Input to GCN does not match

[usykaa]

Hi I have been trying to replicate the results of the paper, but the codes in the provided notebooks do not seem to work, the generated dataset when going into the GCN model for feature extraction will result in shape mismatch

mat1 and mat2 shapes cannot be multiplied (xxx x xxx and 39x256)

Also I have noticed there are some variable names that don't exist, which I changed to accomodate the rest of the code, and I still find myself unable to make the scripts run.

Is this because of mismatch of DGL version? Or are the codes run for experiments not the same code provided in the notebooks?

Many thanks, best regards

[usykaa]

I have been attempting to write my own version of the repository to attempt to replicate the results. I also wonder how were the visualizations in the paper created? The authors did not seem to clairify what does each color of cluster represent.

[ph-mehdi]

Hi I have been trying to replicate the results of the paper, but the codes in the provided notebooks do not seem to work, the generated dataset when going into the GCN model for feature extraction will result in shape mismatch

mat1 and mat2 shapes cannot be multiplied (xxx x xxx and 39x256)

Also I have noticed there are some variable names that don't exist, which I changed to accomodate the rest of the code, and I still find myself unable to make the scripts run.

Is this because of mismatch of DGL version? Or are the codes run for experiments not the same code provided in the notebooks?

Many thanks, best regards

Hi,

Thank you for your feedback regarding the replication of our results. I apologize for the issues you've encountered with the provided notebooks.

1. Variable Names: If you've found variable names that do not exist in the code, it might be due to version differences or updates that were made after the notebooks were created. I recommend checking if you are using the latest version of the code.

2. DGL Version: The version of DGL can definitely impact the functionality of the code. Please make sure you are using the version specified in the documentation or requirements file.

I have reviewed the notebooks and made some changes to ensure they run without issues.

Thank you for your understanding, and I appreciate your effort in trying to replicate the work. If you have any more details to share, please do so, and I'll do my best to assist you.

Best regards,

[ph-mehdi]

I have been attempting to write my own version of the repository to attempt to replicate the results. I also wonder how were the visualizations in the paper created? The authors did not seem to clairify what does each color of cluster represent.

Could you please clarify your question? Are you referring to Figure 5? If so, the colors in this figure represent the 27 disorder classes from the SIDER dataset.

[usykaa]

Hi Mehdi! Thank you very much for you and your team's contribution on the project of BioAct-Het and thank you for the reply, the idea of using class vectors to regularize learning was absolutely brilliant! I have written my own implementation based on your paper and the notebook you kindly provided, and the performance on Tox21 was amazing! Clear local smoothness can be observed in the structure however, I am having a little bit of trouble with SIDER. The visualization results on SIDER was quite dense, and quite frankly the AUC was not very high either ( Ranging from 0.67 to 0.75 across the tasks). Could you please share a little bit of detail on how the visualization on SIDER was produced? Also, here's the visualization I have got on some of the SIDER tasks, I used the embedding from the last layer before the output of the Siamese model and mapped to T-SNE 2D space, coloring it based on whether it was active or not, as you can see the separation of data points are not very clear. (If you are interested I will gladly share my implementation, mainly simplifying the datastructure and optimizing the code to work with Pytorch, I tried my best to retain consistency with your code but I guess it was not possible to make it entirely identical). [image: image.png] [image: image.png] Best regards And thank you for the contribution on github once again! Xing Chen Mehdi Peykan Heyarti ***@***.***>于2024年10月30日 周三18:59写道：

