Demis Hassabis and John Jumper of Google DeepMind won part of the 2024 Nobel Prize in chemistry for AlphaFold, a neural network that predicts protein structures by learning from experimental data. But researchers are increasingly questioning whether this approach can extend to other fields.
What Happened
AlphaFold solved a problem that resisted systematic attack for half a century by training on roughly 170,000 experimentally validated protein structures housed in the Protein Data Bank—a resource that took 53 years of international cooperation and an estimated $21 billion worth of experimental work to assemble. The article notes that most scientific fields lack comparable datasets because experimental results vary widely due to factors like cell line drift, chemical contaminants, and lab humidity changes. Creating consistent, accurate datasets suitable for training neural networks would require new measurement approaches 'none of which will be ready anytime soon,' the source states.
Why It Matters
A handful of fields already meet these data requirements—weather forecasting, much of genomics, and limited areas of chemistry—and may see AlphaFold-style breakthroughs soon. But for most open scientific questions, researchers are turning to a different approach: AI agents. These systems combine an AI reasoning engine with access to multiple tools, allowing them to synthesize diverse outputs and revise conclusions as evidence arrives—mirroring how human scientists actually work under uncertainty. A fundamental architectural shift powered by large language models has dramatically reduced the need for scientifically specialized datasets, potentially opening new avenues for AI-accelerated research across fields that lack the data conditions AlphaFold required.
The Bottom Line
The path to AI-accelerated science may diverge from the DeepMind template. While government support for dataset production and coordination remains critical for some domains, agents represent a more immediately accessible route to augmenting scientific reasoning in fields where massive standardized datasets simply do not exist.