AI for science needs reasoning, not just data

For decades, people have predicted the end of scientific discovery. In 1903, physicist Albert Michelson thought all physical science facts were known. In the 1980s, Stephen Hawking believed theoretical physics might soon be complete. Now, with the rise of AI, this feeling is back, especially after a Nobel Prize was awarded for an AI breakthrough.

In 2024, Demis Hassabis and John Jumper from Google DeepMind won part of the Nobel Prize in chemistry. Their neural network, AlphaFold, predicts protein structures. It learned from thousands of known protein shapes. This problem had puzzled scientists for 50 years, and AlphaFold seemed to solve it.

Hassabis called AlphaFold a "template for how AI can accelerate all of science." This led to many startups using similar AI models for biology, chemistry, and materials. They raised billions of dollars. AlphaFold showed that AI, combined with enough data, could make huge discoveries, even if the underlying reasons weren't fully understood.

The Limits of AlphaFold's Approach

While AI will change science, AlphaFold's method might not be the best model for all fields. Its success depended on very specific conditions that are hard to replicate.

AlphaFold needed the Protein Data Bank, a collection of about 170,000 verified protein structures. Building this database took 53 years of international effort and an estimated $21 billion in experimental work. Such large-scale data collection is incredibly difficult to fund, coordinate, and execute.

Even when resources and cooperation exist, another problem is the difficulty of generating consistent data. Protein crystallography, the main technique for protein structures, is very reliable. It has been key to over 25 Nobel Prizes. However, most experimental science is less consistent. Cell lines change, chemicals have impurities, and lab conditions vary. Creating reliable, accurate, and scalable datasets for AI training in most of biology or chemistry would require new measurement methods and standards, which are not yet ready.

Some fields, like weather forecasting, genomics, and specific areas of chemistry, already meet these data requirements. They might see AlphaFold-style breakthroughs soon. Government support for creating and coordinating these datasets will be crucial. But for most scientific questions, a different approach is needed, at least for now.

The Rise of AI Agents

Scientists have always worked with incomplete data. Biologists looking for new drugs combine different calculations, known structures, and experiments. They use their judgment to weigh each method's strengths and weaknesses. Science is about combining many tools and refining results as new evidence appears. This is how most research happens, but until recently, no software could do this.

Now, AI agents can. An agent is an AI that can reason and use various digital or physical tools. Recent changes in AI architecture, powered by large language models, have made these programs common. They need far less specialized scientific data. This is a major shift for science. It allows for digital tools that can mimic the back-and-forth process of real research. Unlike AlphaFold, which solves a specific problem, agents are generalists. They don't offer a new way to do science; they digitally model how humans discover things.

For example, Google's AI Co-Scientist was tasked with finding out how antibiotic resistance spreads between bacteria. This is a major cause of drug-resistant infections. The system created smaller agents. One agent formed hypotheses from existing research. Another reviewed them critically. A third ranked the best ideas. A fourth refined the top hypothesis. The agent concluded that resistance genes travel on bacterial viruses, using whatever virus could carry them to a new host. This hypothesis was correct. Researchers at Imperial College London had spent a decade reaching the same conclusion through lab work. Their paper, which Co-Scientist had not seen, was still being reviewed.

Overcoming Challenges and Future Impact

AI agents like Co-Scientist are still new. They face challenges like sometimes making up information, having inconsistent judgment, and limits on their memory and input. However, these technical issues are expected to improve. As they do, agents will make science more reliable, consistent, and faster.

Agents could help solve the "reproducibility crisis" in science, where researchers often can't repeat each other's results. For years, scientists have been asked to share their raw data and code to standardize experiments. But this is tedious work that researchers often avoid. Agents, however, automatically record every step they take. This creates an exact record of their methods, making replication precise.

Agents will also improve scientific memory. Knowledge transfer between researchers can be difficult. Graduate students often sift through old, messy lab notebooks to find crucial details. With agents, a lab's entire scientific history could be stored in a central, standardized digital record.

The biggest impact of agents will be speed. If testing an idea takes less time than discussing it, people will just run the test. An agent that can read a thousand papers, design 500 molecules, and learn from failures overnight will drastically cut experimentation costs. This will speed up science and give researchers the freedom to explore bold, unusual questions they wouldn't have risked time on before.

While AlphaFold will lead to amazing discoveries, it won't be the sole driver of scientific progress. The shift to AI agents is a rare kind of breakthrough, a tool that can impact every scientific field at once. Historically, tools like calculus, statistics, spectroscopy, and computers have had this broad effect. Each opened up new problems and redefined their fields. AI agents represent another such transformation.

Deep Dive & References

AI Co-Scientist - Google DeepMind, 2024