Cytiva exec warns AI drug discovery hitting 'data wall' from biased datasets

Drug discovery costs have doubled every nine years since the 1950s (Eroom's Law), with each new drug taking 10–15 years and costing $1–$2.5 billion, facing failure rates above 90%. AI offers hope by shifting from empirical screening to predictive design—modeling molecular interactions before physical testing. Paul Belcher, director of protein research strategy at Cytiva, notes AI can design drug candidates from scratch and eliminate low-quality leads early, saving time and resources. However, AI cannot yet reliably predict kinetics or developability, so every candidate still requires lab validation. This places new pressure on lab teams to characterize more diverse, AI-generated compounds with information-rich, high-throughput methods.

Belcher identifies a critical bottleneck: training data. Most AI models rely on public datasets that only contain positive results, creating a 'data wall' where models converge on similar outputs with diminishing returns. 'No one wants to share their failures,' he says, calling for a 'journal of negative data' to capture failed experiments and non-binding compounds. This publication bias limits AI's ability to learn from failure, making predictions less reliable. To break through, drug companies need proprietary, structured data that includes both successes and failures, and lab systems must evolve to handle the increased throughput demands from AI-driven discovery.