A Talent Decision Was the Difference Behind One of Biology's Biggest Breakthroughs

How a 50-year-old grand challenge in science fell to the team that was deliberately designed to solve it, while better-funded labs kept missing.

On November 30, 2020, the organizers of the Critical Assessment of Protein Structure Prediction (a biennial competition known in the field simply as CASP) declared that a 50-year-old grand challenge in biology had been solved [1]. For five decades, biologists had been trying to predict the three-dimensional shape of a protein from its amino-acid sequence. The shape is what determines what the protein does: how it catalyzes a reaction, binds a drug, builds a tissue. Knowing it changes what is possible in medicine, in agriculture, in materials science. Almost every research biologist had spent some part of their career working on it, directly or indirectly.

The team that solved it was not the most senior in the field. It had not been working on the problem the longest. It did not have access to data or compute the rest of the field could not get. It had been working on this specific challenge for just over four years [2].

The team was Google DeepMind's AlphaFold group, led by Demis Hassabis and John Jumper. Their system, AlphaFold2, was assessed against 145 other entries in CASP14 and produced a median backbone accuracy of 0.96 ångström. The next-best system was at 2.8 ångström. Roughly three times more accurate, and comparable to the experimental methods that take a PhD student a year per structure to produce [3]. Four years later, in October 2024, Hassabis and Jumper shared the Nobel Prize in Chemistry for the work [4].

Every story about AlphaFold focuses on the algorithm. That is the wrong part of the story to learn from. The algorithm is the consequence. The decision that made AlphaFold possible was a hiring decision: who was on the team, what backgrounds they brought, how their work was organized. The same decision sits in front of every CEO right now, in domains far less exotic than protein folding.

Progress did not stall in biology because biologists weren't trying hard enough. It stalled because the people in the room were the wrong combination for the problem in front of them.

Why hadn't 50 years of effort been enough to solve the protein folding problem?

The protein folding problem had been worked on by some of the most capable scientists in the world for half a century, and the work had stalled. The reason was not laziness or lack of resources. The structure of academic labs constrained the structure of the teams that could attack the problem. The field had every other input it needed: a public dataset of around 200,000 experimentally determined protein structures [5], decades of theoretical work going back to Christian Anfinsen's Nobel Prize in 1972 [6], and the CASP competition itself, established in 1994 to give the community a shared benchmark [1].

What it did not have was a team built to put all of those inputs together. Academic biology is organized by department, and departments hire by discipline. What counts as legitimate work, and who gets promoted, is decided inside those disciplinary walls. A structural biologist trying to hire a software engineer for their lab is hiring against the grain of every system they operate inside: the grant body, the tenure committee, the promotion path. The result is predictable. Labs were full of brilliant scientists writing the code themselves, on the side, between experiments. The code was the bottleneck.

The protein folding problem had become a multidisciplinary problem long before the field had multidisciplinary teams. It needed people who could think about machine learning and physics and biochemistry at the same time, and people who could turn those thoughts into infrastructure other people could run. It needed both, on the same team, every day, for years.

A problem becomes structurally unsolvable when the people who could solve it are not allowed to be hired together.

What did DeepMind do differently when they built the AlphaFold team?

DeepMind built a team that academic biology was not structured to build. Roughly one third of DeepMind's total headcount was devoted to an internal organization called Research Engineering [7], led by a senior engineering executive with its own culture and career track. Half of those engineers built shared tooling: benchmarking infrastructure, data pipelines, leaderboards that any researcher in the company could use to test a new idea within roughly a day. The other half were embedded inside research pods of three or four scientists, treated as co-creators of the science rather than as service staff [7].

AlphaFold started inside what DeepMind called the Applied Group, which was majority Research Engineers. In the first nine months of the project, the team spent most of its time engineering and curating data: obsessing about accidental data leakage, choosing the right metrics, building infrastructure that could be used to experiment fast [7]. The deep-learning models came later. The team had been designed to make those models cheap to run and cheap to throw away.

Then, in 2018, after the first version of AlphaFold had won CASP13 but had not yet reached the accuracy needed to be useful to working biologists, DeepMind made the talent decision that turned the project from impressive into transformational. Hassabis expanded the team and installed John Jumper as the new research lead, with an explicit goal: redesign the system from scratch with a completely new architecture [8]. That redesign, supported by a multidisciplinary team Hassabis later described as combining machine-learning experts, great engineers, chemists, biochemists, structural biologists, and biophysicists [9], became AlphaFold2.

