This is the third part of Erik J. Larson’s series on the attempt to understand protein folding using AlphaFold. The first two parts are here and here.

Critics argue that AI cannot simply replace lab work — after all, the ground truth data that AlphaFold relies on comes from those very experiments.  X-ray crystallography and cryo-electron microscopy (cryo-EM) have built the foundation on which the success of AI now stands.

If researchers become too dependent on AI-generated predictions, they risk solving only the easy problems and ignoring the difficult cases lurking in  (or not in) the training data, limiting the system’s ability to generalize beyond what has already been observed. AI does not uncover the unknown so much as it makes highly informed guesses based on what has been seen before. Regardless of how many times we hear this mantra about AI, it remains important: AI’s limitations stem from a few core observations that seem impervious to solution because they’re baked into the approach itself.

Image Credit: Christoph Burgstedt - Adobe Stock

And yet, even the most skeptical critics acknowledge that AlphaFold has given structural biologists an unprecedented resource. The database of predicted structures it has generated is already proving valuable in drug discovery, where researchers study how proteins interact with small molecules in the search for new pharmaceuticals. It is also driving advances in enzyme engineering, enabling scientists to design proteins with novel functions for biotechnology and synthetic biology.

AlphaFold 2 debuted impressively at the biennial CASP (Critical Assessment of Structure Prediction) competition in 2021, where it stunned the structural biology community. Its ability to predict protein structures with accuracy approaching experimental methods was a breakthrough. Scientists quickly recognized that DeepMind’s system had solved many of the folding problems that had frustrated researchers for decades.

This is all good news.

The Problem of Intrinsically Disordered Regions (IDRs)

Here’s the bad news, already hinted at above. AlphaFold 2 — and now AlphaFold 3 — remain fundamentally limited to stable, well-structured proteins that adopt a relatively fixed form. AI-driven protein prediction thrives when proteins behave like elegantly folded origami, locking into a single predictable shape. But biology is messier than that.

Many of the most biologically important proteins don’t adopt a single stable structure at all. Their functions depend on structural fluidity, shifting between conformations depending on biochemical context. These proteins respond to changes in temperature, pH, molecular binding, and post-translational modifications — factors that AlphaFold doesn’t account for dynamically.

Since AlphaFold is designed to predict a single “best-fit” structure, it misinterprets these proteins entirely, either:

1. Producing a low-confidence prediction, effectively admitting defeat.
2. Forcing a disordered protein into a misleading static form, which it never actually adopts in nature.
3. Generating nonsensical outputs, where long stretches of disordered regions appear as gibberish in the predictive model.

The core issue isn’t just missing data — it’s that AlphaFold’s entire approach is built on assumptions that don’t apply to disordered proteins. The AI revolution in protein folding has largely ignored them, and as long as models like AlphaFold remain tethered to static structure prediction, they will continue to do so.

And this is where the revolutionary claims about AI in protein folding start to crumble. Since a huge number of proteins exhibit IDRs in their design — and AI cannot yet model them — talk of a structural biology revolution is premature. AlphaFold and its successors have proven their utility, as structural biologists will readily attest. But the fine print is sobering: most of the protein folding problem remains unsolved, and when it comes to intrinsic disorder, AI has largely drawn a blank.

A problem deeper than just numbers

The sheer prevalence of IDRs alone presents a major problem for AI-driven folding prediction. But the problem runs deeper than just numbers. We’re not just dealing with edge cases throwing off an otherwise sound model. IDRs are not rare exceptions—they are fundamental to biology itself.

Intrinsic disorder in proteins is itself an essential molecular tool, deeply embedded in some of life’s most crucial processes. And AI, so far, has no idea what to do with them.

Why Are IDRs So Important?

So we’ve arrived at the first big takeaway here, contra much hype: far from being an evolutionary afterthought, intrinsically disordered regions (IDRs) are central to protein folding itself. Unlike well-folded proteins, which snap into a single, stable shape dictated by their amino acid sequence, IDRs remain in constant motion, a flexibility that makes them indispensable to cellular function, but it also makes them notoriously difficult to predict, model, and ultimately, target for drug discovery. AlphaFold and other machine learning approaches will tend to ignore these cases — which is why the Google-inspired Dialogues piece failed to mention disordered regions once.

Indeed, many of the body’s most important signaling proteins — including those that control gene expression, relay signals between cells, and regulate the cell cycle — depend on IDRs for function. Unlike their well-folded counterparts, IDRs are not locked into a single role. Instead, they adapt, binding to multiple molecular partners and responding dynamically to changes in their biochemical environment. It’s as if you want to know what your post office mailman is up to when he is not delivering letters — and then you see him flying the plane on your next vacation. Huh?

And signaling isn’t the only important function either. IDRs drive protein‒protein interactions, acting as the connective tissue of complex biochemical networks. Their flexibility allows them to bind multiple partners in different conformations, making them the ultimate molecular adapters. But this same flexibility — the very thing that makes IDRs so functionally powerful — also makes them inherently unstable. Defining the structure of an IDR is like trying to capture the shape of a snake mid-motion — coiling around one object, then uncoiling and wrapping around another, never staying in one configuration for long.

Note: Erik J. Larson writes the Substack Colligo.

Here are the first two articles in this 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.

Here are the fourth and fifth:

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.