vkola-lab / peds2019 Goto Github PK

Hi all, I noticed a very important issue - the ablstm.py script returns scores in a different order than the order of the input sequences.

I tried processing a diverse set of sequences in one file (human, humanized and murine therapeutic sequences) and got scores that were not consistent with your published distributions:

At first I thought it was an overfitting issue, but then I found that I am getting a different result when processing just the first few sequences. When I processed the sequences one by one, the scores now fall into the expected ranges:

Hi vkola-lab

Thanks for providing the code you used to implement your model! I have a question about the data preparation. before using your sequences as input to your model, do you filter for unique sequences? or retain duplicates? the oas database seems to contain lots of duplicates in the paired section which is what I'm interested in modelling.

Thanks!

Hello!
I have some antibody sequence data, and I'm in need of a tool to convert them into the sequence format as shown in your example. I'm unsure about how to properly insert '-' for gaps. Could you possibly provide some guidance or suggestions? I would greatly appreciate your help. Looking forward to your response!

Hi,

Thank you very much for providing the code for your model!

I am trying to reproduce your results for this LSTM method for calculating antibody humanness, as presented in Wollacott et al., 2019.

Could you confirm whether the model on your GitHub is identical to the one used in the Wollacott paper and, if so, how to evaluate test data identically to how you did in the paper? If it is not the same model, could you please advise me how to tune the model on GitHub to obtain the same results reported in the paper and/or provide me with access to the pre-trained model that achieved the results reported in the paper?

I have tried to run the model on the human_test.txt data provided using two approaches based on your documentation (full code included below). With one approach (python ablstm.py eval … model_tmp.npy …), I obtain similar but not identical test set results to what is reported in the result_human_test.txt file. With the other approach (from ablstm import ModelLSTM…), I obtain much higher/poorer scores than expected or reported (~3, instead of ~0-1).

Any guidance on how to reproduce the model and results from the paper would be greatly appreciated. Thank you very much in advance for your time and help!

Best wishes,
Alissa Hummer

My code to run the model on the provided test data:

(1) In the command line:
python3 ablstm.py eval data/sample/human_test.txt saved_models/tmp/model_tmp.npy human_test-output.txt

(2) In a Python script:

from ablstm import ModelLSTM
model = ModelLSTM(embedding_dim=64, hidden_dim=64, device='cpu', gapped=True, fixed_len=True)
                 
tst_fn = './data/sample/human_test.txt'
tst_scores = model.eval(fn=tst_fn, batch_size=512)

Also, I am unable to run the saved_models/tmp/lstm_0.589547.npy model, as I encounter the following error:

File "/data/pegasus/hummer/Hu-mAb-revisions/LSTM-test/peds2019/model.py", line 131, in load
  self.gapped = param_dict['gapped']
KeyError: 'gapped'

Hi

In figure 2b of 'quantifying antibody nativeness' where you plot the roc-auc scores for the human vs non-human classifier, do you modify the LSTM architecture with a new fc layer on top? or do you just train an sklearn logistic regression model on the output scores?

Thanks!

Hi

In the methods section of your nativeness paper, you state that 'These sequences were further clustered at
the 97% identity level to avoid sampling highly related sequences between the training and testing sets'

Could you give me some guidance on how this was done/what tools you use?

Thanks