Add ModelSEED-retrained dGPredictor as an additive reaction-energy source

[freiburgermsu]

Summary

Adds the ModelSEED-retrained dGPredictor as its own additive per-method reaction-energy source, dGPredictor-ModelSEED, alongside the existing Group contribution / eQuilibrator / dGPredictor records. Purely additive — the original KEGG-based dGPredictor record is left untouched, there are no canonical deltag/deltagerr/reversibility changes, and no .tsv or compound-file changes (thermodynamics is a JSON-only field). This continues the additive per-source philosophy of #263.

What this is

dGPredictor (Wang et al. 2021) as shipped under Biochemistry/Thermodynamics/dGPredictor/ was trained on KEGG compound structures. dGPredictor-ModelSEED is the same model retrained on the ModelSEED compound structures: every ModelSEED compound carrying a complete structure is re-decomposed into atom-centered fragments (radius 1 & 2), expanding the group vocabulary, and the BayesianRidge model is refit on the same 4,001 experimental measurements remapped into ModelSEED ID space. It predicts dG for 31,924 reactions (pH 7, I 0.25 M, 298.15 K), including ~11,400 reactions the original KEGG-based model could not reach (compounds with no KEGG cross-reference).

Each reaction now carries both dGPredictor estimates side-by-side, e.g. rxn00001:

"thermodynamics": {
    "Group contribution":    [4.15, 1.22, "="],
    "eQuilibrator":          [-3.46, 0.05, ">"],
    "dGPredictor":           [-3.82, 0.02, ">"],
    "dGPredictor-ModelSEED": [-3.77, 0.87, ">"]
}

New-coverage reactions (e.g. rxn00013) carry a dGPredictor-ModelSEED record where there is no original dGPredictor one — and their canonical deltag is still left untouched.

How

• New staged predictions: Biochemistry/Thermodynamics/dGPredictor/modelseed_retrained_dG.json — {rxn: {dG_mean, dG_uncer}} in kJ/mol, 31,924 reactions.
• New writer: Scripts/Thermodynamics/Update_Reaction_dGPredictor_ModelSEED_Energies.py — stores dGPredictor-ModelSEED [energy, error, operator] (kJ→kcal /4.184); the operator is this estimate's own thermodynamic direction via the shared reversibility_from_energy(). Added to Rerun_Thermodynamics.sh.
• All predictions are recorded as-is; the heavy-tailed extrapolations (macromolecule biosynthesis) carry correspondingly large error bars and never affect the canonical served value.

Data changed

• 31,924 reactions gained a dGPredictor-ModelSEED record; 24,088 reactions unchanged.
• Verified: every modified reaction differs from dev only by the added dGPredictor-ModelSEED key (deep per-reaction JSON equality across all 56,012 reactions; 0 other-field changes), every added value equals dG_mean/4.184, and re-running the writer is idempotent (byte-identical output). Zero changes to canonical deltag/deltagerr/reversibility/notes, the original dGPredictor record, any other field, or any .tsv.

Docs updated in Scripts/Thermodynamics/README.md and Biochemistry/Structures/sources.yaml.

🤖 Generated with Claude Code