Question about the RES task

[YanjingLiLi]

Hi, I have a question about the RES task. For each example, the subunit indices contain atoms and residues from different chains of the protein. Do you just treat them as a complete protein chain to predict the type of residue at the masked place?

[Oxer11]

Yes. The RES task is defined as predicting masked residue types based on local structural enviroments. So we save the subunit indices for each sample in each protein and use it to extract the corresponding part for each sample.

SiamDiff/siamdiff/dataset.py

Lines 237 to 263 in 2bc9682

	def get_item(self, index):
	subunit, protein_index = self.data[index]
	protein = self.protein[protein_index]
	protein = protein.subgraph(subunit.node_index)
	residue_mask = torch.zeros((protein.num_residue,), dtype=torch.bool)
	residue_mask[protein.atom2residue] = 1
	protein = protein.subresidue(residue_mask)
	node_mask = (protein.residue_number[protein.atom2residue] == subunit.residue_num) & \
	(protein.atom_name == data.Protein.atom_name2id["CA"])
	with protein.node():
	protein.ca_mask = node_mask
	with protein.graph():
	protein.label = torch.as_tensor(subunit.label)
	if hasattr(protein, "residue_feature"):
	with protein.residue():
	protein.residue_feature = protein.residue_feature.to_dense()
	protein.residue_feature[protein.residue_number == subunit.residue_num, :] = 0
	protein.residue_type[protein.residue_number == subunit.residue_num] = 0
	if not hasattr(protein, "atom_feature"):
	with protein.atom():
	protein.atom_feature = torch.zeros((protein.num_atom, 1))
	item = {"graph": protein}
	if self.transform:
	item = self.transform(item)
	return item