Rdkit Conf Agent Skill

RDKit Conformer Generation

This skill provides practical command patterns for RDKit 3D/2D conformer generation

using the standardized CLI wrapper: <skill_path>/scripts/rdkit_conf_helper.py.

Key behaviors (important for Agents):

• The script prints environment detection (Python/RDKit/Pandas) by default.
• Multi-conformer sampling: embeds --num-confs conformers (default 10) per molecule

via EmbedMultipleConfs, optimizes each with the chosen force field, and keeps the

lowest-energy one. Set --num-confs 1 to revert to single-conformer behavior.

• 2D fallback: if all 3D embedding attempts fail, Compute2DCoords is used instead

and a [WARN] line is printed to stderr for that molecule.

• Bad/illegal SMILES are skipped entirely and logged to *.skipped.csv (no crash).
• Molecules that fell back to 2D are additionally logged to *.fallback.csv.
• Each run ends with a summary line and absolute output paths:
• [INFO] Done: <N_3d> 3D, <N_2d> 2D-fallback, <N_skip> skipped (total input: <N>)
• [RESULT] conf_sdf=/abs/path.sdf
• [RESULT] conf_xyz=/abs/path.xyz
• [RESULT] fallback_csv=/abs/path.fallback.csv (only if any 2D fallbacks occurred)
• [RESULT] skipped_csv=/abs/path.skipped.csv (only if any SMILES were skipped)

Quick Start

Check CLI help:

uv run <skill_path>/scripts/rdkit_conf_helper.py --help
uv run <skill_path>/scripts/rdkit_conf_helper.py conf --help

Disable environment printing (optional):

uv run <skill_path>/scripts/rdkit_conf_helper.py --no-env conf --smiles "CCO" --output out.sdf

Core Tasks

1) Generate 3D conformers (SDF output, default)

Single SMILES:

uv run <skill_path>/scripts/rdkit_conf_helper.py conf \
    --smiles "CCO" \
    --output /tmp/CCO.sdf

Single SMILES with a custom molecule name:

uv run <skill_path>/scripts/rdkit_conf_helper.py conf \
    --smiles "c1ccccc1" \
    --name benzene \
    --output /tmp/benzene.sdf

From CSV (default SMILES column: smiles):

uv run <skill_path>/scripts/rdkit_conf_helper.py conf \
    --file data.csv \
    --smiles-col smiles \
    --output data.sdf

From CSV with a name column:

uv run <skill_path>/scripts/rdkit_conf_helper.py conf \
    --file data.csv \
    --smiles-col smiles \
    --name-col compound_id \
    --output data.sdf

From SMI (second token per line is used as name automatically):

uv run <skill_path>/scripts/rdkit_conf_helper.py conf \
    --file molecules.smi \
    --output molecules.sdf

2) Control conformer sampling count

Default (10 conformers sampled, lowest-energy kept):

uv run <skill_path>/scripts/rdkit_conf_helper.py conf \
    --file data.csv --output data.sdf

Single conformer (fastest, least thorough):

uv run <skill_path>/scripts/rdkit_conf_helper.py conf \
    --file data.csv --num-confs 1 --output data.sdf

Increase sampling for flexible or macrocyclic molecules:

uv run <skill_path>/scripts/rdkit_conf_helper.py conf \
    --file data.csv --num-confs 50 --output data.sdf

3) Choose force-field minimization

MMFF94s (default, falls back to UFF if unavailable):

uv run <skill_path>/scripts/rdkit_conf_helper.py conf \
    --file data.csv --ff mmff94s --output data.mmff.sdf

UFF (universal force field):

uv run <skill_path>/scripts/rdkit_conf_helper.py conf \
    --file data.csv --ff uff --output data.uff.sdf

Skip force-field optimization (raw ETKDG geometry only):

uv run <skill_path>/scripts/rdkit_conf_helper.py conf \
    --file data.csv --ff none --output data.etkdg_raw.sdf

4) XYZ output

uv run <skill_path>/scripts/rdkit_conf_helper.py conf \
    --file data.csv \
    --format xyz \
    --output data.xyz

5) Tuning embedding for difficult molecules

Large or macrocyclic molecules sometimes fail standard ETKDG; try random initial coordinates:

uv run <skill_path>/scripts/rdkit_conf_helper.py conf \
    --file macrocycles.csv \
    --use-random-coords \
    --max-attempts 500 \
    --output macrocycles.sdf

Use a different random seed (reproducibility):

uv run <skill_path>/scripts/rdkit_conf_helper.py conf \
    --file data.csv --seed 123 --output data.seed123.sdf

Non-deterministic embedding (seed = -1):

uv run <skill_path>/scripts/rdkit_conf_helper.py conf \
    --file data.csv --seed -1 --output data.sdf

6) Suppress hydrogen addition

By default explicit H atoms are added before embedding for more accurate 3D geometry.

