From Hype to Practice: Where AI Actually Helps in Drug Discovery - 2025 Canadian UGM Replay

The November 20 broadcast of CDD’s Canadian User Group Meeting panel brought together experts from across the discovery ecosystem to discuss the practical realities of applying AI in drug discovery.

• Jon Stokes, PhD (Stoked Bio) – Academic founder working on AI-guided antibiotic and antifungal discovery
• Erin Davis, PhD (X-Chem) – CTO overseeing the integration of DEL screening and machine learning at scale
• Josh Pottel, PhD (Molecular Forecaster) – CEO building computational chemistry and modeling services for collaborative R&D
• Moderator: James White, PhD (Collaborative Drug Discovery)

Key Themes

1. Collaboration remains the core challenge.
All three panelists emphasized that interdisciplinary communication, not algorithms, defines success. True progress comes from pairing chemists, biologists, and data scientists who are willing to work through early inefficiencies until they share a common technical language.

2. Knowing when AI works—and when it doesn’t.
Machine learning continues to offer value in pattern recognition and prioritization, but panelists cautioned against overreach. Understanding a model’s domain of applicability is critical; using limited or noisy data often leads to false confidence. As Davis noted, “Some of the molecules that come out of generative models are useful—others are chemically impossible.”

3. Data quality outweighs model complexity.
Stokes stressed that the most productive ML scientists spend most of their time validating, cleaning, and understanding their data. “Eighty percent of the work is checking that the data even makes sense before training anything,” he said. Consistent metadata and unified ontologies—such as those supported by CDD Vault—help ensure reproducibility and enable model reuse.

4. Balancing openness and IP.
While academic labs tend to release model parameters and training data, companies must weigh transparency against business constraints. Pottel described how Molecular Forecaster participates in open prediction challenges to learn from community results while maintaining proprietary code internally. “You can share methods and still preserve value,” he noted.

5. Preparing for the next generation of AI tools.
Looking ahead, the speakers pointed to foundation models such as Boltz-2 and high-throughput DEL-based mapping as potential accelerators. Integrating these tools into multi-parameter optimization pipelines—while maintaining synthetic tractability—will define the next phase of discovery informatics.