Star Scientists Rarely Outshine Their Teams

On July 29, 2026, the Financial Times reported that Google DeepMind had dismantled the research team behind AlphaFold, the artificial intelligence system that transformed protein-structure prediction and helped earn the 2024 Nobel Prize in Chemistry. Most of the original paper’s authors had been reassigned to other work, while nearly a quarter of the full-time Google DeepMind authors had left the company.

The most visible departure was John Jumper, who led the AlphaFold team and shared the Nobel Prize with DeepMind chief executive Demis Hassabis. In June, Jumper announced that he was joining Anthropic. The move poses a question before any organization recruiting a celebrated scientist: how much of the breakthrough travels with the star, how much remains with the team, and how much depends on the particular combination of the two?

Research by Denisa Mindruta, Janet Bercovitz, Vlad Mares and Maryann Feldman offers an unusually precise answer. In collaborations involving star scientists, the star contributed more than the surrounding team in 14.3% of cases. The team, named constellation, contributed relatively more than the star in 9.5%. In about three-quarters of the collaborations, neither side dominated. The researchers’ paper, Stars in Their Constellations: Great Person or Great Team?, was published online in 2024 and appeared in the March 2025 issue of Management Science. 

How much of a discovery belongs to one person?

Scientific achievement is often narrated through individual names. Awards, recruitment campaigns and institutional rankings make the most visible researcher the apparent source of the result. But visibility is not the same as contribution. The paper’s question concerns relative contribution: when a collaboration creates valuable knowledge, which side would be harder to replace?

The researchers examined 555 invention disclosures registered between 1988 and 1999 at a leading U.S. research university with a renowned medical school. An invention disclosure is a formal record submitted to a university at the first stage of possible commercialization. It identifies the people involved in producing a discovery before decisions about patenting or licensing are made.

The dataset contained 1,003 scientists and 248 team leaders. Thirty researchers were classified as stars because their publication and citation records placed them among the top 5% of globally cited scientists in their fields. This setting allowed the authors to study collaborations that produced a concrete research output while observing the attributes of the leader, the constellation and the alternative partners available to each.

How did the researchers separate star from team?

The main measurement problem is selection. Strong scientists tend to work with strong collaborators, and successful teams do not form randomly. A simple comparison between teams with and without stars would therefore mix the effect of the star with the quality of the people who chose to work with that person.

The paper uses a matching model to estimate how leaders and constellations selected one another. The model considers attributes including previous research impact, knowledge profiles, the mix of senior and junior researchers, and the application areas in which the scientists worked. The researchers then use value-capture theory to compare each observed collaboration with the alternative matches available in the same market.

For each collaboration, they replace the leader or the constellation with the second-best feasible match predicted by the model. The fall in expected research impact indicates how difficult the removed party would have been to replace. A large decline after replacing the star signals a relatively irreplaceable star. A large decline after replacing the constellation signals a relatively irreplaceable team.

This approach estimates contribution from the structure of the matching market and the value created by the observed partnership. It does not divide the work into hours, laboratory tasks or lines of authorship. Instead, it asks a more revealing question: how much value would be lost if one side had formed its next-best available collaboration instead?

When does one side dominate?

Dominance in the study has a specific meaning. The star dominates when replacing the star would damage the result more than replacing the constellation. The constellation dominates when the reverse is true. The term describes relative value creation rather than authority, status or behavior inside the team.

As said earlier, The star dominated in 14.3% of collaborations involving a star scientist. The constellation dominated in 9.5%. More than 75% did not fall into either of the categories. In other words, in most collaborations, the star was not the story, and neither was the team. It is the match between the two sides rather than the attributes of either party alone that matters. 

High-value leaders tended to work with high-value collaborators. Some overlap in expertise made cooperation easier, while very high similarity reduced the advantage of combining their knowledge. Strong collaborations also brought together researchers at different levels of seniority and aligned them around related application domains.

The finding therefore goes beyond team size. A larger group does not automatically create more value, and a famous leader does not automatically carry the result. The strongest matches combine distinct bundles of knowledge, experience and research reach.

Why are the best matches hard to replace?

The paper separates two forces. Complementarity determines how much value the collaboration creates together. Substitutability determines the relative contribution of each side by showing how easily the other party could find a comparable partner. A leader can be highly accomplished and still have several close substitutes for a particular team. A constellation can contain excellent scientists and still be one of several groups able to work effectively with the same leader. In those cases, the collaboration may perform well while neither side accounts for most of its value.

The highest joint value appeared when both the star and the constellation offered rare combinations of attributes and worked especially well together. Neither side dominated because replacing either one would weaken the match. Paradoxically, the most valuable collaborations were often those in which it was hardest to say who deserved most of the credit. By contrast, dominance became more likely when one party had many possible substitutes. The dominant side then could claim a larger relative contribution, while the total value created by the collaboration was often lower.

This distinction is important for the way organizations read a record of success. A star’s previous output combines individual ability, access to resources, and repeated matches with particular collaborators. Treating that output as a purely individual track record risks attributing to the star value that actually belonged to the match.

What happens when a star moves?

The findings make the portability of star performance a central recruitment question. A hiring institution may acquire the individual’s knowledge and reputation while leaving behind collaborators whose rare skills helped generate the original results. Hiring the star does not necessarily mean hiring the capability that made the star successful.

The authors draw two practical implications. Organizations recruiting star scientists need to assess the constellation as well as the individual, including whether key collaborators can move, remain connected or be replaced by people with complementary attributes. Institutions allocating patent credit, financial returns and research resources also need to recognize that the less visible members of the team frequently account for an equal or greater share of the value created.

The evidence comes from a research-intensive medical setting and inventions disclosed between 1988 and 1999. The model estimates counterfactual matches rather than following a modern AI laboratory after its members disperse. The percentages therefore describe the collaborations in the study rather than every scientific or corporate team.

DeepMind’s July 2026 restructuring presents the same allocation problem in real time. The original AlphaFold researchers are now spread across Gemini projects, enzyme design, genomics, nuclear fusion, Isomorphic Labs, Anthropic and other destinations. Jumper’s future work will draw attention because the Nobel Prize makes his contribution highly visible. But the more revealing test may be what happens next: whether Jumper can recreate an equally powerful constellation at Anthropic—and whether the scientists who built AlphaFold alongside him can create new breakthroughs in new combinations.