Gaussian process in structural biology and protein science

How it is used

Structural biology and protein science ask how a chain of amino acids folds, holds together and does its job. A central practical question is which changes to that chain will help. Each measurement means building a protein in the laboratory and testing it, so only a handful of the possible sequences can ever be tried. A Gaussian process fits a smooth function over sequence space from the measurements already in hand, and gives both a prediction and a measure of how unsure it is. Researchers use it to predict properties such as whether a protein is made at all, whether it reaches the right place in the cell, how much heat it survives, or how much product it yields.

Two uses recur. One is to rank untested sequences so that the next round of laboratory work is spent on the most informative or most promising candidates. The other is to sort sequences into classes, such as working against not working, before predicting a numerical property only for those that work. The examples show both, often in the same model: active or inactive alongside stability, high against low expression alongside how much protein reaches the membrane. The choice of kernel, the rule saying which sequences count as similar, is where biological assumptions enter; counting the amino acids that differ between two sequences is a common one.

Why researchers use it

The appeal is doing well with small data. Laboratory assays in this field are slow, and a readout measured one sample at a time may allow only tens of experiments against a library of thousands. A method that works from a few hundred measurements, or fewer, fits that constraint. Because it reports uncertainty, it can be used to decide what to measure next, balancing sequences it expects to be good against sequences it knows little about. One example used it to pick a small set of seed sequences chosen to tell the researchers as much as possible about a far larger library.

It stands in for exhaustive testing, and for the chemical intuition of guessing which mutations to combine. Combining mutations is the hard part: effects interact, so a sequence can fail for reasons neither of its parent changes predicts. A model fitted to measured combinations captures some of that interaction without needing a physical account of why. It also stands in for simulating protein structure and energetics, which is costly and does not always predict what a cell will actually produce.

In the record

structural-biologyRobot lab and learning agent redesign an enzyme to survive higher temperaturesA multi-output model with a kernel counting sequence differences classified sequences as active or inactive and predicted heat stability for the active ones.structural-biologyGaussian process models pick light-sensitive ion channels that reach mammalian cell membranesClassification models trained on chimeric protein data separated high from low expression and membrane localisation; a regression model predicted the localisation level.structural-biologyMachine learning picked which engineered enzymes to build for fatty alcohol productionAn information-based criterion chose twenty seed sequences from a large chimera library, which were then built and measured for fatty alcohol production.

Limits

A Gaussian process only interpolates between the sequences it has seen, and a protein library is sparse in a vast space, so predictions far from the training data are guesses with wide error bars. The kernel does the work: if similarity is defined by counting differing amino acids, structural context and interactions between distant residues are largely invisible. Fitting cost grows quickly with the number of measurements. Classification requires approximating a calculation that cannot be done exactly. And the model learns the assay that was run, which may not match the property actually wanted.

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