General Purpose Structure-Based Drug Discovery Neural Network Score Functions with Human-Interpretable Pharmacophore...

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• The BioChemical Library (BCL) is an open-source cheminformatics toolkit.
• Existing BCL tools focus on ligand-based virtual high-throughput screening (vHTS).

Purpose of the Study:

• To expand the BCL capabilities to include structure-based virtual screening.
• To introduce novel scoring functions for protein-ligand interactions.
• To enhance small molecule discovery and design.

Main Methods:

• Developed two new neural network-based scoring functions: BCL-AffinityNet and BCL-DockANNScore.
• Utilized novel distance-dependent signed protein-ligand atomic property correlations.
• Trained models using feed-forward dropout neural networks.

Main Results:

• BCL-AffinityNet demonstrated top performance in affinity prediction and ranking tasks (CASF-2016).
• BCL-AffinityNet showed strong performance on CSAR-NRC HiQ I and II datasets.
• BCL-DockANNScore proved competitive with state-of-the-art methods in docking and screening power.
• Models were decomposable into interpretable pharmacophore maps.

Conclusions:

• The expanded BCL toolkit effectively supports structure-based virtual screening.
• New scoring functions significantly aid in small molecule discovery and lead optimization.
• BCL-AffinityNet and BCL-DockANNScore are valuable additions to computational drug design.