New to LEVITY? Start here! Want to know more about who’s behind LEVITY? Check out this page. 🙏🏼 Not subscribed to the LEVITY podcast on Youtube yet? Do it here. 🎧 More of a listener? The podcast is also available on Spotify, Apple Podcasts and other places.

Time and again science is slowed down by dogma. Ideas and theories become so entrenched that if one even dares to challenge it they are labeled heretics. This has been the case throughout history and it persists to this day.

A very good and recent example of this is that much of what science thought it knew about proteins has turned out to be wrong, or at least only half the story.

The textbook version about proteins is sequence - structure - function. A certain sequence of amino acids gives a protein a three dimensional shape. That shape dictates what role that protein plays in an organism. Structural biologists say that these proteins fold.

This is true for the majority of proteins in biology. These are the type of proteins that the AI model AlphaFold managed to predict having been trained on data that scientists spent decades painstakingly assembling. Two of the minds behind AlphaFold - John Jumper and Google Deepmind CEO Demis Hassabis - was awarded the Nobel prize in chemistry in 2024.

But there’s also a large class of elusive proteins that even AlphaFold can’t say much about. The 3D shape of neatly folded proteins that expensive machinery can reveal is nowhere to be found. It’s like they’re ghosts.

And up until about two decades ago many scientists working in protein biology more or less thought they didn’t really exist at all, or at the very least underestimated their importance.

But they do exist. And they’re not ghosts, but shapeshifters.

Instead of folding into a single, stable form, they exist in a restless, ever-shifting state - a cloud of possible shapes that continuously flows from one to another. The pattern of that cloud depends on temperature, acidity, salt, and the molecules surrounding it. All of it happens at astonishing speed, from billionths of a second to milliseconds, allowing the cell to adapt in real time.

They’re called intrinsically disordered proteins, or IDPs.

Roughly half of all human proteins contain disordered regions, and many are disordered almost entirely.

Far from being evolutionary accidents, they’re central to how cells think, sense, and respond. IDPs act as molecular switchboards - turning signals on and off, linking pathways, and forming temporary assemblies that appear and dissolve as needed. They don’t run every process by themselves, but they’re woven through nearly all of them: gene regulation, stress responses, immune signaling. Their flexibility is what keeps biology responsive.

But that same flexibility has a darker side.

The fluidity that makes IDPs so useful also makes them fragile. As cells age and their repair systems falter, these shape-shifting molecules can start to go astray.

And many major age-related diseases involve them. Tau in Alzheimer’s, α-synuclein in Parkinson’s, and TDP-43 and FUS in ALS all belong to this family. Normally they move freely, forming and dissolving molecular assemblies as needed. With age or stress, they can get stuck - turning into sticky clumps that spread through tissues. Even central stress-response factors like FOXO3 and p53 rely on disordered regions to sense signals and control repair - linking their dynamics directly to aging outcomes.

Now, a team of researchers at Harvard and Northwestern has made what could turn out to be quite the breakthrough. Not only do they show that these mysterious proteins can be understood, but also designed. In a recent paper published in Nature Computational Science, they describe a new computational method that finally cracks the problem of predicting and engineering intrinsically disordered proteins.

The Harvard team built a system that combines physics-based molecular simulations with machine learning. Instead of guessing from data alone, the model learns directly from physical laws. It can predict how a protein’s amino-acid sequence gives rise to a cloud of structures and how to tweak that sequence to control the cloud’s behavior.

In other words, they’ve created a way to design disordered proteins on purpose.

That could have far-reaching consequences. If we can understand and control that behavior, we could maybe start designing therapeutic versions that resist aggregation, or synthetic molecules that stabilize the healthy “liquid” state of the cell.

It also opens doors beyond aging: adaptive biomaterials, smart sensors, new types of molecular drugs that can tune themselves to changing conditions.

Until now, AI in biology has mostly dealt with still images - like frozen shapes of folded proteins. This new generation of models is beginning to capture motion itself, inching closer to a future where we might one day steer life’s processes directly.