In October 2024, the Nobel Committee awarded Demis Hassabis and John Jumper the prize for AlphaFold, DeepMind’s program that predicts the three-dimensional structure of proteins. In 50 years of structural biology, the scientific community had solved approximately 200,000 structures. AlphaFold predicted more than 200 million between 2021 and 2022, in roughly one year. That’s a thousandfold difference, achieved not by recruiting more researchers, but by changing the nature of who does science.

What has happened since is even more profound. AlphaFold was not an agent: it was an extraordinarily powerful tool, but it answered questions posed by humans. The next generation of AI systems formulates the questions itself. It emits hypotheses, designs experiments, analyzes results, and proposes conclusions — all within hours. The researcher verifies, validates, and provides direction. They don’t disappear, but their role shifts.

Structural biology is the full-scale laboratory for this transformation. It will soon become the model for all science.

The Essentials

  • AlphaFold predicted more than 200 million protein structures between 2021 and 2022 in roughly one year, whereas 50 years of structural biology had produced approximately 200,000.
  • Autonomous AI agents — systems capable of formulating hypotheses, designing protocols, and analyzing results without step-by-step human instruction — are now operational in several university and private laboratories.
  • The acceleration raises an unprecedented tension: the volume of results produced by AI agents raises growing questions about whether existing verification infrastructures, from peer review to open databases, can absorb this flow.
  • The challenge of the next decade is not to regulate the speed of production but to build validation systems proportionate to the flow.

AlphaFold Was Only the Beginning

To understand what comes next, we must understand what AlphaFold changed. For decades, solving a protein’s structure required months of X-ray crystallography or cryo-electron microscopy. These techniques remain gold standards, but they are slow and costly. AlphaFold demonstrated that a deep learning model, trained on known structures, could predict unknown structures with a precision that, for the majority of proteins, rivals experimental methods.

The public database built from AlphaFold today covers virtually the entire human proteome and that of dozens of other organisms. It is freely accessible. Researchers in Nairobi or Buenos Aires without access to a synchrotron can now query a protein’s structure in seconds. This is a real shift in the conditions of access to cutting-edge structural biology.

But AlphaFold answered questions. It didn’t ask them.

Agents That Formulate Hypotheses

This distinction is critical. A scientific AI agent doesn’t merely execute a well-defined task. It surveys existing literature, identifies gaps or contradictions, proposes an experiment to resolve them, simulates or analyzes the resulting data, and generates a structured report. The complete cycle — from hypothesis to results — can take hours where a doctoral student would spend several weeks on the same sequence.

Several systems of this type already exist. Coscientist, developed by a Carnegie Mellon team and published in Nature in 2023, automated chemical synthesis end-to-end: literature search, protocol writing, laboratory robot operation, results analysis. The system synthesized known molecules without intermediate human instruction. AI Scientist, published by Sakana AI in 2024, goes even further: it generates research ideas, writes experimental code, conducts computational experiments, and produces draft complete scientific papers, with sections, graphics, and conclusions.

These systems don’t replace scientific judgment. They accelerate it and relocate it. The senior researcher becomes the one who sets directions, verifies protocol rigor, and decides what’s worth publishing. This is a change in position within the knowledge production chain, not an eviction.

This shift in position is comparable to what has occurred in other analytically intensive sectors. Empirical studies on American tech show that junior positions disappear in favor of strengthened senior roles: the machine absorbs repetitive tasks, humans retain higher-level judgment. Science follows the same logic, but with an additional consequence: repetitive tasks in science weren’t merely repetitive — they trained researchers.

The Database Outpaces Validation

The problem isn’t that AI agents produce poor results. It’s that they produce results faster than verification systems can absorb them.

Peer review, as it functions today, is an artisanal process. An article submitted to Nature or Cell waits on average several months before being evaluated by two or three peers who each devote a few hours to the task. This timeline was calibrated for a human submission flow. If AI agents can produce scientific paper drafts by the hundreds per week, the bottleneck is no longer in production but in validation.

