fix: LBP pipeline broken due to wrong Chai-1 FASTA generator

[Xinping-Liu]

Problem

The LBP (and interface) pipeline failed silently at the refold_prepare stage. _prepare_refold_chai1 called make_chai1_fasta_multi_process, which is AME-specific: it looks up task names from an AME CSV and reads pre-built ligand SMILES from assets/ame/standard_nonprotein_fasta/. LBP has no AME CSV, so all FASTA files failed to generate with "Failed to get task name" warnings, and Chai-1 received zero inputs.

Fix

Add make_chai1_fasta_from_backbone_dir to ReFold, a general-purpose method that:

• Reads ligand residue codes directly from HETATM records in each backbone PDB
• Looks up SMILES from the CCD components.cif database (no AME CSV required)
• Reads LigandMPNN-generated protein sequences from inverse_fold/seqs/

Update _prepare_refold_chai1 in pipeline_framework.py to use the new method for LBP/interface tasks. AME continues to use make_chai1_fasta_multi_process unchanged.

Test

Ran end-to-end on 3 LBP backbone structures (× 8 sequences = 24 designs):

• refold_prepare: 24/24 FASTA files generated ✓
• Chai-1 refold: 24/24 structures predicted ✓
• Evaluation metrics (pLDDT, ipAE, ipTM) computed and written to raw_data.csv ✓

Files changed

• refold/refold_api.py: add make_chai1_fasta_from_backbone_dir method
• scripts/pipeline_framework.py: update _prepare_refold_chai1 to use new method