The Need

Discovering molecules that simultaneously satisfy multiple competing design criteria is a resource-intensive challenge across the pharmaceutical, energy, and materials industries. The enormity of chemical space makes exhaustive screening impractical, while existing AI-guided methods either restrict search to predefined compound libraries or rely on tightly coupled architectures that limit flexibility and scalability. A more modular, sample-efficient approach to navigating complex multi-property trade-offs is needed to accelerate molecular discovery.

The Technology

OSU engineers have developed a a modular, two-stage "generate-then-optimize" framework for discovering novel molecules with optimized multi-property profiles. In the first stage, any generative model is used to propose a large, diverse pool of candidate structures. In the second stage, a novel acquisition function selects the most promising candidates for evaluation using a probabilistic, uncertainty-aware approach that enables efficient and scalable batch selection. The decoupled architecture allows each component to be independently updated or swapped without redesigning the overall pipeline.

Commercial Applications

• Drug discovery
• Energy storage
• Catalysis
• Advanced materials

Benefits/Advantages

• Modular and generator-agnostic: Compatible with any generative model type, enabling seamless integration into existing molecular design workflows without architectural redesign.
• Superior sample efficiency: Consistently outperforms state-of-the-art methods, identifying higher-quality multi-property candidates with fewer costly evaluations or experiments.
• Scalable batch selection: Efficiently handles large evaluation batches with no combinatorial overhead, making it directly compatible with high-throughput experimental platforms.
• Unrestricted chemical space exploration: Proposes truly novel structures beyond the limitations of fixed compound libraries, enabling discovery of previously unconsidered candidates.