Databases present the same problem. The Protein Data Bank (PDB) took 50 years to reach 200,000 experimentally validated entries. AlphaFold created its own database (AlphaFold DB) with more than 200 million predictions, distinct from the official experimental PDB — these predictions are not equivalent to experimentally verified structures. The scientific community knows how to distinguish between the two, but downstream users — physicians, engineers, pharmaceutical startups — don’t necessarily.

This is where the real risk lies. Not in an AI agent deliberately fabricating false results, but in an accumulation of plausible, unverified results circulating as if they were. Science already produces this phenomenon on a small scale: the replicability crisis, documented since the 2010s in psychology and biomedical sciences, shows that more than half the results published in certain fields don’t reproduce. AI agents didn’t create this problem, but they have the potential to amplify it considerably.

The Actors Building the Guardrails

This finding is not ignored. Several initiatives are working to build verification infrastructures suited to the new pace.

The EMBL-EBI (European Molecular Biology Laboratory – European Bioinformatics Institute) is developing automated annotation protocols to distinguish experimental structures from predicted structures in its databases. The objective is that each entry bears a confidence indicator readable by automated systems, not only by experts. This is an infrastructure response to an infrastructure problem.

Several scientific journals, including eLife, are experimenting with publication models where articles are made public upon submission, accompanied by open peer evaluations rather than hidden behind the review process. This openly reviewed preprint model is better suited to an accelerated flow: the community can react quickly, corrections are visible, authors’ reputations are publicly invested in each publication.

OpenAI, Anthropic, and several academic laboratories are investing in systems for automated assessment of scientific coherence. The idea is that an AI agent can also serve to verify the results of another AI agent, provided both are trained on different objectives and their architectures diverge sufficiently to produce independent errors. This is a bet on robustness through plurality, with its own limitations.

The question of open access takes on new urgency in this context. If scientific AI agents are developed principally in private laboratories — Microsoft and Amazon absorb the best AI labs without acquiring them, concentrating capacities outside public universities — discoveries produced risk not freely feeding the shared knowledge base. The AlphaFold Nobel crowned a tool whose database is public: this is a notable exception, not the rule in the sector.

The Long Arc: A Decade to Build or Miss the Turning Point

Projecting the impact of scientific AI agents over ten years requires distinguishing two trajectories, one of which doesn’t exclude the other.

The first trajectory, optimistic and plausible: AI agents allow exploration of hypothesis spaces that human science could never have covered for lack of time. In medicinal chemistry, the space of potentially active molecules is estimated at 10^60 compounds. No human research program can traverse it. Agents capable of screening this space, proposing the most promising syntheses and validating them computationally, then handing researchers the best candidates for experimental validation, dramatically accelerate drug development. Conservative estimates from the McKinsey Global Institute suggest that AI applied to life sciences could generate between 400 and 800 billion dollars in annual economic value by 2030 — accounting for clinical validation timelines, which remain human.

The second trajectory, more concerning and equally plausible: if open verification infrastructures aren’t built in parallel, science accelerates without validated knowledge progressing at the same pace. The volume of publications doubles or triples, but the verifiable share stagnates or declines. The distinction between produced result and established result erodes in public communication. Actors lacking access to verification tools — universities in the Global South, under-resourced academic laboratories — find themselves consumers of results they can’t test. Epistemic inequality deepens precisely where data access had widened.

These two trajectories don’t cancel each other out. They coexist already, and the choice isn’t between them but in the institutional energy devoted to each. Production acceleration is already here. Verification infrastructure remains to be built.

The generational question also deserves attention. Doctoral students trained today will be research directors in 2040. If their training unfolds in an environment where AI agents do the repetitive and exploratory work, they won’t learn to do certain things. This could be a net gain — freed from time-consuming tasks, they develop higher-level judgment earlier. It could also be a loss — certain scientific intuitions arise from long immersion in raw data, from patient tinkering that forces understanding of what remains incomprehensible. No data settles this today. Several laboratories, including David Baker’s research group at the University of Washington (2024 Nobel Prize in Chemistry for protein design), deliberately maintain training programs where students learn experimental methods before using computational tools.

