MISATO: machine learning dataset of protein-ligand complexes for structure-based drug discovery

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Large language models (LLMs) show promise in biology and chemistry but lack robust datasets for structure-based drug discovery.
• Precise biomolecule-ligand interaction data is crucial for advancing LLMs in scientific applications.

Purpose of the Study:

• To introduce MISATO, a comprehensive dataset designed to support the development of advanced AI models for drug discovery.
• To provide high-quality, validated data combining quantum mechanical properties and molecular dynamics simulations of protein-ligand complexes.

Main Methods:

• Utilized semi-empirical quantum mechanics to refine ~20,000 experimental protein-ligand complex structures.
• Generated extensive molecular dynamics simulations of protein-ligand complexes in explicit water, totaling over 170 μs.
• Validated experimental data to ensure the reliability of the dataset.

Main Results:

• The MISATO dataset integrates quantum mechanical properties with molecular dynamics simulations for a large number of experimental protein-ligand complexes.
• Demonstrated improved accuracy of baseline machine learning models when trained on the MISATO dataset.
• Established a valuable resource for advancing artificial intelligence in drug discovery.

Conclusions:

• MISATO addresses the critical need for precise biomolecule-ligand interaction data for LLMs.
• The dataset facilitates the creation of next-generation AI models for structure-based drug discovery.
• Provides an accessible entry point for machine learning experts to contribute to drug discovery AI.