DOCKSTRING: Easy Molecular Docking Yields Better Benchmarks for Ligand Design.

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning applications in drug discovery
• Bioinformatics and computational biology

Background:

• Machine learning (ML) methods for drug discovery are rapidly advancing.
• Current ML models are often evaluated using basic physicochemical properties, which poorly represent drug design objectives.
• Molecular docking is crucial for estimating binding affinities but requires specialized expertise, limiting its adoption.

Purpose of the Study:

• To introduce dockstring, a comprehensive bundle for comparing ML models in drug discovery using docking scores.
• To enable robust and meaningful benchmarking of ML models by simplifying docking score computation.
• To provide a realistic evaluation framework that better reflects challenges in drug design.

Main Methods:

• Development of an open-source Python package for easy docking score calculation.
• Creation of a large dataset with docking scores and poses for over 260,000 molecules across 58 targets.
• Establishment of benchmark tasks, including virtual screening and de novo design of selective kinase inhibitors.

Main Results:

• The dockstring Python package offers a streamlined protocol for obtaining reliable docking scores, accessible to non-experts.
• The dataset is the first of its kind to include docking poses and a full matrix format, supporting advanced ML techniques.
• Docking scores proved to be a more relevant evaluation metric than physicochemical properties for drug discovery tasks.

Conclusions:

• Docking scores offer a more realistic and challenging benchmark for ML models in drug discovery compared to physicochemical properties.
• The dockstring bundle democratizes the use of molecular docking in ML-driven drug design.
• This work facilitates the development of more effective ML models for identifying novel drug candidates.