Clinical failure rates over the decades: yikes

Most people accepted the core number and rejected the article’s comparison to cars or airplanes. The prevailing view was that clinical trials are not the equivalent of products rolling off an assembly line. They are the final and most expensive test of whether a finished molecule actually works in humans. That distinction matters because the clinical-stage failure rate sits on top of a much larger invisible funnel of rejected compounds upstream. Several commenters said the stable failure rate is not obviously evidence of stagnation. It can also mean better tools are being used to pursue harder targets and narrower improvements over an already strong standard of care. In oncology, immunology, and other mature areas, a drug can be clinically valuable yet still fail because it cannot beat existing treatment enough in a randomized controlled trial. Where the comments landed was that the stubborn 90% figure is best understood as a mix of three things. First, our preclinical models still do a poor job predicting human response, so too many candidates that were never going to work survive into the clinic. Second, the bar keeps rising. Regulators and companies now screen for risks and off-target effects that older drugs were never filtered for. Third, the easy wins are mostly gone, so companies spend more effort trying to eke out smaller improvements in crowded markets or tackle diseases where the biology is still murky. People were notably skeptical that AI changes the equation soon. Machine learning already helps in parts of the pipeline, but few saw evidence that it can replace the basic problem that we still cannot model the human body well enough to skip costly empirical testing.