Providing cyclic peptides for Chai-1

[ZhuNing9527]

How to predict the structure of cyclic peptides (amide bonds forming a ring)

[aozalevsky]

You can submit it as a ligand using SMILES.

P.S. Not a ChaiDiscovery employee; was just passing by.

[ZhuNing9527]

You can submit it as a ligand using SMILES.您可以使用 SMILES 将其作为配体提交。

P.S. Not a ChaiDiscovery employee; was just passing by.PS 不是 ChaiDiscovery 员工;只是路过。

Thank you. But my macrocyclic peptide has a secondary structure like α-helix. If smiles is used as input, the output may not be able to interpret this information.

[arogozhnikov]

@ZhuNing9527
Idea for you to try: add a max distance restraint between first and last residue of peptide

[ZhuNing9527]

@ZhuNing9527 Idea for you to try: add a max distance restraint between first and last residue of peptide

Thank you for your suggestion. I am trying it. In addition to this, I wonder if I can modify the relative position encoding method to achieve

[arogozhnikov]

@ZhuNing9527

I wonder if I can modify the relative position encoding method to achieve

Yup, you need to modify this place:

chai-lab/chai_lab/data/features/generators/relative_token.py

Lines 36 to 53 in f77f441

	def _generate(
	self,
	token_index: Int[Tensor, "b n"],
	token_residue_index: Int[Tensor, "b n"],
	token_asym_id: Int[Tensor, "b n"],
	) -> Tensor:
	rel_sep, rel_residue, rel_chain = map(
	lambda x: rearrange(x, "b n -> b n 1") - rearrange(x, "b n -> b 1 n"),
	(token_index, token_residue_index, token_asym_id),
	)
	mask = (rel_residue == 0) & (rel_chain == 0)
	rel_sep = torch.clamp(rel_sep + self.r_max, 0, 2 * self.r_max + 1)
	# zero inter-residue and inter-chain
	rel_sep = rel_sep.masked_fill(~mask, 2 * self.r_max + 2)
	return self.make_feature(rel_sep.unsqueeze(-1))