LIT-PCBA: An Unbiased Data Set for Machine Learning and Virtual Screening.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Traditional virtual screening datasets (e.g., DUD, DUD-E, MUV) exhibit chemical biases, leading to overestimated accuracy.
• These biases hinder the reliable evaluation of virtual screening methods and machine learning models.

Purpose of the Study:

• To develop a novel, unbiased dataset (LIT-PCBA) for rigorous benchmarking of virtual screening and machine learning algorithms.
• To create a dataset that minimizes chemical biases and better reflects real-world screening scenarios.

Main Methods:

• Curated 149 PubChem bioassays, removing false positives and assay artifacts.
• Selected 15 protein targets with available X-ray structures and validated screening performance using orthogonal methods.
• Applied asymmetric validation embedding (AVE) to create unbiased training and validation sets.

Main Results:

• The LIT-PCBA dataset comprises 15 targets with 7,844 active and 407,381 inactive compounds.
• It mimics experimental screening decks in hit rate and potency distribution.
• Preliminary screening showed at least one method enriched true actives by a factor of 2 in the top 1%.

Conclusions:

• LIT-PCBA provides a robust, unbiased benchmark for evaluating virtual screening and machine learning methods.
• This dataset will facilitate more accurate assessments and advancements in computational drug discovery.
• The dataset is publicly available for community use in method development and validation.