The AI race to save American science and reshape medicine forever

Renee Wagerzin warns that the United States may lose the scientific competitiveness it has held since World War II, and questions whether the country is committed to preserving that advantage. She argues that a research system so dependent on federal funding is not sustainable. Stronger ties between universities and the private sector, along with more revenue from university intellectual property, could diversify support for science.

Wagerzin says ARPA-H was created to address gaps in conventional research by rapidly pursuing specific health problems, rather than focusing only on general knowledge. Its program managers’ statutory three-year terms were meant to reinforce urgency. She served through an administration transition and into the early weeks of the new administration, when she had an opportunity to make the case for ARPA-H’s continuation. DARPA has lasted nearly 80 years, she notes, and she hopes ARPA-H does too.

At Transfer, Wagerzin’s team is developing computer-vision and perception tools to capture what happens in laboratories beyond what protocols and notebooks record. Small procedural differences—gently inverting a tube rather than shaking it, for example, or labeling tubes before starting—can affect results. Headset video and audio, along with records of equipment settings, can help scientists locate mistakes and preserve techniques. That could spare researchers from repeating work after hours of preparation, such as discovering that a well was skipped only after spending six hours preparing a sample.

The aim is to support scientists, not monitor them: researchers control when recording starts and stops, and the headset design was changed in response to concerns that recording could feel like surveillance. Wagerzin says published papers are only starting points for breakthroughs, not products ready for patients. A lab may show that an approach works without making it scalable or establishing who will take it forward. Involving intended end users during development, and documenting procedures at both the originating and receiving labs, can expose details that published methods leave out and improve the transfer.

AI can produce billions of computational drug designs, Wagerzin says, but that has not translated into billions of clinical trials. Transferring assays between labs and scaling up manufacturing remain obstacles. She believes physical AI infrastructure could eventually enable one-shot manufacturing technology transfer and faster scale-up, helping cures reach patients sooner. The promise depends not just on generating designs, but on reliably reproducing experiments and moving successful work into production.

In the near term, Wagerzin sees lab automation as a way to augment researchers: machines can run repetitive tests, such as checking many magnesium concentrations, while people contribute judgment and hunches. She does not know when fully closed-loop autonomous science will be practical. Her experience with pandemic-era testing through Ginkgo showed that double-digit percentages of labs loaded 96-well plates in the wrong direction, and that human mistakes can escape notice when attention is focused on robots. Transfer tests models by deliberately changing protocols, omitting or adding steps, or using the wrong reagent; it retrains models on new data and uses human review to correct hallucinations and equipment misidentification. Wagerzin says feedback can also reveal weaknesses to AI developers, while excessive experimental labeling may waste computing resources. She describes the current period as consequential for the next 20 to 50 years, though she does not know what the next “Sputnik” will be.