MHC-I initial guess becomes all zeros for two-anchor inputs

[jong3]

Hi PMGen developers,

Thank you for releasing PMGen. We found a possible mismatch between the public MHC-I initial-guess implementation and Methods Section 2.3 of the paper.

We used the README-recommended command with the current master commit (f3f8c546ebb130fcb12b1bdc05c7081225f6fb54):

python run_PMGen.py \
  --mode wrapper \
  --run single \
  --df input.tsv \
  --output_dir output/ \
  --initial_guess

For standard MHC-I inputs with two anchors, AFfine/af2_util.py executes:

if len(full_anchors) == 2:
    all_positions *= 0.  # zero for mhc 1

This appears to zero the complete IG tensor, including all MHC and peptide-anchor coordinates.
Downstream, the IG is added to AlphaFold's default zero prev_pos, so an all-zero IG does not appear to encode anchor–MHC geometry.

However, Methods Section 2.3 states that all MHC coordinates and aligned peptide-anchor coordinates are retained, while only non-anchor peptide positions are masked.

We also noticed that the legacy parser indexes residues only by residue number, without chain ID:
resNo = int(line[22:26])
Because the standardized template chains M, B and P can share residue numbers, simply removing the global zeroing line may still cause chain-coordinate collisions.

Could you please clarify:

1. Is the global zeroing intentional for two-anchor MHC-I inputs?
2. Does the public code correspond to the MHC-I IG implementation used in the paper?
3. What is the recommended code/configuration for retaining MHC and anchor coordinates as described in Methods Section 2.3?

References:

• Paper: https://doi.org/10.1093/bioinformatics/btag381
• Code: https://github.com/soedinglab/PMGen/blob/f3f8c546ebb130fcb12b1bdc05c7081225f6fb54/AFfine/af2_util.py

Thank you!