tsjain

From the paper, o1 = ReLU Batch [W1 W2] R1 o2 = ReLU Batch [W2 W1] R2 If I reverse(_r) the order of R1 and R2, I get, o2_r =...

The only code where I see a loop over k is in ScoreMatchingNoisedAnglesDataset class. I am not sure that I was able to find the use of the wrapped gaussian...

Scaling of edge network with 0.1

1

I did not see where the edge network weights are scaled by 0.1 in the code.

Will ReLU be ever called for single dimensional node_state_dim?

1

Since the default configuration of node_hidden_sizes is a vector of length 1, will the layers in the _for_ loop on lines 38:40 ever get created? https://github.com/Lin-Yijie/Graph-Matching-Networks/blob/f63c575ef9c6132207a52d2285cc4583c9947456/GMN/graphembeddingnetwork.py#L38

In order to exclude residues less than _long_interaction_threshold_, the code checks for _residue_number_ value in the pdb dataframe. The _residue_number_ value is identical for residues with insertion codes. Therefore, it...

As the subject line indicates, I was confused by the division of structural clash loss by the number of atoms. The equation in the paper indicates a overall summation, not...

It wasn't clear from the paper the rationale for increasing this to 20A from the 10A for intra-chain fape. https://github.com/google-deepmind/alphafold/blob/2ca41b1991bbf17be4a86c1face55597ba4c7621/alphafold/model/config.py#L662C22-L662C40

Why isn't there an additional broadcast dimension over the atom indices? Won't this code only calculate pairwise distances between the same atom types in different residues instead of all pairwise...

From the code it appears that the loss for an entire batch is scaled by either the crop length or the average of the sequence length for the batch. Shouldn't...

For the soft or logit design phase, how is the aatype information passed to AF2 structure module? I believe that part needs discrete amino-acid identities in order to assign frames...