OpenSkillIndex← back to search
claude-code

Computational-chemistry-agent-skills unimol

install
source · Clone the upstream repo
git clone https://github.com/jinzhezenggroup/computational-chemistry-agent-skills
Claude Code · Install into ~/.claude/skills/
T=$(mktemp -d) && git clone --depth=1 https://github.com/jinzhezenggroup/computational-chemistry-agent-skills "$T" && mkdir -p ~/.claude/skills && cp -r "$T/molecular-representation/unimol" ~/.claude/skills/jinzhezenggroup-computational-chemistry-agent-skills-unimol && rm -rf "$T"
manifest: molecular-representation/unimol/SKILL.md
source content

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:

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

References