Candidate generation in structural biology and protein science

How it is used

Structural biology studies the three-dimensional shapes of proteins and other large molecules in living things, because a protein's shape largely decides what it does. Protein science asks how a chain of amino acids, the small building blocks strung together to make a protein, folds into that shape, and how to write new chains that fold the way we want. Candidate generation means letting a model propose objects for later testing rather than judging ones that already exist. Here the proposals are amino acid sequences, backbone shapes (the atoms forming a protein's main chain), short peptides, which are very small proteins, and whole sets of possible shapes a floppy protein might take.

The models work on geometry rather than text. In the examples, a protein backbone is turned into a graph: each atom or residue is a point, and the links between points carry distances between the main-chain atoms, how far apart the residues are along the chain, and whether two parts belong to the same chain. A network then reads that geometry and proposes amino acids to fill it. Other examples run proposals in a loop: a structure predictor guesses what a candidate sequence would fold into, a second model scores whether that shape has the wanted repeating pattern, and the combined score drives the next round of changes.

Why researchers use it

The space of possible proteins is far larger than any laboratory can test, and testing is slow and expensive. Researchers reach for candidate generation because it narrows that space cheaply: a model can propose thousands of sequences overnight, and only the promising ones go to simulation or to the bench. It stands in for the older approach of hand-designing proteins from physical energy calculations, which was slower and often produced sequences that did not fold.

The same models also stand in for steps that are not themselves design. One example uses a structure predictor simply to see which ends of a protein stick out loosely, so those parts can be trimmed before an experiment, and reuses the predicted structures as a source of measurements, such as how buried each residue is. Another trains a small network to reproduce the results of slow physical simulations, so that many candidate peptides can be scored without running each simulation. In both cases the model is a fast stand-in for something expensive.

In the record

Limits

A proposal is only a guess. The models learn from solved structures, which are biased towards proteins that crystallise well and fold stably, so sequences for unusual or floppy targets are less reliable. A design that scores well against a structure predictor may simply be exploiting that predictor's blind spots rather than being foldable; confidence scores measure the model's certainty, not the truth. Networks trained to imitate simulations inherit the errors of the simulations, and of the simplified physics behind them. Nothing here replaces laboratory testing, which remains the only real check.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Related

• Candidate generation: the idea, across every field
• AI in structural biology and protein science

Every article