This repository contains scripts for training a generative long short-term memory recurrent neural network on peptide
sequences. The user can provide sets of amino acid sequences to train the model, and finally invoke sampling of
sequences that should be similar to the training data. As such, artificial intelligence is put in charge of de novo design of new
peptide sequences. The code in this repository relies on the keras
package by Chollet and others
(https://github.com/fchollet/keras) with tensorflow
backend (http://tensorflow.org) as well as on
sklearn
(http://scikit-learn.org) and modlamp
(https://modlamp.org).
- README.md: this file
- LSTM_peptides.py: contains the main code in the following two classes:
SequenceHandler
: class that is used for reading amino acid sequences and translating them into a one-hot vector encoding.Model
: class that generates and trains the model, can perform cross-validation and plot training and validation loss.
- requirements.txt: requirements / package dependencies
- LICENSE: MIT opensource license
Clone the directory to your computer by:
git clone https://github.com/alexarnimueller/LSTM_peptides
Then, install all requirements (in requirements.txt
). In this folder, type:
pip install -r requirements.txt
Finally run the model as follows (with your own parameters provided, see the list below):
python LSTM_peptides.py --dataset $TRAINING_DATA_FILE --name $YOUR_RUN_NAME $FURTHER_OPTIONAL_PARAMETERS
dataset
(default=training_sequences_noC.csv
)- file containing training data with one sequence per line
name
(default=test
)- run name for all generated data; a new directory will be created with this name
batch_size
(OPTIONAL, default=128
)- Batch size to use by the model.
epochs
(OPTIONAL, default=50
)- Number of epochs to train the model.
layers
(OPTIONAL, default=2
)- Number of LSTM layers in the model.
neurons
(OPTIONAL, default=256
)- Number of units per LSTM layer.
cell
(OPTIONAL, default=LSTM
)- type of neuron to use, available:
LSTM
,GRU
- type of neuron to use, available:
dropout
(OPTIONAL, default=0.1
)- Fraction of dropout to apply to the network. This scales with depth, so layer1 gets 1*dropout, Layer2 2*dropout etc.
mode
(OPTIONAL, choices=[pretrain. finetune, sample], default=pretrain)- Whether to pre-train (
pretrain
), fine-tune (finetune
) or just sample from a pre-trained model (sample
).
- Whether to pre-train (
valsplit
(OPTIONAL, default=0.2
)- Fraction of the data to use for model validation. If 0, no validation is performed.
sample
(OPTIONAL, default=100
)- Number of sequences to sample from the model after training.
temp
(OPTIONAL, default=1.25
)- Temperature to use for sampling.
maxlen
(OPTIONAL, default=0
)- Maximum sequence length allowed when sampling new sequences. If 0, the longest sequence length of the training
data is
maxlen
- Maximum sequence length allowed when sampling new sequences. If 0, the longest sequence length of the training
data is
startchar
(OPTIONAL, default=B
)lr
(OPTIONAL, default=0.01
)- Learning rate to be used for Adam optimizer.
modfile
(OPTIONAL, default=None
)- If
train=False
, a pre-trained model file needs to be provided, e.g.modfile=./checkpoint/model_epoch_49.hdf5
.
- If
cv
(OPTIONAL, default=None
)- Folds of cross-validation to use for model validation. If
None
, no cross-validation is performed.
- Folds of cross-validation to use for model validation. If
window
(OPTIONAL, default=0
)- Size of window to use for enhancing training data by sliding-windows. If 0, all sequences are padded to the length of the longest sequence in the data set.
step
(OPTIONAL, default=1
)- Step size to move the sliding window or the prediction target
target
(OPTIONAL, default=all
)- whether to learn all proceeding characters or just the last the single next one in sequence
target
(OPTIONAL, default='all')- whether to learn all proceeding characters or just the last 'one' in sequence
padlen
(OPTIONAL, default=0)- number of tailing padding spaces to add to the sequences. If 0, sequences are padded to the length of the longest sequence in the dataset.
refs
, (OPTIONAL, default=True
- whether reference sequence sets should be generated for the analysis
python LSTM_peptides.py --mode pretrain --name train100 --dataset new_sequences.csv --layers 2 --neurons 64 --epochs 100
python LSTM_peptides.py --mode sample --name testsample --modfile pretrained_model/checkpoint/model_epoch_99.hdf5 --sample 100
python LSTM_peptides.py --mode finetune --name finetune10 --dataset finetune_set.csv --modfile pretrained_model/checkpoint/model_epoch_99.hdf5 --epochs 10
When using this code for any publication, please cite the following article:
A. T. Müller, J. A. Hiss, G. Schneider, "Recurrent Neural Network Model for Constructive Peptide Design" J. Chem. Inf . Model. 2018, DOI: 10.1021/acs.jcim.7b00414.