Repository navigation
Bug: PDB parsing drops residues from secondary subchains in multi-chain entities #236
Description
Description
When a PDB file's COMPND record declares an entity with multiple chains (e.g. CHAIN: A, B), the parser only processes the first subchain.
PBD Example:
COMPND MOL_ID: 1;
COMPND 2 MOLECULE: 3-PHYTASE A;
COMPND 3 CHAIN: A, B;
COMPND 4 SYNONYM: 3 PHYTASE A, MYO-INOSITOL-HEXAPHOSPHATE 3-PHOSPHOHYDROLASE
COMPND 5 A, MYO-INOSITOL HEXAKISPHOSPHATE PHOSPHOHYDROLASE A;
COMPND 6 EC: 3.1.3.8;
COMPND 7 ENGINEERED: YES
Residues in remaining subchains get an empty label_seq_id, which causes is_present to be marked as False.
At CIF serialization after inference, residues with is_present=False are filtered out, so secondary subchains are lost.
Meanwhile feature.pkl still contains all residues from the entity. This creates a residue-count mismatch between the two files, causing failures in downstream stages (inverse fold, etc.) that load both and expect consistent dimensions.
path/to/src/boltzgen/task/predict/data_from_generated.py", line 380, in get_feat
design_chains = np.unique(asym_id[chain_design_mask])
IndexError: boolean index did not match indexed array along dimension 0; dimension is 438 but corresponding boolean dimension is 882
Is there a reason the parser deliberately handles only the first subchain?
Activity
Assign label_seq to every subchain of a multi-chain PDB entity