Unimol Agent Skill

Uni-Mol

This skill provides practical command patterns for Uni-Mol molecular representation / training / prediction using the standardized CLI wrapper: <skill_path>/scripts/unimol_helper.py.

Key behaviors (important for Agents):

• The script prints environment detection (Python/Torch/CUDA) by default.
• Bad/illegal SMILES are skipped and logged to *.skipped.csv (no crash).
• Each run ends by printing absolute output paths like:
• [RESULT] repr_npy=/abs/path.npy
• [RESULT] model_dir=/abs/model_dir
• [RESULT] pred_csv=/abs/pred.csv

Quick Start

Check CLI help:

uv run python <skill_path>/scripts/unimol_helper.py --help

Check subcommand help:

uv run python <skill_path>/scripts/unimol_helper.py repr --help
uv run python <skill_path>/scripts/unimol_helper.py train --help
uv run python <skill_path>/scripts/unimol_helper.py predict --help

Disable environment printing (optional):

uv run python <skill_path>/scripts/unimol_helper.py --no-env repr --smiles "CCO" --output out.npy

Core Tasks

1) Extract molecular representations (embedding) to .npy

Single SMILES:

uv run python <skill_path>/scripts/unimol_helper.py repr \
    --smiles "CCO" \
    --output /tmp/ccO.repr.npy

From CSV (default SMILES column is smiles):

uv run python <skill_path>/scripts/unimol_helper.py repr \
    --file data.csv \
    --smiles-col smiles \
    --output data.repr.npy

From SMI:

uv run python <skill_path>/scripts/unimol_helper.py repr \
    --file molecules.smi \
    --output molecules.repr.npy

Force CPU / GPU:

# Force CPU
uv run python <skill_path>/scripts/unimol_helper.py repr --smiles "CCO" --no-gpu --output out.npy

# Force GPU (will warn & fall back if CUDA is unavailable)
uv run python <skill_path>/scripts/unimol_helper.py repr --smiles "CCO" --use-gpu --output out.npy

2) Train a property model (classification / regression / multilabel\_\*)

Regression training (CSV must contain smiles and target columns):

uv run python <skill_path>/scripts/unimol_helper.py train \
    --task regression \
    --input train.csv \
    --smiles-col smiles \
    --target-col target \
    --epochs 50 \
    --output ./model_reg

Classification training:

uv run python <skill_path>/scripts/unimol_helper.py train \
    --task classification \
    --input train.csv \
    --smiles-col smiles \
    --target-col target \
    --epochs 50 \
    --output ./model_cls

Multilabel regression training (explicit multi-target columns):

uv run python <skill_path>/scripts/unimol_helper.py train \
    --task multilabel_regression \
    --input train.csv \
    --smiles-col smiles \
    --target-cols target_0,target_1,target_2 \
    --epochs 50 \
    --output ./model_mreg

Multilabel classification training:

uv run python <skill_path>/scripts/unimol_helper.py train \
    --task multilabel_classification \
    --input train.csv \
    --smiles-col smiles \
    --target-cols y_cls_0,y_cls_1,y_cls_2 \
    --epochs 50 \
    --output ./model_mcls

Target recognition for training:

• Single-task (classification / regression): use --target-col (default target).
• Multilabel tasks: prefer --target-cols (comma-separated).
• If --target-cols is omitted for multilabel tasks, the helper auto-detects columns named target or prefixed with target_ (case-insensitive).

Force CPU:

uv run python <skill_path>/scripts/unimol_helper.py train \
    --task regression \
    --input train.csv \
    --epochs 50 \
    --output ./model_cpu \
    --no-cuda

3) Predict properties to .csv

Predict from CSV:

uv run python <skill_path>/scripts/unimol_helper.py predict \
    --model ./model_reg \
    --input test.csv \
    --smiles-col smiles \
    --output pred.csv

Predict from SMI:

uv run python <skill_path>/scripts/unimol_helper.py predict \
    --model ./model_reg \
    --input test.smi \
    --output pred.csv

Notes:

• Output CSV contains the input rows (for valid SMILES) plus pred / pred_* columns.
• If there are bad SMILES, they are skipped and saved to pred.csv.skipped.csv (or your --error-log path).

Agent Checklist

When using this skill for users:

• Confirm input format:
• .csv requires a SMILES column (default smiles)
• .smi uses the first token of each line as SMILES
• Quote SMILES containing special characters (brackets/parentheses):
• Example: --smiles "C([H])[H]"
• For CSV workflows, verify column names:
• repr: --smiles-col
• train: --smiles-col and --target-col / --target-cols
• predict: --smiles-col
• Watch for skipped SMILES:
• Check *.skipped.csv and decide whether to fix or permanently drop them
• Always capture absolute output paths:
• Look for [RESULT] ...=/abs/path in stdout
• If debugging is needed, enable full traceback:
• UNIMOL_HELPER_TRACE=1 uv run python <skill_path>/scripts/unimol_helper.py ...

References

• Uni-Mol project: https://github.com/fanxiaoyu0/Uni-Mol
• RDKit: https://www.rdkit.org/