Large-Scale Pretraining Improves Sample Efficiency of Active Learning-Based Virtual Screening.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Virtual screening is crucial for identifying drug candidates from large compound libraries.
• Active learning and Bayesian optimization enhance virtual screening efficiency.
• Surrogate machine learning models are essential for predicting compound properties.

Purpose of the Study:

• To evaluate pretrained transformer-based language models and graph neural networks within a Bayesian optimization active learning framework.
• To assess the performance of these models in accelerating the identification of hit compounds.

Main Methods:

• Utilized a Bayesian optimization active learning framework.
• Employed pretrained transformer-based language models and graph neural networks as surrogate models.
• Benchmarked performance on an ultralarge compound library (99.5 million compounds).

Main Results:

• The best pretrained model identified 58.97% of the top 50,000 compounds after screening only 0.6% of the library.
• Achieved an 8% improvement over the previous state-of-the-art baseline.
• Demonstrated superior performance in both structure-based and ligand-based drug discovery scenarios.

Conclusions:

• Pretrained models significantly boost accuracy and sample efficiency in active learning-based virtual screening.
• These models offer a powerful approach to accelerate hit identification in drug discovery.
• The findings suggest pretrained models are valuable tools for navigating vast chemical spaces.