Cohort: 1 

Locations: San Francisco & Palo Alto, United States 

Application Title: AI-enabled discovery framework for neurodegenerative therapeutics 

Team Members: Stanley Prusiner (executive sponsor), Jack Hidary (executive sponsor), Gregory Merz, Nadia Harhen, Andrea Bortolato, Ben Shields, Kelly Richdale

The challenge 

Developing new drug therapies is slow, costly and uncertain. On average, bringing a candidate from early research to clinical use takes more than a decade, with most candidates failing along the way. These challenges are especially acute for neurodegenerative diseases such as Alzheimer’s and Parkinson’s, where disease mechanisms are not fully understood and animal models often fail to predict human outcomes. 

Traditional drug discovery relies heavily on experimental screening. Researchers test hundreds of thousands of molecules in the laboratory to identify a small number of promising leads. Even with modern automation, this process is resource-intensive and constrained by the availability of experimental data. When data is limited or noisy, AI models trained only on historical results struggle to generalise and rarely uncover novel candidates. 

As a result, exploration of chemical space remains narrow, and many potentially useful compounds are never examined. 

The solution 

Physics-guided AI for data-scarce discovery 

The UCSF Institute for Neurodegenerative Diseases partnered with SandboxAQ to develop a discovery framework that combines AI with physics-based simulation. 

Rather than relying only on existing experimental labels, the system uses quantum chemistry and molecular physics models to simulate how molecules interact with biological targets. These simulations generate high-quality signals that guide AI ranking models, effectively creating reliable training data where experimental results are scarce. 

This approach allows the AI to evaluate many more compounds while maintaining scientific grounding, even when real-world data is limited. 

How it works 

Simulating molecular interactions at scale 

The framework uses large quantitative models to estimate how strongly a molecule might bind to a target protein, a key factor in drug effectiveness. These simulations act as a filter, allowing the system to score and rank compounds before laboratory testing begins. 

By combining simulation outputs with AI-based ranking, the system can explore millions of compounds computationally. Researchers then select a much smaller set of high-quality candidates for physical experiments. 

In practice, this shifts much of the early discovery effort from wet-lab trial cycles to in-silico evaluation, reducing time and experimental burden. 

How it differs from conventional approaches 

Reducing reliance on brute-force screening 

Conventional discovery pipelines depend on repeated laboratory testing to compensate for limited predictive insight. AI models trained only on past experiments tend to reinforce existing patterns and struggle to identify novel chemistry. 

The UCSF–SandboxAQ approach integrates physical understanding directly into AI workflows. Simulations provide structured guidance, allowing the system to reason about molecular behaviour rather than extrapolate only from prior data. This enables broader exploration while preserving interpretability. 

Laboratory testing remains essential, but it is applied more selectively and later in the process. 

Impact 

• 5.6 million compounds screened computationally within weeks 
• 36× reduction in experimental effort compared with traditional workflows 
• 30× higher hit rates relative to conventional screening approaches 

Why it matters 

Expanding the search for new therapies 

For diseases where biological insight is limited and failure rates are high, reducing uncertainty early in discovery is critical. The collaboration between UCSF and SandboxAQ shows how physics-guided AI can help researchers explore chemical space more efficiently while lowering experimental cost. 

By shifting early discovery toward simulation-informed evaluation, the framework supports faster progress toward viable therapeutic candidates without increasing risk or reliance on exhaustive laboratory screening.