xyzrecover

xyzrecover is a small RDKit-based package for recovering molecular identifiers from XYZ coordinate files when the bond graph and charge were lost.

It is designed for batches of XYZ files that may contain one molecule, multiple disconnected covalent molecules, or multi-frame XYZ blocks. It uses RDKit's native rdDetermineBonds / xyz2mol implementation to infer connectivity and bond orders, then emits canonical SMILES, explicit-H SMILES, formal charge, formula, InChI, and InChIKey.

Important limitation

XYZ files do not contain bond orders, formal charges, spin state, resonance choice, protonation intent, or non-covalent association metadata. For some geometries, more than one chemically valid assignment is possible. xyzrecover therefore searches candidate charges, scores the successful assignments, and includes alternates in the output instead of pretending the answer is always unique.

When you know the total charge, pass it. That is the single most useful constraint.

RDKit's valence model only handles main-group organic elements. Components containing metals or noble gases are reported with status="unsupported" and smiles=None rather than perceived, because RDKit would otherwise return a silently wrong structure at high confidence (a lone Na⁺ comes back as [NaH]). The supported set is DEFAULT_SUPPORTED_ELEMENTS; override it via PerceptionConfig(supported_elements=...), or disable the check with restrict_to_supported_elements=False (CLI: --no-element-check).

The one exception is ionic metal counterions: an isolated group 1 / group 2 metal atom (Na, K, Li, Ca, Mg, …) is recovered directly as its ion ([Na+], [Ca+2]) using the group oxidation state, bypassing RDKit. Such atoms are split out of any short ionic contact first, so a salt like Na⁺·methyl-phosphate⁻ recovers the metal as [Na+] and perceives the anion separately, and (when you pass total_charge) the charge is balanced across the ion and the anion. This is on by default; disable with recover_ionic_metals=False (CLI: --no-ionic-metals). Transition metals and multi-atom metal fragments are still reported unsupported.

Nitro groups are fixed before bond orders are assigned. Left to itself, RDKit's xyz2mol often returns a nitro group as N([O-])[O-], offset by charges it invents elsewhere ([O+], [S+]), or fails outright, depending on the atom order. Any N with three bonds, two of them to terminal O atoms (nitro, nitrate, nitramine), is therefore assigned [N+](=O)[O-], and RDKit perceives the rest of the component with the group replaced by a neutral stand-in. Nitronate anions come out as the carbanion resonance form ([CH2-][N+](=O)[O-]), which has the same InChI. This is on by default (only with charged fragments, not --radicals-instead-of-charges); disable with fix_nitro_groups=False (CLI: --no-nitro-fix).

Install

python -m pip install rdkit
python -m pip install -e .

Conda also works well for RDKit:

conda install -c conda-forge rdkit
python -m pip install -e .

The package installs a xyzrecover console script; you can also run it as a module with python -m xyzrecover.

Charge options apply to every block. --total-charge and --per-fragment-charges are applied to every XYZ block in every input file. Use one charge state per invocation for heterogeneous batches.

CLI examples

Unknown charge, try charge states -3 through +3:

xyzrecover molecule.xyz --json results.json --csv results.csv --sdf results.sdf

Known total charge for each XYZ block:

xyzrecover acetate.xyz --total-charge -1

A salt or cluster with multiple disconnected components and known total charge:

xyzrecover ion_pair.xyz --total-charge 0 --candidate-charges=-2,-1,0,1,2

Known per-component charges. The order follows connectivity perception (the same order as molecule_index in the output, not the input atom order), so run once unconstrained first to see how the structure splits:

xyzrecover fragments.xyz --per-fragment-charges=1,-1,0

If an asserted charge can't be satisfied for a component, fall back to the candidate charge search for that component instead of failing it:

xyzrecover molecule.xyz --total-charge 2 --charge-fallback

Use extended Hueckel connectivity if your RDKit build supports YAeHMOP:

xyzrecover molecule.xyz --use-hueckel --total-charge 1

Python API

from xyzrecover import PerceptionConfig, recover_xyz_file, records_to_json

config = PerceptionConfig(
    total_charge=-1,                  # optional but recommended when known
    candidate_charges=(-2, -1, 0, 1), # used when fragment charge is unknown
    split_fragments=True,
)

records = recover_xyz_file("input.xyz", config)
for rec in records:
    print(rec.molecule_index, rec.formula, rec.charge, rec.smiles, rec.inchikey, rec.confidence)

print(records_to_json(records))

Each record contains:

• status: ok, failed (RDKit could not assign bonds), or unsupported (component contains an element RDKit cannot handle)
• charge: RDKit formal charge for the selected assignment
• smiles: canonical SMILES, with hydrogens made implicit when possible
• explicit_h_smiles: canonical SMILES retaining explicit hydrogens from the XYZ
• inchi and inchikey
• confidence: high, medium, low, or none
• block_index / molecule_index: which XYZ block (frame) and which recovered component the record came from
• atom_indices: indices into the source block's atoms for this component
• alternates: other successful charge/bond-order assignments, when present
• warnings / errors: notes and failed charge attempts, useful for debugging

Multi-frame XYZ

Each XYZ block is processed independently, so a multi-frame file (e.g. a trajectory of one molecule) yields one set of records per frame, distinguished by block_index. De-duplicate on inchikey downstream if you only want unique species.

Recommended workflow

1. Run with known total charge whenever available.
2. Inspect any record with confidence != "high" or non-empty alternates.
3. For unusual bonding, radicals, metals, transition states, hypervalent species, or very close non-covalent contacts, validate against quantum chemistry output or another bond-perception method.
4. Keep the SDF output because it preserves the recovered 3D coordinates and the inferred bond graph.

What to expect (benchmarks)

Recovery rate and confidence depend strongly on the chemistry. Measured across several datasets (COMP6, ANI-1x/ANI-1xm, GDB, peptides, H-bonded and ion-pair complexes):

Practical gating rule: auto-accept status="ok" with confidence="high" and empty alternates; route everything else (medium/low, non-empty alternates, failed, unsupported) to review or a second method.

Development

Install the dev extras (pytest, ruff, mypy) and run the checks:

python -m pip install -e ".[dev]"
python -m pytest        # tests
ruff check .            # lint
ruff format .           # format

CI runs ruff (lint + format check) and pytest on Python 3.10–3.12.

License

MIT — see LICENSE.