SOTA graphs 2022

[thegodone]

4 topics:

A. transformers:

• MAT ( https://github.com/ardigen/MAT & https://arxiv.org/abs/2002.08264)
• RMAT (https://github.com/gmum/huggingmolecules & https://arxiv.org/abs/2110.05841)

B. Recursive Graphs:

• RGIN ( https://arxiv.org/pdf/1906.12192.pdf + code https://github.com/microsoft/tf2-gnn & https://arxiv.org/abs/1912.11589 + https://github.com/HKUST-KnowComp/NeuralSubgraphCounting, my favorite will be simplest RGIN * )

C. Some finetune versions of existing codes:

• MolNet ( https://arxiv.org/abs/2203.09456 & https://github.com/CIS-group/MolNet)
• MolGIN (https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9189862)
• CMPNN ( https://github.com/SY575/CMPNN & https://www.ijcai.org/proceedings/2020/0392.pdf)
• GemNet: Universal Directional Graph Neural Networks for Molecules (https://github.com/TUM-DAML/gemnet_tf & https://arxiv.org/abs/2106.08903 & https://chemrxiv.org/engage/chemrxiv/article-details/6246035d3affe4aa143c3848 )

D. 3D Equivariants strategies (SE3 => "Chirality friendly" or E => "Achiral ...") essential to have them:

• E(n) Equivariant GNN https://arxiv.org/abs/2102.09844 & https://github.com/lucidrains/egnn-pytorch & https://github.com/vgsatorras/egnn)
• Equivariant GNN for 3D Macromolecular Structure // Steerable-E3-GNN
• SE(3)-equivariant GNN (https://arxiv.org/abs/2106.02347
• https://pubs.acs.org/doi/10.1021/acs.jctc.1c01021 & https://github.com/rinikerlab/EquivariantMultipoleGNN

E. Unsuppervised Graph:

• https://github.com/chao1224/n_gram_graph https://papers.nips.cc/paper/2019/hash/2f3926f0a9613f3c3cc21d52a3cdb4d9-Abstract.html

=> GemNet is very existing model
(*) I think simpler is better so we must focus on low parameter architecture first!

[PatReis]

At the moment we worked on crystal graph models. We added CGCNN.
I want to add E3GNN or ALIGNN next.

[thegodone]

I think E3GNN is a nice addition yes as well as ALIGNN of course

[thegodone]

In the paper CMPNN they said Max pooling and in the code you put softmax is there any reason for that change ?

what is still strange for me they take the [0] value there in the product second term

[PatReis]

Yeah, sorry I should have put "segment_max". I mixed it up, which is why the error you sent me occured. In attention you have softmax and multiply with edges and then pool via sum. But here it seems that is sumed and afterwards the nodes multiplied with max. I would have to check the tensors to see why they have a [0] here...
I run training again...

[thegodone]

I just add this paper that you should look at https://pubs.acs.org/doi/10.1021/acs.jctc.1c01021

[thegodone]

GemNet is very existing model

[thegodone]

I just add this paper that I forget : https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9189862
it's a variant of GIN that is very close to AttFP / Chemprop in performances

I tried in the past to make this code in PyG but never finish it: pyg-team/pytorch_geometric#1729

[thegodone]

I just Add the RMAT that look like a rbf + envelop improvement over MAT logic

[thegodone]

I encourage to do this https://github.com/HannesStark/3DInfomax

[thegodone]

https://arxiv.org/pdf/2111.06283.pdf https://github.com/KarolisMart/DropGNN looks very interesting for MPNN / DMPNN / GIN

[thegodone]

this looks cool : https://github.com/invokerqwer/Chemical-AI https://academic.oup.com/bib/advance-article/doi/10.1093/bib/bbac560/6931719

[thegodone]

Reading https://arxiv.org/pdf/2003.00982.pdf I suggest to add the PE versions (aka graph positional encoding) see AQSOL performances Table 4 (GatedGCN-E-PE) => https://github.com/graphdeeplearning/benchmarking-gnns

[thegodone]

can you tell me what are the next coming architecture in your pipeline ?