A Disordered Protein Won't Hold Still for Its Portrait

Why the most beautiful structure of a protein is sometimes the least useful one — and what photography taught me about seeing dynamics instead of a single frozen shape.

I take a lot of photographs. Somewhere between playing my C. Bechstein B-212 grand piano, and golf, high-fidelity photography is one of the hobbies I keep returning to, and over the years, I have developed a deep respect for two very different machines.

On one side, there is the Hasselblad. A medium-format Hasselblad produces images of almost absurd fidelity — enormous sensors, exquisite glass, resolution, and tonal depth that an entire genre of professional photographers will never give up. When you want to capture a single, still subject in all its glory — a face, a landscape, a product on a table — nothing touches it. It is the gold standard for a perfect frame.

Hasselblad 907X: 100MP medium-format stills and 4K video, vs iPhone 17 Pro Max: 4K Dolby Vision at 24/25/30/60/100/120 fps, Cinematic mode up to 4K/30, Action mode up to 2.8K/60, and 1080p slow motion up to 240 fps — ultimate detail versus all-in-one speed and flexibility.

On the other side, there is the phone in my pocket. By any classical measure, an iPhone sensor is tiny, and its lens is humble. And yet it has quietly become the most consequential camera on Earth. Not because it out-resolves a Hasselblad — it does not — but because it does something a Hasselblad was never designed to do. It fuses dozens of frames in real time, it shoots video, it captures the laugh, the jump, the wave, the moment. It trades a little resolution for an honest record of what moves.

Yours truly, during the Panathaena festival of Arts, Science, and Tech in Greece (2026), with Hussein Kanji of Hoxton, giving a talk to a tech audience about disordered proteins and building a company to tackle them.

I keep coming back to this contrast, because it is the cleanest way I know to explain what my field has gotten wrong about proteins — and what a paper we just published is trying to fix.

The protein world built itself a Hasselblad

For most of the modern history of structural biology, our ambition was Hasselblad ambition. We wanted the perfect still.

We crystallized a protein, hit it with X-rays, and reconstructed a single, gorgeous, atom-by-atom structure. We learned a tidy story along with it: sequence folds into structure, and structure dictates function.

Workflow illustrating the structural biology pipeline, where X-ray diffraction patterns from a protein crystal are computationally processed into 3D electron density to reconstruct the final atomic model.

Find the one true shape, find the pocket, design the drug. The whole edifice of classical, structure-based drug discovery is, in a sense, a Hasselblad pointed at biology — an instrument optimized to capture one immaculate, motionless frame at the highest possible resolution.

An exemplary and oversimplified process of looking for a stable pocket in a well-folded protein that is the cornerstone of computer-aided drug discovery (CADD).

And for proteins that actually hold still, it is glorious. A well-folded enzyme sitting in its crystal is the perfect Hasselblad subject: posed, static, cooperative. Click. You get a masterpiece.

The archive — and the machines we built on top of it

And we did not just take these portraits one at a time. We built an archive.

The Protein Data Bank is the great gallery of structural biology — a single, public repository where, for more than fifty years, every beautiful still we captured has been hung on the wall for everyone to study. Today it holds over 255,000 structures, the overwhelming majority of them solved by exactly the kind of Hasselblad instruments I have been describing: around four out of five by X-ray crystallography, with most of the rest from cryo-EM and NMR (RCSB Protein Data Bank). It is one of the most extraordinary collective achievements in all of science — a quarter of a million frozen frames of life’s machinery, each one painstakingly earned.

And here is where the story takes its most thrilling turn. Because once you have a gallery that large, you can do something new with it: you can teach a machine to paint.

That is, in essence, what AlphaFold did. DeepMind’s AlphaFold2, and now AlphaFold3, learned to predict a protein’s three-dimensional structure straight from its sequence — and they learned it by studying the Protein Data Bank. AlphaFold3 is explicitly built to predict “nearly all molecular types present in the Protein Data Bank,” trained on the very structures deposited there over decades (Abramson et al., Nature, 2024).

The Protein Data Bank (PDB) — built over decades of X-ray crystallography experiments — served as the training foundation for a lineage of structure prediction models: DeepMind's AlphaFold2 and AlphaFold3, followed by the open-source models OpenFold, Chai-1, and Boltz, each of which releases model weights publicly. Together, these models can conjure a plausible 3D structure from a bare amino acid sequence in seconds.

