Missing external deps + incomplete env make repo hard to reproduce

[cytokineking]

First, thanks again for releasing PPIFlow!

I ran into a few snags when trying to set up the full pipeline end‑to‑end. The repo and environment.yml don’t currently include (or link to) key external dependencies that are referenced throughout the notebook/scripts. Issues I hit:

1. Private/undocumented forks used in scripts
   • demo_vhh.ipynb and demo_scripts/run_abmpnn.sh point to a private “ProteinMPNN‑qin” and AbMPNN weights.
   • flowpacker was referenced with a path on your cluster; it is not bundled or linked in the repo.
   • These dependencies are essential to reproduce the notebook’s pipeline.
2. environment.yml is incomplete for the full pipeline
   • No AF3Score / AlphaFold3 build dependencies.
   • No FlowPacker / ProteinMPNN/ AbMPNN requirements.
   • The result is that a “clean install” can run core PPIFlow, but not the full pipeline in demo_vhh.ipynb.
3. Hard‑coded cluster paths in scripts
   • Many scripts point to /lustre/... or your internal environments (e.g., FlowPacker, AF3Score, AbMPNN, Rosetta).
   • This makes it very hard for outside users to reproduce.

Suggested fixes:

• Bundle or link all external repos explicitly (FlowPacker, ProteinMPNN/AbMPNN, AF3Score). If private forks are required, provide a public mirror or release tarball.
• Add an install script (or a requirements_pipeline.txt) that installs the full pipeline, not just core PPIFlow.
• Parameterize paths (env vars or CLI args) instead of hard‑coding /lustre/....
• Optional Rosetta: it appears to be an analysis step only: consider marking it optional and off by default.

Ultimately I was able to get most of the pipeline working (though not the AbMPNN link yet). If you can point to a suitable repo/weights that can be used, I would be happy to create this install script.

Thanks again.

[XikunHuang]

Hi @cytokineking ,

Thank you very much for your kind words and for such a high-quality, detailed issue report. We sincerely apologize for the inconvenience caused by the hard-coded cluster paths and the missing dependencies. This is incredibly helpful for us as we transition the project from an internal research tool to a public-facing repository.

We are currently working on the following updates based on your suggestions:

• Path Parameterization: We are replacing all /lustre/ internal paths with configurable arguments and environment variables.

• Pipeline Documentation: We will explicitly link/bundle the required versions of FlowPacker, ProteinMPNN/AbMPNN and AF3Score to ensure a "clean install" works for the full demo_vhh.ipynb pipeline. We will provide a comprehensive installation script to cover the full end-to-end workflow.

• We will mark Rosetta as an optional dependency to keep the core installation lightweight.

Regarding AbMPNN: Thank you for your patience here. You can find the weights for AbMPNN at this Zenodo record: https://zenodo.org/records/8164693.

We would absolutely love to have your help with the install script! Thanks again for helping us make PPIFlow more accessible to the community.

[cytokineking]

Hi Xikun,
Thanks; I plugged in the AbMPNN weights and it's working fine now. Now that I've been playing in the repo, my ambitions have grown bigger than an install script. I'll shared a fork later for you guys to consider for a potential PR that addresses some issues (e.g., easy install, handle multichain targets, use bindcraft-style hotspot definitions, avoid chain ID clashes, etc) and implements new features (full pipeline automation, resume behavior, parallelization, etc).

[lg1243610235-cmyk]

Dear cytokineking, your PPIFlow-Pipeline works very well. As a beginner, when I run step 3, it prompts that a command for flowpacker is required. I understand that this requires modifying the corresponding YAML file to add the command. Could you please provide examples of how users should add these command?

[cytokineking]

Pull the latest commit; the previous install script did not install flowpacker by default. All dependencies are required for the full pipeline; flowpacker, rosetta, AF3score, etc, and the install script now reflects that. You can re-run the install script and it should patch your environment.

I recommend raising issues about PPIFlow-Pipeline on that fork repo issue page to not clutter the canonical upstream repo issues. I am still working out the kinks and need to provide more documentation, but the full pipeline now works for binders, VHH, and mAbs, including for multichain targets (by concatenating them into a single chain like BindCraft).

Once I am finished with the pipeline, I will submit it to the PPIFlow team for consideration if they wish to merge it with this repo.

[Samuel-gwb]

The PPIFlow-Pipeline is just great !
Thanks for your work !
A question is that af3_db_stub seems containing just empty files. Is this just due to independency of database to run af3score ?
I cannot find place to report issues at the site of the PIFlow-Pipeline repo.

[lg1243610235-cmyk]

AF3Score skip the msa part, ref: "AF3Score: A Score-Only Adaptation of AlphaFold3 for Biomolecular Structure Evaluation"

[cytokineking]

The PPIFlow-Pipeline is just great ! Thanks for your work ! A question is that af3_db_stub seems containing just empty files. Is this just due to independency of database to run af3score ? I cannot find place to report issues at the site of the PIFlow-Pipeline repo.

You need the stub because alphafold3 expect the data pipeline, but we don't install it with the script. so we just make a stub to satisfy that.

[lg1243610235-cmyk]

Thanks for the update! I ran into another issue that seems specific to PPIFlow-Pipeline.
I wanted to open an issue on the fork repo as suggested, but it looks like Issues are currently disabled there.
For now, I’ll briefly describe it here to avoid blocking progress. Please let me know if you prefer another place.
Update (2026-01-29): the latest version from cytokineking can run end-to-end.
However, I found an issue in the first AF3Score → Rosetta step.
Problem:
The PDB files passed to Rosetta after the first AF3Score contain coordinates with values like 0.00, which causes Rosetta to fail.
Tracing this back, the issue comes from the cif2pdb output generated by AF3Score: the PDBs in the pdbs/ directory contain invalid zero coordinates.
Solution:
After AF3Score, when selecting structures to move into filtered_pdbs, use the PDBs from metrics_pdbs/ instead of pdbs/.
Code change:
In /home/lggg/PPIFlow-Pipeline/src/pipeline/steps/external.py,
inside class AF3ScoreStep (around ~10 lines below the class definition), change:
name_map = build_name_map(out_dir)
to:
name_map = build_name_map(out_dir / "metrics_pdbs")
After this change, Rosetta runs normally without coordinate-related errors.

[cytokineking]

I opened an issues page on the fork; let's take the further discussion there so as not to clutter the upstream page. Short answer to your question: this is a bug in the code; we shouldn't point to those pdbs at all, but rather the flowpacker outputs. AF3score is just used for scoring. I will make this fix into the next commit, which is a very large refactor.