Paper Short Abstract

We investigate how AlphaFold and other AI methods shape scientific discovery, focusing on research productivity, exploration of less-studied areas, and translational outcomes. Our large-scale analysis reveals shifts in structural biology and beyond, offering early insights into AI’s broader impacts.

Paper Abstract

Artificial intelligence has seen rapid diffusion across science. Proponents highlight the potential for more efficient and effective scientific discovery, while others raise concerns that AI leads to a narrowing of scientific research. AlphaFold, Google DeepMind’s AI-driven system for protein structure prediction, a long-standing problem in structural biology, has emerged as a high-profile example. In this paper, we examine AlphaFold’s impact in the broader context of AI’s potential to accelerate discovery, diversify research portfolios, and spur innovation. Leveraging a large-scale dataset of publications, patents, protein structures, and clinical trials, we employ difference-in-differences to compare AlphaFold’s effects to other frontier developments, both AI-based and non-AI-based, in structural biology.

Our findings indicate that AlphaFold is associated with notably higher experimental output, especially on novel protein structures. These gains are concentrated among experienced researchers and principal investigators, who also benefit from modest upticks in publication volume and citation metrics. While AlphaFold has not yet led to a notable increase in patenting activity compared to established non-AI methods, clinical citations linked to AlphaFold publications exceed those of many other AI-based approaches. This trend hints at its growing relevance for disease-related inquiries.

By mapping shifts in the organisms and protein structures targeted by researchers who use AlphaFold, we reveal how AI-enabled predictions can lower barriers to exploring less-studied domains. Beyond AlphaFold, our results speak to the broader capacity of AI methods to transform scientific practices, from productivity to translational applications. This underscores the need for continuous monitoring and policy support to harness AI’s promise in science.