It was a staggering achievement, and it kicked off an open-source renaissance: OpenFold faithfully retrained the AlphaFold2 architecture from scratch on roughly 132,000 experimental PDB structures (Ahdritz et al., Nature Methods, 2024), while Chai-1 and Boltz reproduced AlphaFold3-level accuracy and released their weights to everyone — each of them, again, trained on the full PDB (Boltz-1, MIT Jameel Clinic). In just a few years, we went from one frozen frame at a time to models that conjure a plausible frame for almost any sequence we hand them, in seconds.

But notice what every one of these remarkable machines inherited from its teacher. They learned from a gallery of stills. They are, all of them, the most sophisticated Hasselblads ever built — trained on perfect portraits to produce another perfect portrait. Ask AlphaFold for a structure, and it will hand you one gorgeous, confident, motionless frame, because that is the only kind of picture its entire education ever contained.

For a protein that holds still, this is pure magic. For one that doesn’t, for example, human Androgen Receptor (AR), the key driver of prostate cancer, the machine simply paints you the most convincing still it can of a subject that never sits still in the first place — and hands it over with a confidence that hides the very thing you needed to know. Just see the image below, the ‘spaghetti’ parts of the protein are all red and, from a structural biochemistry and drug discovery perspective, unusable.

AlphaFold3 structural model of the full-length human androgen receptor (AR) coloured by per-residue confidence (pLDDT). The globular ligand‑binding domain and portions of the DNA‑binding domain are predicted with high confidence (dark and light blue; pLDDT > 70), whereas the extended N‑terminal transactivation region and other flexible segments appear as highly disordered chains with very low confidence scores (orange to red; pLDDT < 50), consistent with these regions being intrinsically disordered.

And here comes the uncomfortable thing I have written about before, and keep returning to: a huge fraction of the proteins that matter most do not hold still. Roughly a third of the residues in our proteome never fold into a fixed shape at all, and these intrinsically disordered proteins and regions sit at the controlling hubs of biology — transcription (the example of androgen receptor), signaling, the condensates that organize the inside of a cell (Streit, Invernizzi, Bottaro, Tamiola & Lindorff-Larsen, Nature Communications, 2026). They are not posed. They are not cooperative. They are a subject in constant motion.

Asking a Hasselblad to photograph an intrinsically disordered protein is like asking it to photograph a hummingbird in flight. You will get one technically flawless frame of a blur — and you will learn almost nothing about how the bird actually flies.

The trap of the beautiful still

This is the part that took me years to fully internalize, so let me say it plainly.

For a disordered protein, a single high-resolution structure is not just insufficient — it can be actively misleading. It implies a permanence that does not exist. It is a photograph of a thing pretending to be something it is not. You walk away with a stunning image and a false belief: that the molecule looks like that.

It does not look like that. It looks like that, and that, and that — a shifting ensemble of interconverting shapes, most of them fleeting, visited and abandoned over and over. The function lives in the motion, not in any one frame, just like in the attached example of the ACTR protein.

ACTR (Activator for Thyroid Hormone and Retinoid Receptors) is a 71-residue human intrinsically disordered protein (IDP) that functions as a transcriptional coactivator. In isolation, ACTR populates a broad, heterogeneous conformational ensemble and lacks a stable tertiary structure; however, upon binding to the nuclear coactivator binding domain (NCBD) of CREB-binding protein (CBP), it folds cooperatively into three amphipathic α-helices (h1, h2, h3) that form a stable complex.

So you face a genuine trade-off, and it is exactly the photographer’s trade-off. You can have one frame at the highest possible resolution and lose all the motion. Or you can accept slightly lower per-frame resolution and, in exchange, capture the thing that actually matters: the dynamics, the rare moments, the fleeting shapes the molecule briefly assumes when it is doing its job. For a hummingbird — and for a disordered protein — the second option is not a compromise. It is the only honest picture.

A Hasselblad gives you a singular state at high resolution. An iPhone gives you dynamics at sensible fidelity. For proteins that refuse to hold still, the iPhone wins — not despite the trade-off, but because of it.

The iPhone trick is computational

Here is the detail I love most, because it maps so cleanly onto our work.

The iPhone does not beat physics. Its sensor really is small. What makes it extraordinary is computation — it captures many frames, understands how they relate in time, and fuses them into something richer than any single exposure could ever be. The intelligence is not in the glass. It is in how the frames are sampled and combined.

That is almost exactly what we set out to do with disordered proteins, except our “camera” is a physics simulation.

In our new paper in Nature Communications, my colleagues applied an enhanced-sampling method called On-the-fly Probability Enhanced Sampling — OPES — in a multithermal ensemble.

