LIT-PCBA

LIT-PCBA provides a rigorously curated dataset of PubChem-derived dose-response bioassays for unbiased benchmarking of virtual screening and machine learning methods in drug discovery.

Key Features:

• Unbiased Data Compilation: The dataset comprises 15 target sets derived from 149 dose-response PubChem bioassays and contains 7,844 confirmed active and 407,381 confirmed inactive compounds.
• Rigorous Data Curation: False positives and assay artifacts were removed and active and inactive compounds were balanced within similar molecular property ranges to minimize chemical biases.
• Target Selection for Versatility: Target sets were selected based on the availability of at least one X-ray structure in complex with ligands of the same phenotype as PubChem active compounds, supporting ligand-based and structure-based virtual screening.
• Validation through Virtual Screening Methods: Preliminary screenings using 2D fingerprint similarity, 3D shape similarity, and molecular docking identified target sets where at least one method enriched the top 1% of ranked compounds in true actives by a factor of two.
• Asymmetric Validation Embedding (AVE): AVE was applied to further reduce biases and ensure representative separation between training and validation ligand sets.

Scientific Applications:

• Benchmarking virtual screening methods: Provide realistic, unbiased test sets for comparing the performance of 2D, 3D, and docking-based virtual screening algorithms.
• Machine-learning model development and evaluation: Train and evaluate ML classifiers and regression models for compound activity prediction on balanced, curated actives/inactives.
• Comparative assessment of ligand- and structure-based approaches: Enable direct comparison of ligand-based (fingerprint/shape) and structure-based (docking) screening performance using targets with X-ray ligand complexes.

Methodology:

Derived from 149 dose-response PubChem bioassays into 15 target sets; selected targets requiring at least one X-ray ligand complex of the same phenotype; removed false positives and assay artifacts and balanced actives/inactives by molecular property ranges; validated with preliminary virtual screenings using 2D fingerprint similarity, 3D shape similarity, and molecular docking; applied Asymmetric Validation Embedding (AVE) to reduce bias.