Problems of Investment of MolCLR Representation part

[zqwz909]

Hello, I tried to do the experiment in the Investment of MolCLR Representation part of your paper, but the code did not provide the vector to generate the molecular representation. In the encoder part, I used the representation vector generated by the pre-trained gcn, and the result of cosine similarity was very incorrect. How do you use MolCLR to generate molecular representations in this part of the experiment?

[yuyangw]

Hi, thanks for your interest in our work. If I understand your question correctly, you want to extract the representation from GNN. The representations for input molecules are returned as h (https://github.com/yuyangw/MolCLR/blob/master/models/gcn_molclr.py#L158). Hope this helps.

Best, Yuyang

[zqwz909]

Thank you for pointing out. I have another question. Is the feature vector obtained by the gnn network directly used to compare the cosine distance, or does it need to be normalized?

[yuyangw]

It should be normalized.

[zqwz909]

OK, can you tell me the normalization method you adopted? This part of the experiment in the paper doesn't seem to mention.Thank you!

[yuyangw]

You can refer to the cosine similarity function in pytorch: https://pytorch.org/docs/stable/generated/torch.nn.functional.cosine_similarity.html, which already implements normalization.

[RichardLrx]

Hello, I used the pre-training model (pretrained_gin/checkpoints) provided by you and achieved good results in my work. Therefore, I would like to know which data augmentations technology you used to get the pre-training model pretrained_gin. Is atom masking? bond deletion? subgraph removal?

[yuyangw]

The checkpoint is trained with subgraph removal since it achieves the best overall performance as reported in our paper.

[RichardLrx]

Thank you!