alexarnimueller/lstm_peptides

Long short-term memory recurrent neural networks for learning peptide and protein sequences to later design new, similar examples.

Hi 👋, I'm Alex

Connect with me:

Languages and Tools:

lstm_peptides's Issues

About the pseudo-random peptide sequences

Hello Alex,
I recently read your work and was very inspired. I found a way to generate pseudo-random peptide sequences in your code:


self.ran = Random(len(self.generated), np.min(d.descriptor), np.max(d.descriptor)) # generate rand seqs
probas = count_aas(''.join(seq_desc.sequences)).values() # get the aa distribution of training seqs
self.ran.generate_sequences(proba=probas)


But the pseudo-random peptide sequences I generated using this code are completely different from the peptide sequences provided in your appendix. Only 15% of the sequences predicted by the CAMP predictor are AMP, but the pseudo-random peptide sequences you provided exceed 70% Both were predicted to be AMP.
May I know what is this all about? How are your pseudo-random peptide sequences generated?

Looking forward to your reply！

Unexpected behavior with boolean program arguments

Hello Alex,

First and foremost, Thanks a lot for this great repository. I have been using the codebase to fine-tune a pre-trained model on a set of peptides and stumbled upon an unexpected behavior and want to bring it up. As suggested in README, I have run LSTM_peptides.py by setting train argument to False and finetune to True, but the code entered the train branch in the main function (line 727) and started pre-training instead of fine-tuning.

I have dived deeper into the code and found out that argparse module parses the value of train argument as a string, i.e. "False", which is, in turn, casted to boolean by Python and evaluated as True. Thus, the if condition in line 727 (if train:) evaluates to True as long as any non-empty string is provided; triggering the pre-training code.

To isolate and reproduce the error, I have created a small script argparse_test.py with the following content.

import argparse
flags = argparse.ArgumentParser()
flags.add_argument("-t", "--train", default=True, help="whether the network should be trained or just sampled from", type=bool)
args = flags.parse_args()
print(args.train)

When I run this script via a Ubuntu 20 terminal with python3 arparse_test.py --train False (python version is 3.8.10) the output is True. In fact, I have experimented with several values (false, None, True) for train and the code output is always True, except for python3 arparse_test.py --train '', i.e. when train is set to an empty string.

I wonder if I am missing something or the suggested fine-tuning command is supported in certain Python versions. If this is unexpected behavior (which might have been affecting many users), indeed, I'd be happy to create a pull request as a fix.

Looking forward to your reply!

Generating output sequences

Hi. How can I generate the output sequences in fasta or csv files after sampling from trained model?

TypeError: Variable is unhashable. Instead, use tensor.ref() as the key.

Running LSTM_peptide.py script by using
Tensorflow=2.3.1
Scikitlearn=0.23.2
keras>=2.0.2
progressbar2>=3.34.2
modlamp>=3.3.0
to generate novel peptide. It shows this error
Error
raise TypeError("Variable is unhashable. "
TypeError: Variable is unhashable. Instead, use tensor.ref() as the key.
I tried to fix other issues however I am unable to locate this one

requirements

correct entries in requirements

Error in sequence analysis generated in source code

Hello, I've been paying attention to your work recently, but unfortunately, I can't run your code correctly through the examples you provided. When I execute the command python LSTM_peptides.py -- name Train100 -- dataset new_sequences.csv -- layers 2 -- neurons 64 -- epochs 10 sample -- 2, my error goes to the following:

Traceback (most recent call last):
  File "LSTM_peptides.py", line 791, in <module>
    finetune=FLAGS.finetune, references=FLAGS.refs)
  File "LSTM_peptides.py", line 780, in main
    data.analyze_generated(sample, fname=model.logdir + '/analysis_temp' + str(temperature) + '.txt', plot=True)
  File "LSTM_peptides.py", line 400, in analyze_generated
    a.plot_summary(filename=fname[:-4] + '.png')
  File "/home/zh/anaconda3/envs/LSTM_Squence/lib/python3.5/site-packages/modlamp/analysis.py", line 197, in plot_summary
    self.calc_len()
  File "/home/zh/anaconda3/envs/LSTM_Squence/lib/python3.5/site-packages/modlamp/analysis.py", line 177, in calc_len
    d.length()
  File "/home/zh/anaconda3/envs/LSTM_Squence/lib/python3.5/site-packages/modlamp/descriptors.py", line 264, in length
    desc.append(float(len(seq.strip())))
AttributeError: 'list' object has no attribute 'strip'

And some of the grammar in your code is Python 2, and some of the grammar is Python 3. If I don't modify your source code, I won't be able to use Python 2 to create the model, because there will be some unexpected errors in Python 2. I modified some of your code to be Python 3 grammar, so that I can correctly train the model and generate the sequence.
The reason I got the error is because some of the sequences from the d. length () function are strings and some are lists. I don't understand why. Shouldn't the generated sequences be strings? Why is there a list? Please tell me how I can solve this mistake?