Field guides

AlphaFold for experimentalists: a field guide

Structure prediction can now answer questions that used to take a crystallographer years. It can also hand you a beautiful, confident, wrong picture. This guide is about telling the difference.

Who it is for

You run experiments. You have a protein, or two proteins you suspect of binding each other, and you want to know what a predicted structure can honestly tell you before you spend six months at the bench. You do not need to program, and you do not need a background in structural biology.

Chapters

What is in it

1. What a prediction is, and what it is not. The confidence measures, pLDDT and PAE, explained in plain terms, and what each one is actually confident about.
2. Asking a good question. Choosing constructs, the price you pay for truncating a protein, and how many copies of each chain to include.
3. Reading a predicted complex. Interface confidence, the error between chains, and why a real interface looks the same across all five models.
4. The traps. Spurious interfaces and cavities at the cut ends of truncated proteins. Low-confidence regions that only look like structure. Chains that touch without being confidently placed. A distant resemblance in shape mistaken for a shared function.
5. Controls. How to build a small panel of known binders and known non-binders and run it through exactly the same settings as your question, so that a negative result means something.
6. From prediction to experiment. Which mutants to make, and what to test first.
7. Checklists and worked examples from publicly available data.

Why I wrote it

My lab studies how parasite proteins take over host cells, which means we spend a great deal of time asking whether protein A binds protein B. Over the past few years I have watched structure prediction change how we work, and I have also watched it mislead us, more than once, in ways that were obvious only in hindsight. Every trap in this guide is one that caught somebody I know.