Regenerate the pH 7 protonation bundles with Marvin 26.1

[freiburgermsu]

Closes the item Marvin_26.1_pKa_Run.md left under "Not done here": protonations/ still held only the 23.4 bundle. Companion to #289 / #292.

53,127 compounds, 141,715 rows — the row count 23.4 shipped, exactly, in every source. ChEBI, KEGG and Rhea cover an identical compound set. MetaCyc swaps one in each direction. Zero unparseable, zero non-standard InChI, zero InChIKeys disagreeing with the InChI beside them.

source	compounds	rows	23.4 rows
ChEBI	11,544	30,488	30,488
KEGG	16,275	46,911	46,911
MetaCyc	25,071	63,665	63,665
Rhea	237	651	651
total	53,127	141,715	141,715

Two things to weigh before merging

1. The tautomer step is missing

ChargeMol.java took the dominant tautomer at pH 7 before the major microspecies. That needs TautomerizationPlugin, licensed under the Isomers Plugin Group; these regenerations run on a Protonation-group licence. Verified, not assumed:

$ cxcalc "CC(=O)CC(=O)C" majortautomer -H 7
chemaxon.license.api.LicenseException: No valid license has been found.
Product name: Isomers Plugin Group

Net charge at pH 7 is unchanged for 79.7% of shared compounds and different for 20.3%, skewed toward deprotonation (11.8% one unit lower, 4.1% two or more; 3.6% one higher).

That 20.3% is an upper bound on the tautomer effect, not a measurement of it. It also contains the genuine 23.4 → 26.1 engine change, which the pKa work already showed moves site counts on roughly one shared set in seven — a pKa crossing 7 flips the protonation state on its own. The two cannot be separated without an Isomers licence, and the report does not pretend otherwise. Restoring the step is a few lines if that licence is obtained.

2. Marvin's InChI and SMILES importers lose stereochemistry

Likely the most reusable finding here, and not documented anywhere obvious. Round-tripping the source InChI with no protonation at all — import, export, compare:

import path	reproduces the source InChI
MolImporter.importMol(<InChI>)	61%
MolImporter.importMol(<SMILES>)	39%
MolImporter.importMol(<RDKit molblock>)	99.6%
RDKit alone (control)	99.7%

The loss is on import, not export — what goes is the /t and /m parity layers. Every structure is therefore bridged through RDKit into a molblock before Marvin sees it, which is what ChargeMol.java was doing when it read args[0] as a mol file rather than a string.

Measured on the compounds where protonation changes nothing, so output should equal input:

source	no-op compounds	this run	23.4
ChEBI	7,128	99.0%	96.6%
KEGG	8,346	99.0%	97.8%
MetaCyc	13,738	98.6%	96.0%
Rhea	135	98.5%	98.5%

Unbridged this scores 55–61% — and every coverage count still looks perfect. inchi:AuxNone,SAbs on export matters just as much: without SAbs, 30.9%.

Smaller notes

• InChIKey is derived from the InChI actually written. Marvin's inchikey export ignores SAbs and disagrees with its own inchi export — for POLYMER_10033 it emits ...-ARPYZQPTNA-N where the InChI it just wrote hashes to ...-VFUOTHLCSA-N, which is what 23.4 shipped. 0 disagreements across all 44,294 InChI rows.
• One engine, not the CLI. Plugin and cxcalc majorms agree on 296/300 ChEBI compounds; all 4 exceptions are organometallics where cxcalc silently drops the metal fragment (C35H34MgN4O5 → C35H34N4O5). Same reason ChargeMol.java fragmented and re-fused.
• Query molecules are recovered, not dropped. aromatize() on query atoms leaves bonds no SMILES writer can express, and MetaCyc spells its protein-bound cofactors with ~ any-bonds. Without handling both, 156 MetaCyc compounds would have been lost silently.
• Charge on wildcard-adjacent atoms is inherited. 23.4 wrote *[NH2+][C@@H](... for POLYMER_12621; so does this. Correcting it is a curation decision about what * means, so it is recorded and left alone.
• Failures, by id. Four compounds yield no row; three are H⁺ (InChI=1S/p+1, no heavy atom — 23.4 has no row either, so agreeing is correct). The one real failure is MetaCyc HypC-Dimer-Fe-CO2, an ArrayIndexOutOfBoundsException in PkaPlugin.run() — one row out of 141,715.
• Re-running rewrites a bundle byte for byte.

Not done here

• Compound records untouched — Update_Compound_Structures_Formulas_Charge.py is the step that rewrites them.
• Print_Structure_Formula_Charge.py has not been re-run over these files; the formula/charge columns are Marvin's own.
• ChEBI keeps bare ids in this bundle, matching the 23.4 protonation file (the pKa bundles use CHEBI_). The id-format migration stays separate.

Full detail in Biochemistry/Structures/_reports/Marvin_26.1_Protonation_Run.md.

---

Generated by Claude Code