XGraphBoost

XGraphBoost integrates Graph Neural Networks (Graph Convolutional Networks, Gated Graph Neural Networks, Directed Message Passing Neural Networks) with the XGBoost gradient boosting algorithm to predict molecular properties by extracting graph-based molecular features for classification and regression tasks such as target binding affinities, side effects, and target missing probabilities.

Key Features:

• Graph Neural Network Integration: Leverages GCNs, GGNNs, and DMPNNs to represent molecules as graphs with atoms as nodes and bonds as edges for automatic feature learning.
• XGBoost Algorithm: Uses the XGBoost gradient boosting framework to build ensemble prediction models from extracted features.
• Automatic Feature Extraction: Replaces hand-crafted molecular descriptors by deriving features directly from molecular graph representations via GNNs.
• Scalability and Efficiency: Combines GNN-based representation learning with XGBoost prediction to achieve stable performance while addressing GNN training resource demands.
• Versatility in Applications: Supports both classification and regression tasks for diverse molecular property predictions including binding affinities, side effects, and target missing probabilities.

Scientific Applications:

• Drug discovery and development: Facilitates prioritization of candidate molecules by predicting properties relevant to therapeutic efficacy and safety.
• Molecular screening: Enables screening of compounds with similar characteristics to known drugs using graph-derived feature representations.
• Property prediction: Applies to prediction of target binding affinities, side effects, and target missing probabilities for compound evaluation.

Methodology:

Molecular structures are converted into graph representations, GNNs (GCNs, GGNNs, DMPNNs) extract features from these graphs, and the extracted features are input to an XGBoost classifier/regressor to build the predictive model.