Can a computer guess the shape of a protein?

In Issue 2 we said a drug works by fitting into a pocket on a protein. That raises an obvious practical question: how do you find out what shape the pocket is?

For most of the history of biology, the answer was: with enormous difficulty.

The folding problem

A protein starts as a chain. Your cells read a gene and link amino acids together in order, like beads on a string. There are twenty kinds of bead.

Then the chain folds up. Not randomly — the same sequence folds into the same intricate shape every time, in a fraction of a second, and that shape is what determines everything the protein does.

So the sequence must contain the shape. We have been able to read sequences cheaply for decades. And yet going from the sequence to the folded shape turned out to be one of the hardest problems in science. The number of possible configurations for even a small protein is larger than is worth thinking about.

Instead, scientists measured shapes experimentally — growing protein crystals and firing X-rays at them. It works. It also took something like a year of skilled work per protein, and plenty of proteins refused to crystallise at all.

What changed

In 2021, a system called AlphaFold was shown to predict protein structures from sequence at an accuracy that was, for many proteins, comparable to experimental measurement.

The approach, very roughly: it had learned from every structure scientists had already solved, plus a powerful signal hidden in evolution. If two positions in a protein touch each other in the folded shape, then across thousands of species they tend to change together — because a mutation in one is tolerated only if the other compensates. Those correlations are a map of what touches what, written across millions of years of evolution.

Then the predicted structures were released publicly, for free, for over 200 million proteins — most of the proteins anyone had sequenced. Not quite all of them, as you are about to find out.

What this did not solve

Here is where it is worth being careful, because this gets overstated constantly.

A shape is not a drug. Knowing the pocket is the start of the problem, not the end. You still have to design a molecule that fits, that is not toxic, that survives your stomach, that reaches the right tissue, and that actually helps a patient. Those remain hard.

It predicts one shape. Real proteins move, flex, and change shape when they bind things. A single static structure is a useful simplification, not the whole truth.

Confidence varies. AlphaFold reports how sure it is for each region, and some parts of some proteins come back as little more than a guess. Floppy regions genuinely have no fixed shape to predict. Ignoring those confidence scores is a classic beginner's mistake.

It does not tell you what to target. It tells you what a protein looks like. Whether that protein matters in a disease is the question from Issues 2 and 4, and no structure will answer it.

What actually happened is that a step which used to take a year now takes minutes — and the bottleneck moved somewhere else. That is how progress usually works, and it is less dramatic and more interesting than "AI solved biology."

Try this

1. Go to the AlphaFold Protein Structure Database: https://alphafold.ebi.ac.uk
2. Search for filaggrin — the protein from Issue 1.

You will not find it.

That is not a mistake on your part, and it is the most useful thing in this issue. Human filaggrin is 4,061 amino acids long, and the database excludes proteins above a length cutoff. The protein with the strongest genetic evidence in eczema is missing from the database that supposedly covers everything.

Now do the comparison the absence sets up:

3. Search JAK1 instead — a protein that is drugged in eczema, and is in the database.
4. Rotate the model, then turn on confidence colouring. Blue is high confidence, orange and yellow are low.
5. Now look up lysozyme, a small compact enzyme of 148 amino acids, and compare.

Lysozyme is almost entirely blue. JAK1 has blue regions joined by orange stretches that genuinely have no fixed shape. And filaggrin is not there at all.

You have just seen, directly, why some proteins are far harder to work with than others — and why filaggrin being enormous and structural is part of why nobody has drugged it.

One question to take away

If predicting structures is nearly free now, what is the next bottleneck — and who is working on it?

---

Source: Jumper et al., "Highly accurate protein structure prediction with AlphaFold," Nature, 2021. PMID 34265844.