MF-PCBA

MF-PCBA aggregates 60 multifidelity datasets from PubChem BioAssay (PCBA), combining primary and confirmatory high-throughput screening (HTS) modalities to enable molecular representation learning and multifidelity integration for drug discovery.

Key Features:

• Dataset collection: Sixty multifidelity datasets derived from PubChem BioAssay (PCBA).
• Data modalities: Each dataset contains primary screening and confirmatory screening modalities reflecting conventional HTS workflows.
• Scale: The resource comprises over 16.6 million unique molecule-protein interactions.
• Multifidelity focus: Explicit support for integrating low-fidelity (primary) and high-fidelity (confirmatory) measurements.
• Molecular representation learning: Enables tasks that combine low- and high-fidelity data for molecular representation learning.
• Data curation: Assembly includes rigorous filtering processes to ensure data quality and relevance.
• Method evaluation: Includes evaluation of recent deep-learning-based methods for multifidelity integration.

Scientific Applications:

• Drug activity prediction: Improve prediction of compound activity by leveraging multifidelity HTS measurements.
• Experimental design optimization: Inform more efficient follow-up and confirmatory screening strategies.
• Machine learning benchmarking: Serve as a benchmark for developing and evaluating multifidelity integration and molecular representation learning methods.
• Computational and experimental drug discovery: Provide a large-scale resource for computational analyses and experimental planning in drug discovery.

Methodology:

Data acquisition from PubChem BioAssay (PCBA), pairing of primary and confirmatory screening modalities, rigorous filtering for quality and relevance, and evaluation of deep-learning-based multifidelity integration methods.