Biology Is the Model, Not the Exception

What’s happening in structural biology isn’t an isolated case. It’s the first domain where the transition has been rapid and documented enough to be readable — protein structure prediction was a well-defined problem with measurable success criteria, making it a natural target for deep learning. But the logic applies to all experimental science.

In materials physics, AI agents already identify candidates for next-generation batteries and superconductors by traversing composition spaces that experimental chemistry would take decades to explore. In astronomy, data from the James Webb telescope represent a volume that human astronomers couldn’t process manually in reasonable time: automated systems detect anomalies, flag unusual galactic candidates, and generate hypotheses for observation teams. In epidemiology, predictive models identify pathogens at risk of species jump before epidemics emerge.

The difference from previous decades isn’t that AI enters science — that’s been the case since the 1990s. The difference is that current systems operate on the entire logic of the scientific cycle, not on an isolated step. Formulating a hypothesis, designing the experiment, analyzing results, drafting conclusions: each step was until now the domain of a researcher. These are now automatable in a growing spectrum of contexts.

The challenge is therefore not to slow this transformation or lament it. It’s to build, in parallel and with the same level of urgency, the open infrastructures that allow verification of what machines produce. Open-access journals with transparent evaluation. Databases annotated and maintained by public institutions. Automatic replication protocols. Training programs that don’t reduce researchers to passive supervisors but give them tools to question and correct the agents they supervise.

Science has always advanced through discontinuous acceleration. The printing press multiplied circulation of ideas and created new verification problems that the Republic of Letters took a century to solve. High-throughput genome sequencing flooded databases with variants whose clinical interpretation remains incomplete thirty years later. Scientific AI agents pose the same structural question, at greater speed. The answer lies not in the instrument, but in the institutions surrounding it.

Sources

  1. Sciences et Démocratie / Nature — L’IA accélère la découverte scientifique : promesse et vertiges: https://www.sciences-et-democratie.net/intelligence-artificielle-33106-ia-accelere-decouverte-scientifique-promesse-vertiges
  2. AlphaFold Protein Structure Database (EMBL-EBI / DeepMind): https://alphafold.ebi.ac.uk
  3. Boiko et al., « Autonomous chemical research with large language models » (Nature, 2023) — Coscientist
  4. Lu et al., « The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery » (Sakana AI, 2024)
  5. RCSB Protein Data Bank — historical statistics of structure deposits
  6. McKinsey Global Institute — The economic potential of generative AI (2023)
  7. Nobel Committee for Chemistry — announcement of 2024 prize, Royal Swedish Academy of Sciences
  8. Nobel Prize 2024 Chemistry – Official announcement: https://www.nobelprize.org/prizes/chemistry/2024/press-release/
  9. EMBL – AlphaFold Nobel 2024: https://www.embl.org/news/science-technology/alphafold-wins-nobel-prize-chemistry-2024/
  10. RCSB PDB – Milestone 200,000 entries: https://www.rcsb.org/news/639b9e337f8444f313d20414
  11. AlphaFold DB – Official site: https://alphafold.ebi.ac.uk/
  12. PubMed – AlphaFold DB in 2024 (214 million structures): https://pubmed.ncbi.nlm.nih.gov/37933859/
  13. ScienceDirect – Thousandfold factor AlphaFold vs PDB: https://www.sciencedirect.com/science/article/pii/S0966842X25001106
  14. Nature – Coscientist (original paper): https://www.nature.com/articles/s41586-023-06792-0
  15. Carnegie Mellon – Official announcement Coscientist: https://www.cmu.edu/news/stories/archives/2023/december/cmu-designed-artificially-intelligent-coscientist-automates-scientific-discovery
  16. Nature – AI Scientist (Sakana AI, 2026): https://www.nature.com/articles/d41586-026-00899-w
  17. Nature – AlphaFold (original paper 2021): https://www.nature.com/articles/s41586-021-03819-2