reduce learning rate when a metric has stopped improving.
-g GPU, --gpu GPU Set gpu IDs that can be used. For example: 1,2,3.
-d, --do_evaluation_after_each_epoch
Do evaluation after each epoch during model training.
-f OUTLIER_RATIO, --outlier_ratio OUTLIER_RATIO
The percentage of data (outlier) will be removed from the training data using a two-step training.
$ python autort.py predict -h
usage: autort.py [-h] [-t TEST] [-o OUT_DIR] [-p PREFIX] [-s ENSEMBLE] [-a]
AutoRT
optional arguments:
-h, --help show this help message and exit
-t TEST, --test TEST Input data for testing
-o OUT_DIR, --out_dir OUT_DIR
Output directory
-p PREFIX, --prefix PREFIX
-s ENSEMBLE, --ensemble ENSEMBLE
-a, --radamAn example training data is available in the example/data folder as shown below. The training data should have at least 2 columns x and y in a tab-delimited text format. The x column contains peptide sequences and y column contains retention time. The unit of retention time can be minute or second. If the unit of retention time is minute, users should set "-u m" and if the unit of retention time is second, users should set "-u s".
$ head example/data/PXD006109_Cerebellum_rt_add_mox_all_rt_range_3_train.tsv
x y
AAAAAAAAAAAAAAAGAAGK 58.056936170212765
AAAAAAAAAK 4.394
AAAAAAAAATEQQGSNGPVK 26.975
AAAAAAAAQ1HTK 8.2107
AAAAAAAAQMHTK 15.553
AAAAAASLLANGHDLAAAMAVDK 73.801
AAAAADLANR 18.85030769230769
AAAAAGAGLK 13.3316
AAAAALSGSPPQTEK 27.125375000000002For a peptide with modification(s) (modified peptide), it must be formatted as described below:
Each of the modified amino acids must be represented using a different integer number (0-9) in the peptide sequence and keep consistent in both training and testing (or prediction) data. For example, if there are four modifications in a dataset: oxidation on M, phosphorylation on S,T and Y, then we can use 1 represents oxidation of M, 2 represents phosphorylation of S, 3 represents phosphorylation of T and 4 represents phosphorylation of Y.
Then a peptide like this:
AS(phosphorylation)FDFADFST(phosphorylation)PY(phosphorylation)STM(oxidation)AAK
must be formatted as
A2FDFADFS3P4ST1AAK
Then the setting for parameter -p looks like this -p 1,2,3,4.
An example testing data is available in the data folder as shown below. The first column "x" is required. The second column "y" is optional.
$ head example/data/PXD006109_Cerebellum_rt_add_mox_all_rt_range_3_test.tsv
x y
DSQFLAPDVTSTQVNTVVSGALDR 84.633
LLDLYASGER 51.07025
EMPQNVAK 16.787
GSPTGSSPNNASELSLASLTEK 62.60036363636364
LGLDEYLDK 57.118833333333335
FNGAQVNPK 20.01075
AVTISGTPDAIIQCVK 66.29124999999999
SSPVEYEFFWGPR 81.73833333333333
AIPSYSHLR 29.95825
An example to show how to train a highly accurate RT model using a small dataset from a global proteome experiment:
Please note that the length of the longest peptide in the training data must be <= 60 which is the maximum length supported by the based models in the github folder models/general_base_model/.
$ cd example
$ cat transfer_learning.sh
## training
python ../autort.py train -i data/28CPTAC_COprospective_W_VU_20150810_05CO037_f01_normal_train.tsv -o tf_model/ -e 40 -b 64 -u m -m ../models/general_base_model/model.json -rlr -n 10
## prediction
python ../autort.py predict -t data/28CPTAC_COprospective_W_VU_20150810_05CO037_f01_normal_test.tsv -s tf_model/model.json -o tf_prediction/ -p test
The training took less than 20 minutes using one Titan Xp GPU on a Linux server.
The parameter setting for -e, -b, -sm, -rlr and -n for the above examples worked well for many data.
cd example
## training
python ../autort.py train -e 100 -b 64 -g ../models/base_model/model.json -u m -p 1 -i data/PXD006109_Cerebellum_rt_add_mox_all_rt_range_3_train.tsv -sm min_max -l 48 -rlr -n 20 -o PXD006109_models/After the training is finished, the trained model files are saved in the folder PXD006109_models/.
Then we can use the trained models for prediction as shown below:
## prediction:
python ../autort.py predict -t data/PXD006109_Cerebellum_rt_add_mox_all_rt_range_3_test.tsv -s PXD006109_models/model.json -o PXD006109_prediction/ -p test
The prediction result will be saved in file PXD006109_prediction/test.csv and looks like below. The column y_pred contains the predicted retention time.
$ head PXD006109_prediction/test.csv
x y y_pred
DSQFLAPDVTSTQVNTVVSGALDR 84.633 87.11121
LLDLYASGER 51.07025 51.25605
EMPQNVAK 16.787 17.024113
GSPTGSSPNNASELSLASLTEK 62.60036363636364 61.83924
LGLDEYLDK 57.11883333333334 57.66608
FNGAQVNPK 20.01075 19.809322
AVTISGTPDAIIQCVK 66.29124999999999 68.5102
SSPVEYEFFWGPR 81.73833333333333 82.20954
AIPSYSHLR 29.95825 31.344095
The training took less than 12 hours using one Titan Xp GPU on a Linux server.
Example: example/Experiment_specific_RT_prediction_using_AutoRT.ipynb
Example: example/Phosphorylation_experiment_specific_RT_prediction_using_AutoRT_Colab.ipynb
Wen, B., Li, K., Zhang, Y. et al. Cancer neoantigen prioritization through sensitive and reliable proteogenomics analysis. Nature Communications 11, 1759 (2020). https://doi.org/10.1038/s41467-020-15456-w