Enhanced sampling methods in molecular dynamics (MD) simulations are designed to overcome the fundamental timescale problem: standard unbiased MD trajectories become trapped in low-energy local minima, unable to surmount high free-energy barriers within tractable simulation times. By applying a bias potential that progressively flattens the free-energy landscape across conformational space, enhanced sampling methods enable a single simulation to escape local traps and reversibly visit rare, low-population states — including biologically functional conformations — that would otherwise remain inaccessible.

The technical achievement is that a single simulation replica diffuses across a wide temperature range and is reweighted back to physiological conditions, instead of running the brittle, parameter-heavy replica-exchange machinery these simulations normally demand (Streit et al., 2026). Think of it as one clever, computationally fused exposure rather than a cumbersome rig of cameras all firing at once.

Comparison of replica-exchange (REST2) and OPES multithermal enhanced sampling strategies for intrinsically disordered proteins. In REST2, multiple parallel simulation replicas — each held at a distinct temperature — must continuously exchange configurations, imposing significant computational overhead and parameter tuning. In OPES multithermal sampling, a single simulation replica diffuses autonomously across the full temperature range via a converged bias potential and is subsequently reweighted to physiological conditions (310 K), eliminating the need for replica exchange. Critically, OPES accesses rare, low-population folded conformational states not sampled by REST2 or unbiased MD, as demonstrated for the intrinsically disordered protein ACTR (Streit et al., Nature Communications, 2026).

And the OPES “footage” holds up against reality. Compared to the older REST2 approach and to plain simulations, it produces consistent ensemble averages, converges faster, and — crucially — explores broader conformational space, catching low-population states the other methods miss (Streit et al., 2026). Lower drama, more honest motion.

The shot of a disordered protein we caught

We pointed this iPhone-style camera at ACTR, a 71-residue disordered transcriptional coactivator that only folds when it meets its partner. In its free state, ACTR looks like exactly what the textbooks promise: a floppy, structureless chain. Photograph it with a Hasselblad and you would conclude there is nothing there to see.

But the motion told a different story. Our simulations revealed a rare, low-population set of transiently structured states — populated only a few percent of the time — in which multiple α-helices fold cooperatively and form genuine tertiary contacts, a fleeting, binding-competent shape the protein visits and then lets go (Streit et al., 2026). This is the laugh in the candid photo. The wave. The moment a posed portrait would never have caught.

And these are not numerical ghosts. The states are reversibly sampled, separated from the disordered ground state by only a modest free-energy barrier, and — the part that matters most to me — consistent with extensive NMR and SAXS data we used to validate the ensemble (Streit et al., 2026). The physics and the experiments agree on what the candid shot shows.

Why should anyone outside a biophysics lab care?

Because those fleeting, partially folded conformations may harbor exactly the kind of binding pockets that make an “undruggable” disordered protein suddenly addressable.

You cannot design a drug against a pocket you have never seen. And you will never see this one with a Hasselblad.

What does all this actually mean for computational protein research?

None of this means the Hasselblad is obsolete. For a folded enzyme with a deep, stable pocket, the perfect high-resolution still is still the right tool, and I would reach for it every time. The mistake was never owning a Hasselblad. The mistake was believing it was the only camera worth owning — and pointing it at subjects it was never built to capture.

The lesson photography taught me, and the lesson this paper makes concrete, is the same one:

Resolution is not the same as understanding.

A slightly lower-resolution picture that captures motion can tell you infinitely more than a flawless picture of a single frozen instant — when the truth of your subject is the motion.

Resolution is not understanding. The droplet is frozen in flawless detail — yet tells you nothing about the rain. The dancer is a blur — yet you can feel the leap.

For a third of the proteins in your body, the truth is the motion. Just like the human body, the proteins we care about are always in motion. So we stopped trying to take the perfect photograph and started shooting the honest, computationally fused, slightly lower-resolution footage instead. It is the least glamorous-sounding upgrade in the world. Trade a little resolution, catch the moment.

Acknowdlegments

• Big thanks to Kareem Reeda for reading, reviewing, and providing critical suggestions.

• I want to cordially thank our Scientific Advisory Board member, Prof. Kresten Lindorff-Larsen of the University of Copenhagen, for his continued scientific support and for spearheading our efforts to advance understanding of disordered proteins.

• I want to express my sincere gratitude to Dr. Julian Streit, who conducted this project jointly with our Peptone team: Dr. Michele Invernizi and Dr. Sandro Bottaro.