Navigating Ultralarge Virtual Chemical Spaces with Product-of-Experts Chemical Language Models

Area of Science:

• Computational chemistry
• Drug discovery informatics
• Artificial intelligence in chemistry

Background:

• Ultralarge virtual chemical spaces offer billions of potential drug candidates.
• Current chemical language models struggle to navigate specific spaces and ensure synthetic accessibility.
• Efficient exploration of chemical space is crucial for accelerating drug discovery.

Purpose of the Study:

• To introduce a novel approach for navigating ultralarge virtual chemical spaces using product-of-experts (PoE) chemical language models.
• To enable controlled compound generation with desired properties and high synthetic accessibility.
• To demonstrate the utility of PoE models in drug discovery for targeting specific receptors and properties.

Main Methods:

• Developed a modular and scalable product-of-experts (PoE) framework for chemical language models.
• Combined a prior model pretrained on the target chemical space with expert and anti-expert models.
• Fine-tuned models using external property-specific datasets for targeted compound generation.

Main Results:

• PoE chemical language models successfully generated compounds with desirable properties, including favorable docking to dopamine receptor D2 (DRD2).
• Generated compounds demonstrated predicted ability to cross the blood-brain barrier (BBB).
• The majority of generated compounds remained within the specified target chemical space, ensuring relevance and accessibility.

Conclusions:

• PoE chemical language models offer a powerful and controllable method for navigating and generating compounds within ultralarge virtual chemical spaces.
• This approach significantly enhances the efficiency and success rate of drug discovery by ensuring synthetic accessibility and desired molecular properties.
• The study highlights the potential of advanced AI techniques in revolutionizing the exploration of chemical diversity for therapeutic development.