Fix chem-msms-predict with public MassSpecGym checkpoints (instrument arg + ms_pred patches)

[hugogontijomachado]

Summary

Make chem-msms-predict (ICEBERG forward prediction) actually run with the public
MassSpecGym checkpoints on a standard install. Two changes:

1. predict_msms.py: add --instrument (default Orbitrap)

The public MassSpecGym ICEBERG checkpoints condition on an instrument feature
(Orbitrap | QTOF). The skill never passed one, so instrument defaulted to None,
was serialized to the candidates TSV as an empty cell, read back as NaN, and crashed the
generator with RuntimeError: value cannot be converted to type int64 without overflow.
Added an --instrument argument (default Orbitrap) threaded into iceberg_prediction.

2. conda-envs/msms-agent/install.sh: patch 3 upstream ms_pred bugs after clone

With ms_pred pip-installed and run on CPU with pytorch_lightning ≥ 2.0, ICEBERG forward
prediction hits three upstream bugs (none macOS-specific):

• iceberg_elucidation.py invokes predict_smis.py via a cwd-relative path (only
  resolves from a repo checkout).
• predict_smis.py uses pl.utilities.seed.seed_everything (removed in PL 2.0).
• predict_smis.py calls torch.cuda.set_device(gpu_id) unconditionally (fails on CPU).

A post-clone patch step applies the three fixes so the env is usable out of the box.

Note: the install.sh patch is a temporary bridge. I submitted the same three fixes
upstream — coleygroup/ms-pred#35. Once that merges, this patch step can be dropped
(or replaced by bumping the ms_pred pin).

Testing

Verified end-to-end on macOS arm64 (CPU): glyphosate MS/MS for [M-H]- (precursor
168.0067) and [M+H]+ (170.0213) at CE 10/20/40/60 eV produce sensible spectra
(168→124, 170→88, 170→124, water losses).

[mlederbauer]

Hi! Creator of this skill here. Thank you very much for this addition that enables a user to run ICEBERG on CPU, and the missing instrument flag. I am kindly asking to merge PR hugogontijomachado#1 into your feature branch with some additional edits, after which this branch id ready to merge into main.

conda run -n ms-gen python .agents/skills/chem-msms-predict/scripts/predict_msms.py \
    --smiles "c1ccccc1C(=O)OCCN" \
    --gen_ckpt downloads/iceberg_dag_gen_msg_best.ckpt \
    --inten_ckpt downloads/iceberg_dag_inten_msg_best.ckpt \
    --collision_energies 20 40 \
    --adduct "[M+H]+" \
    --instrument "Orbitrap" \
    --output_dir results/msms_prediction

Thank you very much! cc @bowen-bd

[bowen-bd]

Thanks @hugogontijomachado for the contribution and @mlederbauer for the review!
Can @hugogontijomachado please let us know when @mlederbauer's PR has been merged into this PR?

[hugogontijomachado]

Thanks @mlederbauer for the review and @bowen-bd for taking a look!

@bowen-bd — answering your question: @mlederbauer's PR (hugogontijomachado#1) is now merged into this branch. ✅

Before merging I rebuilt the env from the updated install.sh on macOS (arm64) in a throwaway env to verify. The new macOS path (torch 2.3.0 + DGL-torch2.3 + torchdata stub) builds cleanly — and is actually nicer on arm64 than pinning torch 2.2.2, since PyG ships universal2 torch-scatter/torch-sparse wheels for 2.3.0 (no source build needed).

The rebuild surfaced one runtime issue: with ray unpinned in core_env.yaml, the env resolves ray 2.55 without the tune extra, so predict_smis.py crashes at runtime — ModuleNotFoundError: pyarrow and Can't import ray.tune (ms_pred imports ray.tune.integration.pytorch_lightning, removed in ray ≥ 2.8). I pushed a small follow-up pinning ray[tune]==2.7.2 (+ setuptools<81, since ray 2.7.2 still imports pkg_resources). After that, ICEBERG forward prediction runs end-to-end (glyphosate [M-H]- → precursor 168.0067, spectrum produced).

Branch should be ready to merge into main — thanks both!