Uni-Mol Tools - Molecular Property Prediction CLI

Package: cli-anything-unimol-tools Command: python3 -m cli_anything.unimol_tools

Description

Interactive CLI for training and inference of molecular property prediction models using Uni-Mol Tools. Supports 5 task types: binary classification, regression, multiclass, multilabel classification, and multilabel regression.

Key Features

• Project Management: Organize experiments with named projects
• 5 Task Types: Classification, regression, multiclass, multilabel variants
• Model Tracking: Automatic performance history and rankings
• Smart Storage: Analyze usage and clean up underperformers
• JSON API: Full automation support with --json flag

Common Commands

Project Management

# Create a new project
project create --name drug_discovery

# List all projects
project list

# Switch to a project
project switch --name drug_discovery

Training

# Train a classification model
train --data-path train.csv --target-col active --task-type classification --epochs 10

# Train a regression model
train --data-path train.csv --target-col affinity --task-type regression --epochs 10

Model Management

# List all trained models
models list

# Show model details and performance
models show --model-id <id>

# Rank models by performance
models rank

Storage & Cleanup

# Analyze storage usage
storage analyze

# Automatic cleanup of poor performers
cleanup auto

# Manual cleanup with criteria
cleanup manual --max-models 10 --min-score 0.7

Prediction

# Make predictions with a trained model
predict --model-id <id> --data-path test.csv

Data Format

CSV files must contain:

• SMILES column: Molecular structures in SMILES format
• Target column(s): Values to predict (name specified via --target-col)

Example:

SMILES,target
CCO,1
CCCO,0
CC(C)O,1

Task Types

1. classification: Binary classification (0/1)
2. regression: Continuous value prediction
3. multiclass: Multiple class classification
4. multilabel_classification: Multiple binary labels
5. multilabel_regression: Multiple continuous values

JSON Mode

Add --json flag to any command for machine-readable output:

python3 -m cli_anything.unimol_tools --json models list

Output format:

{
  "status": "success",
  "data": [...],
  "message": "..."
}

Interactive Mode

Launch without commands for interactive REPL:

python3 -m cli_anything.unimol_tools

Features:

• Tab completion
• Command history
• Contextual help
• Project state persistence

Test Data

Example datasets available at: https://github.com/545487677/CLI-Anything-unimol-tools/tree/main/unimol_tools/examples

Includes data for all 5 task types.

Requirements

• Python 3.8+
• PyTorch 1.12+
• Uni-Mol Tools backend
• 4GB+ RAM (8GB+ recommended for training)

Installation

cd unimol_tools/agent-harness
pip install -e .

Documentation

• SOP: UNIMOL_TOOLS.md
• Quick Start: docs/guides/02-QUICK-START.md
• Full Documentation: docs/README.md

Testing

cd docs/test
bash run_tests.sh --unit -v    # Unit tests (67 tests)
bash run_tests.sh --full -v    # Full test suite

Performance Tips

• Start with 10 epochs for initial experiments
• Use smaller batch sizes if memory is limited
• Monitor storage with storage analyze
• Use models rank to identify best performers
• Clean up regularly with cleanup auto

Troubleshooting

• CUDA errors: Reduce batch size or use CPU mode
• CSV not recognized: Verify SMILES column exists
• Low accuracy: Try more epochs or adjust learning rate
• Storage full: Run cleanup auto to free space

Related

• Uni-Mol Tools: https://github.com/dptech-corp/Uni-Mol/tree/main/unimol_tools
• Uni-Mol Paper: https://arxiv.org/abs/2209.11126
• CLI-Anything: https://github.com/HKUDS/CLI-Anything