Swift Dock: A Machine Learning Framework for Accelerating Molecular Docking Calculations Using Neural Networks

Swift Dock is an open-source framework that uses LSTM neural networks and traditional machine learning regression models to predict molecular docking scores. Its core goal is to train models on small-scale samples and then use them to predict docking results for large chemical libraries, thereby significantly accelerating the drug screening process. Key technologies include LSTM with attention mechanisms, XGBoost, and various molecular fingerprint features like MACCS and Morgan.

Molecular docking is a core step in drug development, but traditional methods (e.g., AutoDock, Glide) involve complex physicochemical calculations and conformational searches, leading to high computational costs. For large chemical libraries, this can take weeks or months. Swift Dock addresses this by training ML models on small explicit docking datasets to predict scores for other molecules, avoiding expensive calculations for every candidate.

Swift Dock offers two main workflows:

1. LSTM Neural Network Workflow: Built on PyTorch with attention mechanisms, it handles SMILES sequence data and supports features like MACCS fingerprints, one-hot encoding, and their combinations.
2. Traditional ML Regression Workflow: Includes XGBoost, decision tree regression, and SGD regression, requiring pre-generated .dat fingerprint files. Key molecular features: MACCS (166 pharmacophore features), Morgan (fine-grained local structure), one-hot encoding (SMILES sequence preservation).

Swift Dock provides comprehensive evaluation metrics: R², RMSE, MAE. It also includes analysis tools like SHAP (feature importance), t-SNE (feature space visualization), and Tanimoto similarity analysis (chemical space coverage assessment). These help understand model behavior and optimize strategies.

Swift Dock applies to multiple drug development stages:

• Virtual Screening: Accelerate candidate selection from large libraries.
• Molecular Optimization: Guide structure modifications for better binding affinity.
• ADMET Prediction: Extendable to assess absorption, distribution, metabolism, excretion, and toxicity.
• Target Druggability: Evaluate target potential via model performance.

Limitations: Prediction accuracy depends on training data quality/coverage; may perform poorly on novel structures (suggest active learning to expand datasets). Current focus is on docking score prediction (no conformation generation or binding mode analysis). Future plans: Integrate generative models for de novo molecular design.