DeepDDS

DeepDDS predicts synergistic drug combinations in cancer by applying a graph neural network with an attention mechanism to integrate drug molecular structure and cancer cell gene expression data.

Key Features:

• Graph Neural Network (GNN): Employs a graph neural network to model drug molecular structures and predict interactions affecting cancer cell viability.
• Attention Mechanism / Graph Attention Network: Integrates an attention mechanism to focus on relevant atomic and substructure features within molecular graphs to improve prediction accuracy.
• Feature Embeddings: Uses embeddings derived from drug molecule structures and gene expression profiles as model inputs.
• Multilayer Feedforward Neural Network: Processes feature embeddings through a multilayer feedforward neural network to identify synergistic drug combinations.
• Interpretability via Atomic Feature Correlation: Analyzes the correlation matrix of atomic features within the graph attention network to reveal chemical substructures that contribute to predicted synergy.
• Benchmarking and Performance: Evaluated against classical machine learning methods and other deep learning approaches on benchmark datasets and reported over 16% improvement on an independent AstraZeneca dataset.

Scientific Applications:

• Computational screening of drug combinations: Predicts synergistic drug pairs to prioritize candidates for experimental validation in cancer studies.
• Prioritization for experimental validation: Ranks combinations that may effectively inhibit specific cancer cells for downstream wet-lab testing.
• Oncology combination therapy discovery: Supports identification and optimization of combination therapies in cancer research.

Methodology:

Integrates a graph neural network and attention mechanism to process embeddings from drug molecular structures and gene expression profiles, feeds these embeddings into a multilayer feedforward neural network, evaluates performance versus classical machine learning and other deep learning methods on benchmark datasets, tests on an independent AstraZeneca dataset, and inspects a correlation matrix of atomic features within the graph attention network for interpretability.