The people who built AlphaFold

Look at Jumper's path. Physics and mathematics undergraduate at Vanderbilt. Marshall Scholarship to Cambridge to start a PhD in theoretical condensed matter physics, which he left after a year with a Master's because the work was not a fit [10]. Three years at D. E. Shaw Research running molecular dynamics simulations on proteins and supercooled liquids. Then a PhD in theoretical chemistry at the University of Chicago, working on machine learning for protein folding and dynamics [11]. By the time DeepMind recruited him in late 2017, he was a physicist who had become a chemist who had become a machine-learning researcher. No academic department in the world is structured to hire that person directly for a tenure-track role. DeepMind's structure was.

DeepMind's competitive advantage in solving a biology problem was not biological. It was a team they could legally and culturally assemble that no academic lab could.

Why couldn't another well-funded lab have done the same thing?

Another lab could not have done the same thing because hiring sets the ceiling of innovation, and most labs and most companies do not realize they are choosing their ceiling when they are choosing their next hire. The data was public. The relevant algorithms, including transformers, attention, and neural networks, were all in the open literature. The compute budget for the first version of AlphaFold was within the reach of an academic project [7]. Every input that could have been copied was copied. The team was not.

A team is not a copy-paste artifact. It is the accumulated result of years of hiring decisions, role definitions, promotion criteria, and budget allocations. By the time a CEO or a head of department realizes their team cannot solve the problem in front of them, the team has already been built and is already producing the result it was structured to produce. Changing the result requires changing the team. And changing the team requires admitting that the previous decisions, made in good faith, were not enough.

This is the part most leaders do not want to be true. It is much more pleasant to believe that more effort, more deadlines, more pressure, or one more strategy off-site will close the gap. None of those things change who is in the room. They just exhaust the people who are.

Standard team versus a team designed to win

Hiring sets the ceiling of innovation. Effort only decides how close to the ceiling you get.

What does it look like to make a hire that raises the ceiling instead of reinforcing it?

A hire that raises the ceiling is a hire made against the existing team's center of gravity, with the explicit goal of making a problem solvable that is not solvable today. It is structurally uncomfortable. It usually involves overruling at least one objection from people who would prefer a candidate who looks like the team they already have.

In the AlphaFold case, the ceiling-raising hires were not single individuals. They were a design choice that ran throughout the team. Engineers were not a service function. The biologists were there as core members, not as a token addition to a machine-learning team. Each pod had the combination of skills needed to make a real decision without convening another meeting [7]. The decision-making latency was shorter than the latency of competing labs by months. Over a four-year project, that compounded into a result no one else could reach.

In a company, the equivalent is rarely the marquee senior hire that gets announced on LinkedIn. It is the third or fourth hire on a team that finally gets the composition right: the operator embedded with the strategist, the engineer embedded with the product lead, the legal mind embedded with the commercial team. The work the team can do shifts qualitatively. They stop running into the same recurring problem because the problem is no longer outside the team's combined skill surface.

A team is the accumulated result of every hiring decision you have made. By the time you can see the result clearly, the decisions that produced it are already years old.

What is the leadership moment most CEOs miss when progress stalls?

The leadership moment most CEOs miss is the realization that pushing harder with the same team is not a strategy. It is a confession that the team is the limit, presented as if it were a plan. Almost every executive can tell a story about a stalled product, a missed quarter, or a quality problem they tried to solve by adding pressure: more reviews, tighter deadlines, more weekend work, all before anyone considered changing the composition of the team. The pressure rarely produced the breakthrough. It often produced attrition, which produced a worse version of the same team a year later.

This pattern is expensive in every direction. The Society for Human Resource Management estimates that replacing an employee can cost between 50% and 200% of their annual salary, depending on seniority and specialization [12]. On a senior leader compensated at $300,000, that puts the cost of a single wrong hire somewhere between $150,000 and $600,000. That figure does not count the work that did not get done while the seat was occupied by the wrong person. It does not count the cost of not making the right hire in time: the projects that never shipped, the breakthroughs that went to a competitor, the talent that left because the team around them was not strong enough to keep them engaged.

The leadership move is to ask a different question than the one most leadership teams ask in a strategy review. The usual question is: how do we get more out of this team? The better question is: which problems are we structurally unable to solve with this team, and which of those will matter in twelve months? The first question accepts the ceiling. The second one is the only one that can change it.

