Why AlphaFold Is Not the Template for AI in Science
A new analysis argues that the data conditions that made AlphaFold possible are rare and will take decades to replicate in most scientific fields, shifting attention toward AI agents as the more practical path to accelerating research.
👁 117 reads
The data infrastructure that made AlphaFold a Nobel Prize-winning breakthrough is so difficult to reproduce that most scientific fields will not meet its conditions for decades, according to an analysis published by MIT Technology Review - a finding that cuts against billions of dollars of investment in biology and chemistry foundation models.
- The Protein Data Bank, which underpinned AlphaFold, took 53 years and an estimated $21 billion to assemble - conditions that most scientific fields cannot replicate
- Experimental variability in biology and chemistry - drifting cell lines, trace contaminants, humidity changes - makes it practically impossible to build the clean, scalable datasets needed to train foundation models in most areas
- AI agents, rather than data-hungry foundation models, are identified as the more viable near-term approach to speeding up scientific discovery
The argument matters because the science-AI investment wave of recent years was built largely on a single premise: that AlphaFold showed the way. In 2024, Demis Hassabis and John Jumper of Google DeepMind received part of the Nobel Prize in chemistry for the neural network, which predicts the three-dimensional structures of proteins by learning from thousands of experimentally measured shapes. DeepMind itself called AlphaFold “the template for how AI can accelerate all of science to digital speed.” A wave of startups building foundation models for biology, chemistry, and materials discovery raised billions of dollars in the years that followed.
A $21 Billion Foundation Most Fields Cannot Afford
What that framing glossed over is the singular nature of the resource AlphaFold was trained on. The Protein Data Bank - the dataset of roughly 170,000 experimentally validated protein structures that underpinned the model - took 53 years of international scientific cooperation to assemble. By a recent estimate cited in the analysis, the experimental work required cost approximately $21 billion. Efforts of that scale, the analysis notes, are “infamously difficult to fund, next to impossible to coordinate, and hugely time-consuming to execute.”
The protein structure problem also benefited from an unusually reliable experimental technique. Protein crystallography is so dependable that over 25 Nobel Prizes have relied on it, making it possible to accumulate data that is consistent, accurate, and precise enough to train a modern neural network. That combination - decades of coordinated funding, open data sharing, and a reproducible measurement method - is not the norm in science. It is the exception.
In most of experimental biology and chemistry, results vary in ways that defeat the requirements of large-scale model training. Cell lines drift between labs and over time. Chemicals carry trace contaminants. Lab humidity shifts. The cumulative effect is that even well-resourced fields struggle to produce datasets that are consistent enough, accurate enough, precise enough, and scalable enough to train a foundation model reliably. Building the measurement infrastructure and standardized protocols needed to fix those problems, the analysis concludes, is work that will not be ready anytime soon.
Where Foundation Models Can Still Win
The critique is not a blanket dismissal of AI in science. The analysis identifies a narrow set of fields where the AlphaFold conditions are already met or close to it: weather forecasting, much of genomics, and limited areas of chemistry. Those domains may see AlphaFold-style breakthroughs shortly, if they have not already. The distinction matters for investors and developers evaluating where to direct resources - genomics, with its large, standardized, and increasingly open datasets, looks structurally different from, say, organoid biology or heterogeneous chemical synthesis.
Commercial data ownership is a separate barrier the analysis flags. Even where the required experimental data exists in principle, much of it sits behind proprietary walls. Pharmaceutical and materials companies routinely retain experimental datasets that would, in aggregate, be sufficient to train competitive models - but those datasets are not going to be pooled into a public resource on any near-term timeline. The protein structure community’s decades-long commitment to open deposition in the Protein Data Bank was, in retrospect, an anomaly shaped by specific historical and cultural conditions in structural biology.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →The Case for AI Agents Instead
Against this backdrop, the analysis makes the case for AI agents as the more practical vehicle for accelerating science in the near term. Rather than learning patterns from massive curated datasets, agents can reason, plan, and interact with tools - running simulations, searching literature, designing experiments, and iterating on results. The agent approach sidesteps the data-scarcity problem because it does not require a Protein Data Bank equivalent to exist before it can be useful. It can operate in fields where data is sparse, heterogeneous, or locked in natural-language form across millions of papers.
For developers and small-business owners building tools on top of large language models, the implication is practical: the near-term commercial opportunity in science-adjacent AI is likely to lie in agentic workflows - literature review automation, hypothesis generation, experimental design assistance, and data extraction from unstructured sources - rather than in training new domain-specific foundation models from scratch. The latter requires infrastructure that, outside a handful of fields, simply does not yet exist.
The deeper point the analysis makes is about the sociology of science as much as the technology. AlphaFold was not just a machine learning achievement; it was the downstream product of half a century of community-building, funding, and open-science norms. Replicating the output without replicating those conditions is not a strategy - it is a wish.
More on this from QUANTSLANT
news·Peer Review Is Breaking Down as Paper Volumes Climb 5.6% a Year