Papyrus: a large-scale curated dataset aimed at bioactivity predictions

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning in drug discovery

Background:

• Publicly available ligand-protein bioactivity data is rapidly growing, presenting opportunities for machine learning.
• Data quality and accessibility challenges hinder the effective use of this data for research.
• Researchers spend significant time adapting and identifying suitable datasets for their specific needs.

Purpose of the Study:

• To construct a comprehensive and standardized dataset for machine learning in drug discovery.
• To address the challenges of data heterogeneity, quality, and accessibility in bioactivity data.
• To provide a benchmark dataset for developing predictive models and facilitating research.

Main Methods:

• Aggregation of multiple large public datasets (e.g., ChEMBL, ExCAPE-DB) with smaller, high-quality datasets.
• Standardization and normalization of approximately 60 million data points for machine learning suitability.
• Demonstration of data filtering techniques and application in quantitative structure-activity relationship (QSAR) and proteochemometric modeling.

Main Results:

• Creation of the Papyrus dataset, comprising ~60 million standardized ligand-protein bioactivity data points.
• Successful integration and harmonization of diverse public bioactivity datasets.
• Illustrative examples of data filtering, QSAR, and proteochemometric modeling using the Papyrus dataset.

Conclusions:

• The Papyrus dataset serves as a valuable, accessible benchmark for constructing predictive models in drug discovery.
• Standardized and curated bioactivity data significantly enhances the efficiency of machine learning applications.
• This resource aims to accelerate research by providing a reliable foundation for computational studies.