…

Hi I have been trying to replicate the results of the paper, but the codes in the provided notebooks do not seem to work, the generated dataset when going into the GCN model for feature extraction will result in shape mismatch mat1 and mat2 shapes cannot be multiplied (xxx x xxx and 39x256) Also I have noticed there are some variable names that don't exist, which I changed to accomodate the rest of the code, and I still find myself unable to make the scripts run. Is this because of mismatch of DGL version? Or are the codes run for experiments not the same code provided in the notebooks? Many thanks, best regards Hi, Thank you for your feedback regarding the replication of our results. I apologize for the issues you've encountered with the provided notebooks. 1. Variable Names: If you've found variable names that do not exist in the code, it might be due to version differences or updates that were made after the notebooks were created. I recommend checking if you are using the latest version of the code. 2. DGL Version: The version of DGL can definitely impact the functionality of the code. Please make sure you are using the version specified in the documentation or requirements file. *I have reviewed the notebooks and made some changes to ensure they run without issues.* Thank you for your understanding, and I appreciate your effort in trying to replicate the work. If you have any more details to share, please do so, and I'll do my best to assist you. Best regards, — Reply to this email directly, view it on GitHub <#2 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AOSEX55GMUSOOQCMYOQIQ5LZ6C3YLAVCNFSM6AAAAABQOSPMFOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDINBWGU4DKNBVG4> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[ph-mehdi]

Hi Mehdi! Thank you very much for you and your team's contribution on the project of BioAct-Het and thank you for the reply, the idea of using class vectors to regularize learning was absolutely brilliant! I have written my own implementation based on your paper and the notebook you kindly provided, and the performance on Tox21 was amazing! Clear local smoothness can be observed in the structure however, I am having a little bit of trouble with SIDER. The visualization results on SIDER was quite dense, and quite frankly the AUC was not very high either ( Ranging from 0.67 to 0.75 across the tasks). Could you please share a little bit of detail on how the visualization on SIDER was produced? Also, here's the visualization I have got on some of the SIDER tasks, I used the embedding from the last layer before the output of the Siamese model and mapped to T-SNE 2D space, coloring it based on whether it was active or not, as you can see the separation of data points are not very clear. (If you are interested I will gladly share my implementation, mainly simplifying the datastructure and optimizing the code to work with Pytorch, I tried my best to retain consistency with your code but I guess it was not possible to make it entirely identical). [image: image.png] [image: image.png] Best regards And thank you for the contribution on github once again! Xing Chen Mehdi Peykan Heyarti @.>于2024年10月30日 周三18:59写道：
…
Hi I have been trying to replicate the results of the paper, but the codes in the provided notebooks do not seem to work, the generated dataset when going into the GCN model for feature extraction will result in shape mismatch mat1 and mat2 shapes cannot be multiplied (xxx x xxx and 39x256) Also I have noticed there are some variable names that don't exist, which I changed to accomodate the rest of the code, and I still find myself unable to make the scripts run. Is this because of mismatch of DGL version? Or are the codes run for experiments not the same code provided in the notebooks? Many thanks, best regards Hi, Thank you for your feedback regarding the replication of our results. I apologize for the issues you've encountered with the provided notebooks. 1. Variable Names: If you've found variable names that do not exist in the code, it might be due to version differences or updates that were made after the notebooks were created. I recommend checking if you are using the latest version of the code. 2. DGL Version: The version of DGL can definitely impact the functionality of the code. Please make sure you are using the version specified in the documentation or requirements file. I have reviewed the notebooks and made some changes to ensure they run without issues. Thank you for your understanding, and I appreciate your effort in trying to replicate the work. If you have any more details to share, please do so, and I'll do my best to assist you. Best regards, — Reply to this email directly, view it on GitHub <#2 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AOSEX55GMUSOOQCMYOQIQ5LZ6C3YLAVCNFSM6AAAAABQOSPMFOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDINBWGU4DKNBVG4 . You are receiving this because you authored the thread.Message ID: @.>

Hi Xing Chen!

Thank you very much for your kind words and for sharing your progress with the project. It’s great to hear about the impressive results you achieved on Tox21.

However, I noticed that the two images you tried to include in this message aren’t visible here.

Regarding the SIDER dataset, I ran the code for the "Association_based_strategy" notebook in the GitHub repository and obtained results identical to those shared there. The t-SNE visualization section also executed smoothly, and the outputs were similar to those in the provided notebook.

It’s possible that certain modifications in your PyTorch-based implementation may have affected the results. To ensure consistency, please consider using the same libraries specified in the repository to achieve comparable results.

Thank you