What will it take for AI to change drug discovery? (www.writingruxandrabio.com)

This piece argues that AI alone won’t rescue drug discovery unless it’s paired with better data and regulatory change. The author revisits Eroom’s Law — drug approvals per billion dollars of R&D have roughly halved every nine years — and warns against two self-delusions: equating more hypotheses with progress, and over-relying on model-system data that don’t capture human physiology. Current AI wins in biology (e.g., AlphaFold, antibody design) succeed where inputs, outputs and datasets are dense and well-defined, but these accelerate workflows we already could do. Most programs still fail in humans (~10% succeed) because preclinical models lack predictive validity; clinical development costs exceed $1B, so quality of hypotheses—not quantity—is the bottleneck. Technically, the article stresses we need causal, dynamic, multi-scale in-human data (not just static multi-omics) to train models that truly forecast clinical outcomes. Human genetics is a concrete example: target mechanisms supported by human variants roughly double–triple clinical success rates. Practical implications: pursue regulatory and trial-reform (the “Clinical Trial Abundance” idea) to make interventional, in-human data cheaper and faster, and treat AI as a complement to, not a substitute for, better clinical data. Otherwise AI risks “clogging” pipelines with mid-quality leads and amplifying past mistakes from reductionist approaches.