Pushing harder with the same team is not a strategy. It is a confession that the team is the limit, presented as if it were a plan.

What is the broader principle this story illustrates?

The broader principle is this: organizations do not grow into success. They hire into it. The shape of the team determines the shape of what the team can build, and the shape of the team is set by decisions made long before the work begins. AlphaFold is the cleanest possible demonstration because the problem was famous, the data was shared, and the result was binary. Either the protein folded as predicted, or it didn't. There was nowhere to hide.

Most corporate problems are less clean. Signal is muddier. Timelines stretch. Metrics are negotiable. All of that makes the talent decision easier to defer and harder to attribute. The underlying mechanism is still the same. A team designed for the present cannot reliably deliver the future. A team designed to anticipate the future, staffed with combinations of skill that look unnecessary today, produces results that look impossible later.

This is what we mean by Talent-Led Growth: a growth model in which organizational outcomes are determined by deliberate talent decisions rather than tools, tactics, or effort alone. Organizations do not grow into success — they hire into it — by selecting for future states, designing teams as systems, and removing friction so people can compound impact over time.

Every great team has, at some point, made a hire that did not fit the team they had, because they were hiring for the team they were trying to become. The AlphaFold team made that decision many times, against the grain of an entire scientific field that was structured to make it impossible. The leaders who win the next decade are the ones who recognize that the most important business decision they will make this year is a hiring decision they have not yet authorized.

Organizations do not grow into success. They hire into it.

References

1. Jumper, J. et al. (2021). "Highly accurate protein structure prediction with AlphaFold." Nature 596, 583–589. The CASP14 organizers recognized AlphaFold2 as a solution to the protein structure prediction problem; the announcement was made on November 30, 2020 at the start of the CASP14 conference. nature.com
2. Google DeepMind, "AlphaFold: a solution to a 50-year-old grand challenge in biology" (November 30, 2020) — confirms project initiation in 2016 and the four-year arc to CASP14. deepmind.google
3. Jumper, J. et al. (2021), Nature — primary source for the CASP14 results: 146 entries total, median backbone accuracy 0.96 Å r.m.s.d.95 versus 2.8 Å for the next-best method. Jumper, J. (2025, Chicago Maroon interview) — each known protein structure takes roughly a year of a PhD student's time to produce. nature.com
4. NobelPrize.org — Press release, The Nobel Prize in Chemistry 2024, awarded jointly to David Baker, Demis Hassabis, and John Jumper. Announced October 9, 2024. nobelprize.org
5. Jumper, J. — Chicago Maroon interview (2025): approximately 200,000 known protein structures in the public training set, with around 14,000 new structures added per year. chicagomaroon.com
6. NobelPrize.org — The Nobel Prize in Chemistry 1972, awarded in part to Christian B. Anfinsen for the connection between amino acid sequence and biologically active conformation — the foundational result that frames the protein folding problem. nobelprize.org
7. Crossan, S. (2024). "Engineering for Science." Former DeepMind product lead on why AlphaFold happened at DeepMind: roughly one third of headcount as Research Engineering, embedded pods, the Applied Group composition, and the engineering-first first nine months. stevecrossan.medium.com
8. Gairdner Foundation — John Jumper laureate page: the 2018 expansion, Jumper becoming research lead, and the from-scratch redesign into AlphaFold2; confirms he was recruited in late 2017. gairdner.org
9. Hassabis, D. & Jumper, J. — "QnAs with Demis Hassabis and John M. Jumper," PNAS (2023): "We needed machine learning experts, of course, and great engineers, but also chemists, biochemists, structural biologists, and biophysicists." pnas.org
10. Vanderbilt University, "Boundary-Spanning Genius" (2024) — biographical profile of John Jumper: Marshall Scholarship, departure from the Cambridge PhD after a year, D. E. Shaw Research, and the move to a chemistry PhD at Chicago. news.vanderbilt.edu
11. John M. Jumper — biographical entry: BS Physics and Mathematics (Vanderbilt, 2007); MPhil Theoretical Condensed Matter Physics (Cambridge, 2010); PhD Theoretical Chemistry (Chicago, 2017). wikipedia.org
12. Society for Human Resource Management (SHRM), "The Myth of Replaceability" — replacing an employee costs 50% to 200% of annual salary depending on level, including recruiting, training, lost productivity, and onboarding ramp-up. shrm.org

Published by Teamtailor — building the talent operating system for organizations that intend to win on hiring.