HydraScreen: A Generalizable Structure-Based Deep Learning Approach to Drug Discovery.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Accelerated drug discovery requires robust computational tools.
• Accurate prediction of protein-ligand interactions is crucial for structure-based drug design.
• Existing methods may face challenges with bias and generalization.

Purpose of the Study:

• To introduce HydraScreen, a deep learning framework for safe and robust accelerated drug discovery.
• To develop an end-to-end pipeline for high-throughput screening and lead optimization.
• To enhance the interpretability and impartiality of machine learning models in drug discovery.

Main Methods:

• Utilized a state-of-the-art 3D convolutional neural network for molecular structure and interaction representation.
• Implemented an end-to-end pipeline for high-throughput screening and lead optimization.
• Developed a novel interaction profiling approach to detect model and data biases.

Main Results:

• Achieved top-tier results on CASF-2016 benchmarks for affinity and pose prediction (Pearson's r = 0.86, RMSE = 1.15, Top-1 = 0.95).
• Demonstrated effective generalization across novel proteins and ligands via temporal split.
• Interaction profiling enhanced interpretability and reinforced model impartiality.

Conclusions:

• HydraScreen offers a safe and robust framework for accelerated drug discovery.
• The framework shows strong performance in predicting protein-ligand binding.
• Future work can focus on enhancing the robustness of machine learning scoring functions.