About This Episode
This episode of ByteSight features Dr. Paul Agapow, a scientist-turned-industry AI specialist who now works across AI, statistics, simulation, and decision-making in biotech and pharma with a focus on helping drugs reach the market faster. His background spans everything from academic research to real-world clinical trial challenges, with a refreshingly grounded view of what AI can (and can’t) do.
This conversation explores a deceptively simple question: What does it take to get AI from pilot to real clinical impact? From biased datasets and misleading correlations to „last mile” adoption barriers in clinical trials, Paul and Manasi dig into why AI success is as much about people, context, and trust as it is about models and performance metrics.
About Dr. Paul Agapow
Dr. Paul Agapow bridges the worlds of biology and artificial intelligence to transform how medicines are discovered and developed. An immunologist turned AI leader and strategist, he has led teams at Imperial College, AstraZeneca and GSK, bringing together data, computation, and biology to accelerate drug development and reduce risk. Over his career, he has brought this hybrid approach to optimising clinical trials, developing digital biomarkers, and identifying patient subpopulations, always with a focus on reducing uncertainty and improving decision-making in complex biomedical systems.
The Wolf–Husky Problem in Biomedicine
Paul introduces a classic AI failure mode: a model trained to tell wolves from huskies ends up learning „snow in the background“ rather than the animal. The uncomfortable lesson is that models can deliver convincing outputs while learning incidental shortcuts.
As Paul puts it:
I would hold that biomedicine is just full of wolf-husky problems.
In medical data, shortcuts can hide everywhere, including site effects, scanning position, population selection, hospital workflows, and countless „minor details” that models detect faster than humans. The episode shows why data context and careful validation matter as much as model architecture.
Clinical Trials and the “Last Mile” of AI
The discussion moves into the reality of drug development:
Nothing gets into proper medical use unless it passes through the eye of the needle of a clinical trial.
Paul explains why clinical trials are complex, expensive, and often chaotic and how AI can help, from patient finding in health records to trial simulation („digital twins“), better monitoring, and smarter subpopulation analysis. But he also emphasizes the real bottleneck: the „last mile”, which is to get tools into workflows where users aren’t engineers, and where outcomes are high-stakes.
Final Takeaway
AI won’t deliver impact just because it is powerful. It needs the foundations: good data, correct assumptions, explainability, and trust, especially in healthcare where the consequences of being wrong are real. Paul’s message is practical: build systems that are understandable enough to be trusted, useful enough to be adopted, and robust enough to work beyond the „easy” datasets because without adoption, there is no impact.
Learn more about Dr. Paul Agapow
