Add the SSI-DDI Model

[benedekrozemberczki]

• Please read the paper first. It is here.
• There is also code-release with the paper here.
• After that read the contributing guidelines.
• If there is an existing open-source version of the model please take a look.
• ChemicalX is built on top of PyTorch 1.10. and torchdrug.
• A similar model is EPGCNDS which usesGraphConvolutions to generate drug representations. Take a look at the layer definition here. You should use the layers from torchdrug not the models.
• The library heavily builds on top on torchdrug and molecules in batches are PackedGraphs.
• There is already a model class under ./chemicalx/models/
• Context features, drug level features, and labels are all FloatTensors.
• Look at the examples and tests under ./examples/ and ./tests/.
• Add auxiliary layers as you see fit - please document these, add tests, and add these layers to the main readme.md if needed.
• Add typing to the initialization and forward pass.
• Non-data-dependent hyper-ammeters should have default values.
• Please add tests under ./tests/ and make sure that your model/layer is tested with real data.
• Write an example under ./examples/. What is the AUC on the test set? Is it reasonable?

[YuWVandy]

Where can I find the contribution guideline and looks like the input of the original model in that paper only has two graph data and different interaction types without context features, do I need to incorporate also the context features into the current framework? Or do I need to find a way to transform the context features into a specific interaction type? Sorry that this is my first time using such a cooperative mode of GitHub and there might be some silly questions.

[benedekrozemberczki]

1. Context features are unused in this model.
2. There is a public reference implementation out.

[YuWVandy]

1. Context features are unused in this model.
2. There is a public reference implementation out.

Thanks! Where can I find the public reference implementation out? Is it the documentation of this ChemicalX or just the original code associated with this paper?

[benedekrozemberczki]

https://github.com/kanz76/SSI-DDI/blob/master/models.py

Do not rely on PyTorch Geometric.

[YuWVandy]

Sure, I will try to build myself the Graph Attention Layer, which is one of the most important building block for SSI-DDI

On Feb 2, 2022, at 12:22 PM, Benedek Rozemberczki @.**@.>> wrote:

https://github.com/kanz76/SSI-DDI/blob/master/models.pyhttps://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fkanz76%2FSSI-DDI%2Fblob%2Fmaster%2Fmodels.py&data=04%7C01%7Cyu.wang.1%40Vanderbilt.Edu%7Cb51098f241e14c071cba08d9e6790151%7Cba5a7f39e3be4ab3b45067fa80faecad%7C0%7C0%7C637794229648404353%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=EgYoJG2aVysTsNjx70olhWWrjJEEhEcCU2elXCjW93I%3D&reserved=0

Do not rely on PyTorch Geometric.

— Reply to this email directly, view it on GitHubhttps://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2FAstraZeneca%2Fchemicalx%2Fissues%2F11%23issuecomment-1028226190&data=04%7C01%7Cyu.wang.1%40Vanderbilt.Edu%7Cb51098f241e14c071cba08d9e6790151%7Cba5a7f39e3be4ab3b45067fa80faecad%7C0%7C0%7C637794229648404353%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=l4aKvc6N%2BEQcYmZLGlfN8Pkd06SgAfAFbsIeqd5qCxQ%3D&reserved=0, or unsubscribehttps://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fnotifications%2Funsubscribe-auth%2FAM2OPGTDBC3QD6FZ6R6AEA3UZFY7HANCNFSM5LWMCK3A&data=04%7C01%7Cyu.wang.1%40Vanderbilt.Edu%7Cb51098f241e14c071cba08d9e6790151%7Cba5a7f39e3be4ab3b45067fa80faecad%7C0%7C0%7C637794229648404353%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=cGVinCCZ1bWUYhoSLd8Lv2Rl6xTbb5z7emK9WS1GOHM%3D&reserved=0. Triage notifications on the go with GitHub Mobile for iOShttps://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fapps.apple.com%2Fapp%2Fapple-store%2Fid1477376905%3Fct%3Dnotification-email%26mt%3D8%26pt%3D524675&data=04%7C01%7Cyu.wang.1%40Vanderbilt.Edu%7Cb51098f241e14c071cba08d9e6790151%7Cba5a7f39e3be4ab3b45067fa80faecad%7C0%7C0%7C637794229648560567%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=TxAOv%2BXCkUZENlgTOT3L6IDSHX%2BWke2yO1M4OC5CFNw%3D&reserved=0 or Androidhttps://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Fplay.google.com%2Fstore%2Fapps%2Fdetails%3Fid%3Dcom.github.android%26referrer%3Dutm_campaign%253Dnotification-email%2526utm_medium%253Demail%2526utm_source%253Dgithub&data=04%7C01%7Cyu.wang.1%40Vanderbilt.Edu%7Cb51098f241e14c071cba08d9e6790151%7Cba5a7f39e3be4ab3b45067fa80faecad%7C0%7C0%7C637794229648560567%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=oRJMMBAcD47UTfhmOaCAKyuJW%2BzhpJFIw2fZM%2FYgino%3D&reserved=0. You are receiving this because you commented.Message ID: @.***>

[benedekrozemberczki]

Torchdrug already has that layer.

[YuWVandy]

The original paper defines different M learnable matrices for different interaction types in Equation (11), I guess here in our problem, can we assume only one interaction type is considered since the input is only the leftmolecules and rightmolecules.

[benedekrozemberczki]

Yes.

[YuWVandy]

Just finished coding the model part, which is all based on torchdrug, pytorch and the current existing packages of chemicals, and currently working on testing the model. One question is that does the current pipeline support GPU training? If not, I guess I will just modify it a little bit on my end and try to train using GPU to test my implementation.

[YuWVandy]

For testing whether our implementation is right, can I directly write a testing case following the previous script written in the test folder on github? But one thing is since it currently does not support GPU training, it will take me so long time to get the results. Can someone help me with that?