Closing the data loop in AI-driven drug discovery

Drug development has always been brutally slow and expensive, and it keeps getting worse. Since the 1950s, costs have roughly doubled every nine years, a trend researchers call Eroom's Law. A new drug now takes 10 to 15 years and $1 billion to $2.5 billion to reach market, with more than 90% of candidates failing along the way. Pharma companies are betting heavily on AI to change that math, and Paul Belcher, director of protein research strategy at Cytiva, has watched the shift happen from inside the lab.

The clearest early win is in hit identification, the process of finding molecules that bind to a disease target. Belcher describes a move away from brute-force physical screening toward predictive design, where AI proposes candidates and models their behavior before anyone touches a bench. That removes the old ceiling on how many compounds a company can physically test. But AI still can't reliably predict things like kinetics or developability, so every generated candidate still needs lab validation — and that's straining workflows built for cruder, yes-or-no screening rather than detailed profiling of AI-generated compounds at scale.

The bigger issue, according to Belcher, is data quality. Many models trained on public datasets are hitting what he calls a data wall: everyone draws from the same pool, so results converge and improvements shrink. Worse, those datasets skew almost entirely toward successful experiments, since researchers rarely publish failures. Belcher jokes there should be a journal of negative data, because right now it just sits buried in lab notebooks, never feeding back into research. Without that balance, models learn what works but not what doesn't, which limits how reliable their predictions can really be.

Fabrication adds another layer of risk. Belcher points to Elisabeth Bik's 2016 finding that nearly 4% of biomedical papers contained duplicated or manipulated images — and that was before generative AI made faking data far easier. Tools like Cytiva's Image Integrity Checker, which borrows hashing techniques from blockchain to verify scientific images, are starting to catch interest from publishers wanting cleaner data before it ever reaches a model.

Belcher's long-term vision is the autonomous lab: systems that run around the clock, testing, optimizing, and feeding results straight back into AI models with minimal human involvement. Getting there depends less on smarter algorithms and more on plumbing — interoperable instruments and data that can actually move between systems, rather than sitting locked inside standalone machines. No AI-designed drug has yet won full FDA approval, though Belcher expects that within two to three years. And with frontier AI training costs more than doubling annually since 2016 according to a Stanford study, he thinks the industry will eventually settle into a balance between computational prediction and physical lab work, rather than one replacing the other entirely.