The 50-Year Grand Challenge

Anfinsen's experiment: sequence encodes structure

Our story starts in 1957, in Christian Anfinsen's lab at the National Institutes of Health. Anfinsen was studying ribonuclease A, a small enzyme that cuts RNA. Like all proteins, ribonuclease A needs its specific 3-D shape to function.

Anfinsen unfolded the protein completely using a chemical called urea (which disrupts hydrogen bonds and hydrophobic interactions) and a reducing agent (which broke its disulfide bonds). The protein was now a floppy, unstructured string — totally inactive.

Then he simply removed the denaturing agents, allowing the protein to refold. Remarkably, the enzyme's activity came back — nearly completely. The protein had refolded into its correct shape spontaneously, without any external guide or template.

This was electrifying — and immediately raised an obvious question: if the sequence encodes the structure, can we compute the structure from the sequence? Surely, given enough physics and computing power, we could just simulate how the chain folds? That question launched a 50-year odyssey.

Levinthal's paradox: why brute-force simulation fails

In 1969, Cyrus Levinthal put the computational problem into sharp relief with a devastating calculation. Consider a modest protein with just 100 amino acids. Each amino acid in the chain has several rotatable bonds with a few stable positions each. If we assume just 3 possible conformations per amino acid, the total number of possible shapes for a 100-residue protein is:

Yet proteins fold to their correct shape in milliseconds to seconds. A protein doesn't randomly sample every possible conformation — it couldn't possibly. So how does it fold so fast?

This is Levinthal's paradox: the search space is astronomically large, yet proteins navigate it in biological time. The paradox's resolution is that proteins don't search randomly.

The folding funnel: not random, but directed

The resolution to Levinthal's paradox came from thinking about energy landscapes. Instead of imagining all possible protein shapes as a flat map to be exhaustively searched, picture the energy landscape as a multidimensional funnel.

The energy landscape for a protein is shaped like a funnel:

• The rim (top): Unfolded states — many possible, all high energy because hydrophobic residues are exposed to water.
• The walls (sides): Partially folded intermediates — progressively lower energy as hydrophobic cores begin to form.
• The bottom: The native folded state — the global free-energy minimum where hydrophobic residues are buried, hydrogen bonds are maximised, and the protein is stable.

The funnel shape means that as the protein folds, each step is on average energetically favourable. The chain doesn't search randomly; it's guided downhill by thermodynamics. Local structures form first (helices and sheets nucleate rapidly), then these compact into the final tertiary structure.

Understanding the funnel qualitatively solved the physical paradox — proteins fold fast because they follow energy gradients, not because they're lucky. But it did nothing to make computational prediction easy. The energy landscape has too many dimensions, and simulating it accurately requires knowing the precise contributions of every molecular interaction.

Experimental methods: powerful but slow

While theorists grappled with the computational problem, experimentalists developed powerful (but laborious) methods for actually measuring protein structures.

X-ray crystallography

The workhorse technique for most of the 20th century. The idea: a beam of X-rays is fired at a crystal of purified protein. X-rays scatter off the electrons of each atom, and the diffraction pattern is recorded on a detector. By applying a mathematical operation called the Fourier transform, scientists can reconstruct the 3-D electron density map, and from that, place atoms.

The catch: you have to grow a crystal of the protein first. Not all proteins crystallise — membrane proteins and intrinsically disordered proteins are notoriously difficult. Crystallisation is often described as more art than science, and can take months of optimisation. Even once crystals exist, data collection and structure solution can take weeks to months.

NMR spectroscopy

Nuclear magnetic resonance (NMR) spectroscopy measures how atomic nuclei (typically ¹H, ¹³C, ¹⁵N) respond to magnetic fields. In a protein, each nucleus resonates at a slightly different frequency depending on its chemical environment — which depends on its neighbours in 3-D space. By analysing patterns of nuclear Overhauser effects (NOEs — signals between nuclei that are close in space), structural constraints can be extracted and a model built.

The catch: NMR is limited to smaller proteins (typically below ~50 kDa, roughly 450 residues) because larger proteins give overlapping, unresolvable signals. It's also very instrument-intensive and requires isotopic labelling of the protein.

Cryo-electron microscopy (cryo-EM)

The revolutionary newcomer. Proteins in solution are rapidly frozen in a thin film of vitreous ice (so fast that water molecules don't have time to crystallise), preserving them in native-like conformations. An electron beam is passed through the sample, and images are collected from thousands of randomly oriented protein particles. Computational algorithms average these images and reconstruct the 3-D structure.

Cryo-EM exploded after the "resolution revolution" of 2013 — when detector technology and computational methods improved dramatically. By 2017, structures at atomic resolution (below 2 Å) became routine for large proteins and complexes. Jacques Dubochet, Joachim Frank, and Richard Henderson received the 2017 Nobel Prize in Chemistry for developing cryo-EM.

The catches: Still requires expensive instruments (>$5 million), specialist expertise, and significant data collection and processing time. Smaller proteins (<100 kDa) remain challenging.

The sequence-structure gap: a widening chasm

As genome sequencing became faster and cheaper (culminating in Next-Generation Sequencing in the 2000s), the number of known protein sequences exploded. By 2020:

A gap of over 1,000-fold. And growing faster every year, because sequencing is becoming exponentially cheaper while structural determination methods improve only linearly. It was physically impossible to close this gap experimentally. Computational prediction was the only viable path.

Early computational approaches

Researchers weren't waiting for AlphaFold — they had been developing computational prediction methods for decades. The main approaches were:

Homology modelling

If your target protein has a known relative (a homologue) with a solved structure, you can use that structure as a template and thread your sequence onto it. Works well when there's a closely related protein (>30% sequence identity), but fails when no suitable template exists — so-called "orphan" proteins or novel folds.

Fragment assembly

Divide the chain into overlapping short fragments (~9 residues), look up each fragment's likely local structure from a database of solved protein structures, then assemble and optimise. Rosetta (David Baker's lab, University of Washington) became the gold standard. Good for small proteins but computationally expensive and inaccurate for large ones.

Physics-based simulation (molecular dynamics)

Simulate the protein's atoms bouncing around according to Newtonian mechanics, hoping they'll stumble into the native fold. Fundamentally limited by computational cost — even the fastest supercomputers could only simulate microseconds of folding time, while folding takes milliseconds. Dedicated hardware (Anton, from D.E. Shaw Research) extended this, but not enough.

The CASP benchmark

To measure progress objectively, the Critical Assessment of Protein Structure Prediction (CASP) competition was founded in 1994. Every two years, experimental groups provide target sequences whose structures are known but not yet published. Computational groups submit predictions, which are scored after the experimental structures are released.

CASP provided a rigorous, fair leaderboard. For 25 years, progress was slow and incremental — methods improved, but the hardest targets (proteins with no template, or large multi-domain proteins) remained largely unsolved. Then came CASP14 in 2020, and everything changed.