Repository navigation
ESMFold2: SMILES ligands are written as LIG, which is a different molecule in the CCD #421
Description
ESMFold2 labels a ligand given as SMILES LIG in the returned MolecularComplex and in to_mmcif(). In the CCD, LIG is an unrelated component, so the label points at the wrong molecule:
LigandInput(id="B", smiles="O=C1NC(=O)C(N1)(c1ccccc1)c1ccccc1") # phenytoin, C15 N2 O2
complex.sequence # ['ALA', 'GLY', 'LIG']
from biotite.structure.info import residue
residue("LIG") # 29 atoms, heavy atoms C15 N3The mmCIF carries no _chem_comp entry for it, so a reader that looks component IDs up in the CCD sees that molecule rather than the ligand. The other tools I checked do not write the bare LIG for SMILES ligands:
| Tool | Name written for a SMILES ligand | CCD entry with that name |
|---|---|---|
ESMFold2 (main) |
LIG |
yes, a different molecule (C15 N3) |
| Boltz | LIG1, LIG2, … |
none |
| Chai-1 | LIG + asym id |
none |
| AlphaFold 3 | LIG_<chain id> |
none |
| OpenFold3 | LIG0, LIG1, … (per distinct SMILES) |
none |
| AtomWorks | UNL |
"Unknown ligand" placeholder, no atoms |
The label comes from tokenize_ligand_smiles in esm/models/esmfold2/prepare_input.py. Naming each distinct SMILES ligand with a non-CCD ID, e.g. LIG1, LIG2 as Boltz does, would remove the collision.
(Names read from source: Boltz b1ebfc4, chai-lab 66c38d1, alphafold3 f43dfb8, OpenFold3 270425a, AtomWorks fcf7af8c, ESM 43b4548; CCD lookups use biotite's bundled CCD.)
Activity