Property prediction in structural biology and protein science

How it is used

Structural biology and protein science study the shapes that protein chains fold into and what those shapes let a protein do. Many of the things a researcher wants to know are measurable numbers attached to a protein that already exists: how much heat it survives before coming apart, how firmly it grips a partner molecule, whether a given change to one of its building blocks helps or harms it. Measuring each of these in the laboratory is slow and costly. Property prediction trains a model on proteins whose values have already been measured, then asks it for a number on proteins that have not been measured. The input is usually the protein's sequence of building blocks, its three-dimensional shape, or both.

The examples here show several model families doing this. Graph networks treat a protein structure as a web of connected points and learn from the local neighbourhood around a change. Protein language models learn statistical habits from very large collections of sequences, and some are conditioned on structure as well. Gaussian processes fit a smooth surface through a study's own measurements and also report how uncertain each estimate is, which lets a laboratory choose what to test next. Smaller learned components also read raw instrument data: in one case, networks scored peaks in a nuclear magnetic resonance spectrum as real signal or artefact and placed overlapping ones apart.

Why researchers use it

The reason researchers reach for these models is arithmetic. The number of possible changes to even a small protein is far larger than any laboratory can test, and the field's central questions, which variant is more stable, which binds better, which region has no fixed shape at all, are questions about choosing from that huge set. A model that returns a number in seconds stands in for an experiment that takes days, so it is used to narrow the field before anyone picks up a pipette. The examples show it used that way: to trim loose ends off a sequence before designing an assay, to screen the effects of mutations on binding, and to pick the next batch of variants to make.

In the record

structural-biologyResearchers measure folding stability for hundreds of thousands of protein variantsAlphaFold predicted structures for candidate sequences, which were used to trim loosely attached ends before assay design and to supply burial, contact and shape features for analysis.structural-biologyDeep learning tool predicts how single mutations weaken protein partnershipsA graph-based network trained on measured binding free energy changes predicted how mutations alter protein-protein interactions, with reverse mutations added to the training data.structural-biologyRobot lab and learning agent redesign an enzyme to survive higher temperaturesA Gaussian process fitted to the study's own measurements classified sequences as active or inactive and predicted thermostability for the active ones.structural-biologySoftware predictions map floppy stretches of touch-sensing PIEZO channelsSix sequence-based predictors scored every building block of two human PIEZO proteins for disorder, and a consensus profile marked regions where most predictors agreed.

Limits

These models learn from measurements that already exist, and those collections are uneven: well studied protein families and common kinds of change are over-represented, so a prediction on something unusual can be confidently wrong. Where no experimental structure exists, a predicted one is used instead, and its errors pass into the property estimate. Many models give a single number with no honest sense of its uncertainty. Agreement between several predictors, as in the disorder example, is reassuring but not proof: predictors trained on overlapping data can share the same mistake. Laboratory measurement remains the arbiter.

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