DeepMol-WorkBuddy-skill: Extending Computational Chemistry Capabilities for Desktop AI Agents

This article introduces the DeepMol-WorkBuddy-skill project, which encapsulates the DeepMol computational chemistry framework into a skill package for WorkBuddy desktop AI agents. It addresses the implementation challenges of AI for Science in the computational chemistry field, allowing researchers without programming backgrounds to complete end-to-end molecular machine learning workflows (including data loading, feature engineering, model training, and interpretability analysis) via natural language interaction.

AI for Science has made significant progress in computational chemistry and drug discovery, but its implementation faces three major obstacles: 1. High technical barriers (requiring mastery of Python, RDKit, deep learning frameworks, etc.); 2. Complex workflows (including multiple steps like data preprocessing and feature extraction); 3. Fragmented tools (different libraries needed for different tasks). The DeepMol-WorkBuddy-skill project aims to lower the usage threshold by encapsulating DeepMol into an AI agent skill package.

DeepMol-WorkBuddy-skill is a skill package for WorkBuddy desktop AI agents, with its core being the transformation of DeepMol framework capabilities into AI-executable process knowledge. DeepMol is built on libraries like RDKit, TensorFlow/PyTorch, scikit-learn, and DeepChem, providing high-level APIs. The skill package includes: 1. SKILL.md (an 11-step workflow guide); 2. API reference documentation; 3. Practical CLI scripts (environment verification, tutorials, QSAR pipelines, etc.).

The skill package equips WorkBuddy with the following capabilities: data loading and preprocessing (supports CSV/SDF and large datasets), compound standardization (multiple strategies), molecular feature engineering (Morgan fingerprints and other descriptors), feature selection and dimensionality reduction (PCA/t-SNE, etc.), data splitting (random/stratified/K-fold/scaffold splitting), model building (traditional ML/deep learning/DeepChem models), hyperparameter optimization (grid/random search + Optuna), model interpretability (SHAP visualization), imbalanced data handling (SMOTE, etc.), and pipelines & AutoML (end-to-end pipelines + automatic optimization).

Environment requirements: Python ≥3.13, DeepMol 1.2.1. Installation requires bypassing scikeras dependency issues, with steps including core library installation, Torch installation, etc. Verification command: python scripts/check_install.py. Usage examples: QSAR modeling pipeline, batch feature engineering, ADMET prediction (pre-trained models). After integrating with WorkBuddy, natural language interaction is possible (e.g., "Train a QSAR model").

The skill package has undergone practical validation: the test environment includes DeepMol 1.2.1, Python3.14, and Windows11; all scripts run successfully; API documentation accurately reflects actual behavior, avoiding toolchain instability issues and ensuring research efficiency.

Advantages of this model: 1. Lower the threshold for using professional tools (natural language interaction); 2. Ensure execution of best practices (AI operates according to expert workflows); 3. New paradigm of human-machine collaboration (researchers define problems, AI executes details); 4. Ensure reproducibility (standardized pipelines + save/load mechanisms).

DeepMol-WorkBuddy-skill demonstrates an important direction for AI for Science: encapsulating complex scientific tools into AI skill packages to lower technical barriers and improve research efficiency. In the future, more similar skill packages covering fields like bioinformatics and materials science are expected to emerge.