MolData, a molecular benchmark for disease and target based machine learning.

Area of Science:

• Computational drug discovery
• Molecular machine learning
• Artificial intelligence in medicine

Background:

• Deep learning advances feature extraction in computational drug discovery.
• Existing databases like PubChem and ChEMBL have complex bioassay descriptions hindering machine learning application.
• Biological and chemical knowledge is crucial for effective data handling in drug discovery.

Purpose of the Study:

• To create a comprehensive, disease- and target-based dataset from PubChem for machine learning in drug discovery.
• To facilitate and accelerate molecular machine learning by providing curated bioassay data.
• To democratize access to molecular machine learning resources.

Main Methods:

• Collected and curated a large-scale dataset (MolData) from PubChem, linking screening results to specific diseases and targets.
• Organized data into 30 unique categories of targets and diseases.
• Performed correlation analysis on bioassays and benchmarked over 30 machine learning models using multitask learning.

Main Results:

• Developed MolData, containing approximately 170 million drug screening results from 1.4 million unique molecules.
• Identified valuable information for drug repurposing across diseases like cancer, metabolic disorders, and infectious diseases.
• Established a benchmark for over 30 models trained on disease and target categories.

Conclusions:

• MolData significantly advances computational drug discovery by providing a large, accessible, and well-annotated dataset.
• The dataset and benchmark models accelerate molecular artificial intelligence development for practical drug discovery applications.
• Facilitates drug repurposing research by revealing correlations between molecular screening data and diseases.