Use --no-hs to keep the molecule as-is (heavy atoms only):

uv run <skill_path>/scripts/rdkit_conf_helper.py conf \
    --file data.csv --no-hs --output data.noh.sdf

7) Custom log file paths

uv run <skill_path>/scripts/rdkit_conf_helper.py conf \
    --file data.csv \
    --output data.sdf \
    --error-log logs/skipped.csv \
    --fallback-log logs/used_2d.csv

______________________________________________________________________

3D Embedding Pipeline Details

For each molecule, the script runs the following steps in order:

• Parse SMILES via Chem.MolFromSmiles.
• Add hydrogens (Chem.AddHs) -- skipped with --no-hs.
• Multi-conformer 3D embedding (EmbedMultipleConfs, --num-confs candidates,

default 10): tries ETKDGv3, then ETKDGv2, then ETDG, then ETDG+useRandomCoords

as a fallback chain until at least one conformer is embedded.

• Force-field minimization (if --ff is not none): each successfully embedded

conformer is individually optimized. MMFF94s transparently falls back to UFF if

parameters are unavailable for that molecule.

• Lowest-energy selection: the conformer with the minimum post-optimization energy

is retained; all others are discarded. If --ff none, the first embedded conformer

is kept without energy ranking.

• 2D fallback (if all 3D attempts yield zero conformers): generates a flat 2D layout

via Compute2DCoords (Z=0 for all atoms), prints a [WARN] to stderr, and records

the molecule in the fallback log.

______________________________________________________________________

Output Format Notes

SDF output (--format sdf, default):

• Standard V2000 multi-molecule SDF, one conformer per molecule.
• Molecule name (from --name, --name-col, or auto-generated mol_<i>) is

written to the SDF header line.

• Compatible with most cheminformatics tools (RDKit, OpenBabel, Schrodinger, etc.).

XYZ output (--format xyz):

• Concatenated XYZ blocks (element, x, y, z per atom).
• Molecule name is written as the comment line (second line of each block).
• Coordinates are in Angstroms.
• Note: if --no-hs is used, hydrogen atoms are absent from the XYZ.

Fallback log (*.fallback.csv):

• Written only when at least one molecule fell back to 2D.
• Columns: idx, smiles, name, dim (always 2), ff (always 2d_fallback), note.

Skipped log (*.skipped.csv):

• Written only when at least one SMILES was skipped.
• Columns: idx, smiles, error.

______________________________________________________________________

Agent Checklist

When using this skill for users:

• Confirm input format:
• .csv requires a SMILES column (default smiles)
• .smi uses the first token per line as SMILES, second token (if present) as name
• Quote SMILES containing special shell characters (brackets/parentheses):
• Example: --smiles "C@@H(F)Cl"
• For CSV workflows, verify column names:
• --smiles-col for the SMILES column
• --name-col (optional) for molecule identifiers to embed in SDF/XYZ headers
• Check the [INFO] Done: summary line for the 3D/2D/skip breakdown.
• If 2D fallbacks occurred, inspect *.fallback.csv:
• Consider --use-random-coords or --max-attempts tuning for the affected SMILES.
• 2D conformers have Z=0 and are not suitable for 3D-based applications (docking, 3D QSAR).
• Always capture absolute output paths:
• Look for [RESULT] ...=/abs/path in stdout.
• If debugging is needed, enable full traceback:
• RDKIT_CONF_HELPER_TRACE=1 uv run <skill_path>/scripts/rdkit_conf_helper.py ...

References

• RDKit conformer generation guide: https://www.rdkit.org/docs/GettingStartedInPython.html#working-with-3d-molecules
• ETKDG paper: Riniker & Landrum, J. Chem. Inf. Model. 2015, 55, 2562
• ETKDGv3: Wang et al., J. Chem. Inf. Model. 2020, 60, 2044