AI Companies Enter Drug Development Without Evidence It Works

Anthropic announced Claude Science this week as an AI workbench for drug development, built on a familiar frame—artificial intelligence excels at predicting molecular properties, therefore AI will accelerate pharmaceutical discovery.

The inference is so clean it feels inevitable. It is also wrong about what matters.

Drug development has a graveyard problem. Of every 5,000 compounds screened in the lab, only 250 move into animal testing—of those, 5 enter human trials. One gets approved. That 10% success rate has barely shifted in thirty years despite exponential improvements in computational chemistry, crystal structure prediction. Machine learning's ability to model molecular interactions.

The prediction part works fine. The drugs still fail where it matters—a compound can have perfect binding affinity to its target protein and still fail in a human because it doesn't cross the blood-brain barrier or metabolizes too quickly in the liver or triggers an immune response no amount of molecular modeling predicted. Then there is manufacturing, which introduces variables computational models don't touch—sterility, consistency, cost. A drug that works in theory can be impossible to produce reliably at scale, and then comes the regulatory apparatus. The FDA doesn't care how elegant your molecule is.

The companies that have genuinely moved the needle on drug approval rates are the ones that solved manufacturing, streamlined trials, or found regulatory pathways that worked—not the ones that predicted their way there. Genentech did it through biotech partnerships with Roche. Moderna did it by betting the entire company on one platform before anyone proved mRNA vaccines worked in humans. They built. Anthropic's move signals something real. Major AI companies see biotech as the next frontier and want to own the entry point.