This paper studies how a major advance in artificial intelligence reshapes the production of scientific knowledge. I exploit the public release of AlphaFold2—a deep-learning system that predicts protein structures with near-experimental accuracy—as a sharp and field-wide reduction in the cost of structural information in structural biology. I construct a new protein-level dataset linking the universe of proteins to their experimental structural characterization, AlphaFold2 coverage, and the complete corpus of associated scientific publications. Following the release, proteins receiving larger informational shocks experience substantial increases in scientific activity: the probability of publication rises by 60–80 percent, and publication counts more than double relative to pre-release trends. These gains are not uniform. Publication responses are strongest for proteins that had partial, but incomplete, experimental structural information prior to AlphaFold2, indicating increasing returns to existing scientific capital. Rather than displacing experimentation, AlphaFold2 complements wet-lab research: experimental validation activity increases for proteins with high predicted coverage. The composition and organization of research teams also shift. Projects on high-coverage proteins involve more specialized and computationally oriented contributors, and resulting publications engage more intensively with structural and computational questions while maintaining experimental inquiry.
Contact Person: Michael Rose
Subscription to the invitation mailing list and more information on the seminar page.
