Recurrent neural networks in structural biology and protein science

How it is used

Proteins are chains of amino acids that fold into shapes, and the shape and the chemistry of particular positions decide what the protein does. A recurrent network reads such a chain in order, keeping an internal state that it updates at each residue, so what it makes of one position depends on the stretch it has already seen. The same habit of reading sequence in context underlies the protein language models that now supply much of this work: they are trained on large collections of sequences and emit, for every residue, a vector of numbers summarising its setting. In the examples here those vectors feed small downstream networks, or a paired network that compares two proteins.

The tasks are mostly labelling positions and regions. One study searched antibody sequences for the asparagine and glutamine sites liable to a chemical change called deamidation, using per-residue vectors from a pretrained model as the input to a two-layer network. Another marked the lysines that carry a chemical tag, taking a window of residues around each one. A third located a flexible loop inside a family of protease inhibitors, comparing three ways of encoding the same sequence. A fourth scored how alike two proteins of unknown function might be, pairing a recurrent layer with a distance between embeddings.

Why researchers use it

Much of what a protein scientist wants to know is written in the sequence but not locally: whether a residue is modified, or buried, or part of a binding surface depends on neighbours near and far along the chain. Experiments that settle such questions, mass spectrometry or a solved structure, are slow and cover a small share of known sequences. Reading a sequence in order, with memory, is a way of letting distant context reach a position without hand-writing rules about what should count.

It also stands in for feature engineering. The alternative, visible in one example here, is to choose the quantities yourself: charge, hydrophobicity, helix and sheet propensities, and so on, computed over candidate segments. That works where the chemistry is understood and the rules are simple, but it fixes in advance what the model may notice. A learned embedding from a pretrained model instead arrives already carrying patterns drawn from many sequences, so a small network trained on a few thousand labelled sites can reuse them.

In the record

Limits

Reading a chain in order says nothing directly about the folded shape, so two residues close in space but far apart in sequence are hard to relate, and these methods can miss effects that depend on the fold. Labels are scarce and uneven: sites not yet annotated are usually treated as negatives, which is often wrong, and under-sampling to balance classes discards real data. Models trained on well-studied families may not transfer to unusual ones. A prediction marks a position as worth testing; it does not establish that anything happens there.

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

Related

• Recurrent neural network: the idea, across every field
• AI in structural biology and protein science

Every article