But while AlphaFold has revolutionized the field, it differs in some key ways from other AI models, like ChatGPT. As Tomasz Włodarski, PhD, from the Institute of Biochemistry and Biophysics of the Polish Academy of Sciences explains, AlphaFold doesn’t actually “understand” the process it models. It’s an impressive tool, but it’s not without limitations.

A Historic Milestone in Protein Science

Proteins are fundamental to life. For more than a century, scientists have known that these complex molecules are essential to countless biological functions. However, it wasn’t until the 1960s that we had the tools to begin understanding their intricate 3D shapes. John Kendrew and Max Perutz, pioneers in the field, were awarded the Nobel Prize in Chemistry in 1962 for using X-ray crystallography to reveal the first protein structures in 3D.

"This discovery set the stage for a major leap in biochemistry,” Włodarski says. "By understanding protein structure, we could start to understand their function, which was a game-changer."

But the process of discovering new protein structures was painstaking and slow. Researchers had to conduct expensive experiments, and it could take years to solve the structure of even one protein. Despite decades of research, by the 1960s, scientists had only uncovered 225,000 protein structures—a tiny fraction compared to the 250 million amino acid sequences that had been cataloged.

The Dream of Faster Protein Structure Discovery

To speed up the process, the Critical Assessment of Protein Structure Prediction (CASP) competition was launched in 1994. Every two years, scientists tested their ability to predict the 3D structure of proteins based solely on their amino acid sequence. Early on, predictions were wildly inaccurate—only about 20% correct. By 2016, the accuracy had improved to 40%. But progress was slow, and for a while, many scientists doubted whether structural predictions would ever be reliable.

The game-changing moment came in 2018. DeepMind, a promising startup at the time, unveiled its AI-powered algorithm, AlphaFold. At the 2018 CASP-13 conference, AlphaFold amazed the scientific community by predicting protein structures with over 60% accuracy. And unlike other tools, AlphaFold’s creators made the program available for free online. The world was electrified.

AlphaFold2: A Leap into the Future

In 2020, DeepMind launched AlphaFold2, and the results were nothing short of groundbreaking. AlphaFold2 achieved 90% accuracy in predicting protein structures—matching experimental results. This was a huge breakthrough in molecular biology and dramatically changed the way scientists approach protein research. “It was a shock,” says Włodarski. “No one expected it to work this well.”

The implications were enormous. AlphaFold2 quickly became an indispensable tool for biologists, sparking a new wave of research. “We started to dig into old projects from decades ago and suddenly found answers to questions we hadn’t been able to solve before.”

AlphaFold Protein Database: A Treasure Trove of Knowledge

Building on its success, the creators of AlphaFold took the next big step in 2022, launching the AlphaFold Protein Structure Database, which houses predicted structures for over 250 million proteins. This vast database is a goldmine for scientists around the world, enabling them to quickly look up the structure of any protein they’re studying.

“Now, every biologist can check this database to find the probable structure of the proteins they’re working on,” says Włodarski. "This has sped up research across many fields."

Does AlphaFold Replace the Need for Structural Biologists?

At first, the rapid progress of AlphaFold sparked some concern. Why pursue a PhD in structural biology if an AI could solve the problem in minutes? But Włodarski points out that AlphaFold doesn’t replace the need for biologists. “It’s a model, not a perfect answer,” he explains. “There are still cases where AlphaFold gets it wrong. And even when it’s right, the model needs to be verified through experiments.”

Moreover, some proteins don’t have a clearly defined structure. For example, certain proteins may adopt multiple shapes depending on their environment or function. AlphaFold simply predicts the most probable structure, but it doesn’t explain the process of how a protein takes shape.

The Limits of AlphaFold: It Doesn’t “Understand” Protein Folding

While AlphaFold’s predictions are impressive, it’s important to understand its limits. Włodarski explains that AlphaFold doesn’t offer any insights into the process of protein folding—the series of events that happen as a protein takes its final shape inside a cell.

“That’s where AlphaFold falls short,” he says. “It doesn’t teach us the biophysics behind folding, like why proteins sometimes fold incorrectly. That’s crucial because misfolded proteins are behind neurodegenerative diseases like Alzheimer’s.”

AlphaFold is essentially a very advanced pattern recognition tool that uses existing data to predict protein shapes. It doesn’t learn from its own predictions, nor does it simulate the underlying physical processes. As Włodarski explains, it lacks an understanding of the physics at play during protein folding, which could help scientists understand the root causes of diseases linked to protein misfolding.

Understanding Protein Folding: The Next Frontier

Although AlphaFold can predict a protein’s final shape, it doesn’t yet offer insight into how proteins fold in real-time inside the cell. Włodarski’s own research focuses on understanding this complex process, which takes place in the ribosome, a molecular machine inside the cell.

“AlphaFold can’t simulate this process,” he says. “We need real-world data from the cell environment to understand what happens in the ribosome. It’s an incredibly complicated process.”

Over the years, scientists have discovered that protein folding in the ribosome differs significantly from what happens in a test tube. As Włodarski puts it: "Some of the processes we see in the lab don’t happen in the cell. And that’s what we’re still trying to figure out."

The Road Ahead: AI and Biology’s Bright Future

While AlphaFold has opened up exciting new possibilities, there’s still much work to be done. For example, Włodarski’s research suggests that by modifying the ribosomes in bacteria, it might be possible to make proteins fold faster. This could lead to more efficient protein production, which is crucial for developing new medicines.

“In the future, we might use AI to speed up protein production by optimizing ribosomes,” he says. “Artificial intelligence is already transforming biology, but there’s still a lot of research that requires human insight.”

As technology moves forward, AI models like AlphaFold will continue to reshape our understanding of proteins. However, as Włodarski points out, we will still need researchers to understand the processes behind protein folding and how they relate to diseases.

(PAP)

PAP - Science in Poland, Urszula Kaczorowska

uka/ bar/ amac/ kap/

tr. RL