Multitask deep networks with grid featurization achieve improved scoring performance for protein-ligand binding

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Deep learning models enhance scoring in structure-based drug discovery.
• Multitask deep networks outperform single-task networks in pharmaceutical applications.
• Grid featurization integrates inter- and intra-molecular information for protein-ligand complexes.

Purpose of the Study:

• To evaluate three novel multitask deep networks (standard, bypass, progressive) for predicting protein-ligand binding affinities.
• To compare the performance of these networks with AutoDock Vina and MM/GBSA.
• To assess the potential of combining grid featurization with multitask deep networks for improved scoring.

Main Methods:

• Developed and tested three multitask deep learning architectures: standard, bypass, and progressive networks.
• Employed grid featurization to represent protein-ligand complex coordinates as fingerprints.
• Evaluated scoring performance using Pearson correlation coefficient and mean absolute average error.
• Assessed screening ability through re-docking pose analysis and Area Under the Curve (AUC).

Main Results:

• The progressive network with grid featurization achieved the highest Pearson correlation coefficient (0.74) and lowest mean absolute average error (0.98).
• All tested networks improved screening ability for re-docking poses.
• The progressive network demonstrated superior AUC (0.87) compared to AutoDock Vina (0.52).

Conclusions:

• The progressive network, when combined with grid featurization, represents a powerful rescoring approach.
• This method enhances screening results obtained from conventional docking software.
• The findings support the utility of advanced deep learning techniques in structure-based drug discovery.