WelQrate: Defining the Gold Standard in Small Molecule Drug Discovery Benchmarking.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Bioinformatics

Background:

• Deep learning has significantly impacted computer-aided drug discovery, yet lacks standardized evaluation practices.
• Current AI efforts in drug discovery are hindered by insufficient benchmarking, slowing real-world application and innovation transfer.

Purpose of the Study:

• To introduce WelQrate, a novel gold standard for benchmarking small molecule drug discovery.
• To address the need for a sound model evaluation framework in AI-driven drug discovery.

Main Methods:

• Developed WelQrate, a curated collection of 9 datasets across 5 therapeutic target classes, using expert-designed hierarchical curation pipelines and domain-driven preprocessing (e.g., Pan-Assay Interference Compounds filtering).
• Proposed a standardized model evaluation framework encompassing high-quality datasets, featurization, 3D conformation generation, metrics, and data splits for reliable virtual screening benchmarking.
• Conducted benchmarking experiments using the WelQrate collection to assess model performance, dataset quality, featurization, and data splitting impacts.

Main Results:

• The WelQrate dataset collection ensures high-quality data through rigorous curation and preprocessing.
• The standardized framework provides a reliable method for evaluating AI models in drug discovery.
• Benchmarking revealed the influence of different models, data quality, featurization, and splitting strategies on performance.

Conclusions:

• Recommends WelQrate as the new gold standard for small molecule drug discovery benchmarking.
• Emphasizes the importance of standardized evaluation frameworks for AI in drug discovery.
• WelQrate datasets, curation codes, and scripts are publicly available to foster reproducible research.