masashitsubaki / cpi_prediction Goto Github PK

I am referring to your paper. Can you provide data for versions 1:3 and 1:5? My email is [[email protected]]. Thanks a lot for your help.

I used the dataset "human" in your paper, run your code and calc AUPR. The result was that the AUPR is almost 1. I think its too high, a little weird. I am new in this area, could you give some explanation?
Thank you. @masashitsubaki

I just follow the author's advice to train my own datasets,but i have a problem during the training as following:

Training...
Epoch Time(sec) Loss_train AUC_dev AUC_test Precision_test Recall_test
C:\Anaconda3\envs\my-rdkit-env\lib\site-packages\sklearn\metrics_classification.py:1221: UndefinedMetricWarning: Precision is ill-defined and being set t
o 0.0 due to no predicted samples. Use zero_division parameter to control this behavior.
_warn_prf(average, modifier, msg_start, len(result))
C:\Anaconda3\envs\my-rdkit-env\lib\site-packages\sklearn\metrics_classification.py:1221: UndefinedMetricWarning: Precision is ill-defined and being set t
o 0.0 due to no predicted samples. Use zero_division parameter to control this behavior.
_warn_prf(average, modifier, msg_start, len(result))
1 18.5711581 2328.3497102856636 0.4819866010391031 0.5351374101374101 0.0 0.0

so how to deal with this problem?Thanks!

Hi, I have a question about the "attention_cnn" module.
In general, it uses "softmax" rather than "tanh" to get the weights. But as below, your code it uses "tanh", so I am a little confused. Are there some reasons to use like that? thanks!

Thanks for this excellent method for predicting protein-compound interactions. I am just wondering how to make predictions for some compounds? I have trained a model using my own data, the training curve seems to be right. So, how can I make predictions for some other data using this trained model?

Greeting sir,
Thank you for this work.

How can I use this model for prediction?

Thanks

Hi, it seems that your code can't do "new drug" or "new target" problem, for every drug learns an embedding vector, so "new drug" type can't be predicted and "new target" as the same. Is this true?
Thank you:)

Hi,

Could you please provide the script for the negative sample construction? I found that the negative sample construction link provided by Liu et al., (2015) is invalid. Hence, I guess maybe you have the original code.

Thank you very much!

excuse me，I have completed data preprocessing and training, and generated the model file. But how to use the model?

Hello, excuse me, how to use this model after training? For example, predict the interaction of only one protein with a small number of compounds.

Is the data division of 8: 1: 1 used to stop the training of the model with one of them as the validation set?

Traceback (most recent call last):
File "C:/Users/user/Downloads/15647NeonBand.RarZipExtractorPro_g3b9h1p9bdemw!App/CPI_prediction-master/CPI_prediction-master/code/newtest1.py", line 153., in
setting) = sys.argv[1:]
ValueError: not enough values to unpack (expected 14, got 0)
################
I am getting this type of error. How can I solve it? Can you provide me the original code to me please....
My email id [email protected]

I am referring to your paper. Found that the data you provided is only the positive:negativate=1:1 version. Can you provide data for versions 1:3 and 1:5?My email is [email protected] you very much.

Thanks for sharing the code!

I read the code carefully, but I found some questions.

• 1 How about the end2end test? what the end2end mean is, that mol and protein in, and interaction out.

• 2 The molecular input without any feature, will it make any sense, with GNN?

• 3 Where is edge features? Description in the paper, but no line in the code.

Torch version: 1.1.0

As the title, i trained the code directly on the dataset human, but got the results below, which were different from the paper, is the code not same with the paper?

The code uses GPU...
Training...
Epoch Time(sec) Loss_train AUC_dev AUC_test Precision_test Recall_test
1 1.043790566000098 55.68943989276886 0.75 0.9583333333333334 0.6 1.0
2 1.8972396280005341 55.123449206352234 0.8125 1.0 0.6 1.0
3 2.755670120001014 54.722046077251434 0.8125 1.0 0.6 1.0
4 3.6050277929971344 54.029464304447174 0.9375 0.9583333333333334 0.6 0.5
..........................................
98 84.71962930099835 0.011957645416259766 1.0 0.7916666666666667 0.75 0.5
99 85.56498634699528 0.01192164421081543 1.0 0.7916666666666667 0.75 0.5

python version and requirments version please

I'm referring to your paper. I followed the setting mentioned in your paper " We randomly divided 102 DUD-E targets into 72 targets as a training dataset and 30 targets as a test dataset" to process DUD-E dataset, but I couldn't get the results as your paper. I want to know what I can do to get the best results, if I need to change the hyper-parameters?
I hope to your reply, thank you!