From the editor: Recently, Erik J. Larson wrote a five-part series for Mind Matters News on the use of artificial intelligence programs to unravel the ins and outs of protein folding in cells. Because disease impacts protein folding, improved knowledge is a vital strategy. The problem is that many important proteins are not regular but rather “intrinsically disordered,” which makes AI-based research more of a challenge. Larson has written, “Physical systems that evolve over time but don’t follow a fixed formula have always presented a deep challenge to AI. The problem of outliers or “edge cases” has frustrated AI scientists and engineers (and now structural biologists) for decades, and there’s no good answer yet.”

Wake Forest University biophysicist Jed Macosko. who researches this area, feels that there are grounds for optimism. The letter he sent to Mind Matters News follows:

There are many articles one could read about AI, and several about the area of AI that most affects my biophysics research, namely, AlphaFold. So, I generally pass on opportunities to read more articles about AI unless something catches my attention. Last week, Denyse O’Leary sent me an email with the subject line “AI in biology: The future AI didn’t predict – Erik J. Larson.” Since I know and admire Erik Larson and was curious about “the future that AI didn’t predict,” I read the article.

It turned out that this article was the fifth and final installment of a series of articles. I only read as far as the first clue of what “future” it was talking about. Since AI hasn’t been active in biology for very long, I was honestly wondering what it could possibly have predicted that was already revealed to be wrong.

Image Credit: Christoph Burgstedt - Adobe Stock

By the time I got to the thirteenth paragraph, I found the answer: The “future“ AI could not predict was the future of a given protein as it undergoes nanosecond-by-nanosecond changes to its structure.

I couldn’t agree more!

But, when I skipped down to the final sentence of the article, I read the following: “AlphaFold is a “revolution” in structural biology? Hardly.”

That seemed too strong. I started thinking about analogies to the type of revolution we are seeing in biology thanks to AI tools. The first analogy I thought of came straight from the article itself: Level 5 self-driving cars. Larson says, correctly, “Level 5 self-driving cars remain an elusive goal because the real world doesn’t conform to the practice or training data.”

The comparison with self-driving cars

As a refresher, Level 5 self-driving cars, if they existed, would drive indistinguishably from a good human driver. They would be able to go off-roading, drive proficiently in a foreign country, and have no trouble boarding a ferry, parallel parking, or pulling into an oil change garage. Level 4 cars, in contrast, already exist and do a great job of picking people up and dropping them off where they want to go, as long as where they start and end happens to be within a specific geographical area.

AlphaFold did to the community of protein folding researchers what Waymo did to Uber drivers in San Francisco, LA and Phoenix (and to some of them in Austin). But it also did more. I searched my mind for an analogy that better captures what AlphaFold did and why I think it really is a revolution in structural biology.

For some reason, I thought of the word “bematist.” I came across this word when I was giving a Columbus Day physics lecture. This term refers to the professional walkers in ancient societies who were paid to pace out long distances.

Unfortunately for Christopher Columbus, no one paid a bematist to walk overland from the coast of Portugal to the coast of China. If they had, he would have realized that the distance was 5,000 miles further than people had estimated. Then he would have known that his journey across water in the opposite direction would be much longer than anticipated. As it happens, there were unknown continents in the way. But for that fact, would have run out of food and water long before sighting land!

Waymo’s revolution was like the invention of a hodometer. That was an ancient version of the odometer, such as shown in the figure below. It allowed someone who had not been trained in the art of bematism to do what a bematist did. Heck, if the road you wanted paced out was all downhill, you could push your hodometer over the lip of the hill and introduce the world to the first self-walking bematist!

But AlphaFold’s revolution is more than this

If we stick with the bematist analogy, Alphafold is more like the jump between bematists and GPS. Just as the GPS can immediately tell you the distance between any two points on the globe, so also AlphaFold can tell you the structure of any protein with a known amino acid sequence.

However, this analogy is not perfect. It breaks down in two ways. First, there was a community of protein folding researchers who, from before I was in grad school and for the next three decades, worked incredibly hard to “crack” the code from amino acid sequence to protein folding. I doubt any bematists held conferences to discuss how they could “crack” the code of figuring out how far it is from point A to point B with the click of a button!

Image Credit: Christoph Burgstedt - Adobe Stock

Second, as Larson points out, 30–50% of human proteins contain at least a section of intrinsic disorder: an amino acid sequence that fails to consistently fold into a single, stable structure. This would be like GPS only working for two points that have land between them, but failing to measure distances over water.

But even if we take these two imperfections in our analogy into account, do you think that any bematist, when asked if GPS is a “revolution” in measurement, would answer “Hardly”? No way! Even though bematists might be sad that 30-50% of the possible point-A-to-point-B requests weren’t completely answerable by the new technology, they would still be overjoyed by the new development and consider it a true revolution.

AlphaFold and intrinsically disordered proteins (IDPs)

AlphaFold and its spin-offs represent a true revolution. In the end, it will be able to solve many questions about intrinsically disordered proteins.

For comparison, yes, “Level 5 self-driving cars remain an elusive goal.” Of course! But then Level 4 cars are all we really need. We don’t need to hail a Waymo to go off-roading!

Similarly, we don’t need to know the answers to all the “off-road” questions about intrinsically disordered proteins to better understand how they work.

Intrinsically disordered proteins are essential to life and their disorder is exactly what life needs. But how they allow for life is not going to be an off-road question. It will be a question that can mostly be answered without resorting to physics or mathematics that we do not yet understand. It will be a question that improved AI systems will be able to answer.

Here are the five articles in Erik J. Larson’s series:

AI in biology: AI meets intrinsically disordered proteins. Protein folding — the process by which a protein arrives at its functional shape — is one of the most complex unsolved problems in biology. The mystery of protein folding remains unsolved because, as is so often the case with AI narratives, the reality is much more complicated than the hype.

AI in biology: What difference did the rise of the machines make? AI works very well for proteins that lock into a single configuration, as many do. But intrinsically disordered ones don’t play by those rules. The resulting problems aren’t a temporary bug — they’re a basic limitation of training a machine learning model on a dataset where proteins always fold neatly.

AI in biology: So is this the end of the experiment? No. But a continuing challenge is that many of the most biologically important proteins don’t adopt a single stable structure. Their functions depend on structural fluidity. The core issue AI isn’t just missing data — AlphaFold’s overall approach won’t work here.

AI in biology: The disease connection — when proteins go wrong Some of the most crucial proteins for human health—the ones we need to understand most urgently—are the very ones that AI has the hardest time modeling. The issue is not simply that AI struggles with intrinsically disordered regions — it is that the very premise of IDR behavior contradicts the way these models operate. This isn’t just a flaw — it’s a fundamental crack in the foundation of AI’s “revolutionary” claims.

AI in biology: The future AI didn’t predict. It doesn’t look like the past. Physical systems that evolve over time but don’t follow a fixed formula have always presented a deep challenge to AI. The problem of outliers or “edge cases” has frustrated AI scientists and engineers (and now structural biologists) for decades, and there’s no good answer yet.