Experimental design in structural biology and protein science

How it is used

Structural biology and protein science study how chains of amino acids fold into shapes, and how those shapes do work: catalysing reactions, sitting in membranes, binding other molecules. Changing even a few positions in a sequence can alter all of it. The number of possible sequences is astronomically larger than the number anyone can build and measure, so researchers must choose. Here a model is trained on whatever measurements already exist, then asked which untested sequences to make next. In the examples this takes several forms: picking seed sequences from a combinatorial library of chimeras, ranking mutants for fitness, selecting variants that are both active and stable, and screening designed peptides before any wet work begins.

Two families of model recur. Gaussian processes, which predict a value and also how uncertain that prediction is, pair naturally with rules for choosing the next experiment, since a candidate can be worth testing either because it looks good or because little is known about it. One example fits such a model to classify sequences as active or inactive while also predicting thermal stability; another trains classifiers for whether a membrane protein expresses and reaches the cell surface. The second family is neural networks trained on sequence, including large models pre-trained on many known proteins and then adapted with a small number of laboratory measurements, or small networks trained on the output of physical simulations rather than experiments.

Why researchers use it

The binding constraint in protein engineering is usually the assay, not the computation. Some readouts are slow or low-throughput, so only a few dozen variants can be measured in a round. In one example, searching a combinatorial chimera space was attempted precisely because the gas chromatography readout could not cover it exhaustively. A model that proposes a small, informative batch converts an impossible exhaustive search into a few feasible rounds. The same logic explains the appeal of selecting a seed set that is maximally informative about a whole library: the aim is not to find the best variant immediately but to learn the landscape cheaply.

Uncertainty estimates stand in for the judgement an experienced engineer would otherwise supply about which guesses are worth the bench time. Where measurements are scarce, researchers borrow strength from elsewhere: from pre-trained sequence models, from alignment-based predictors used to generate stand-in labels, or from coarse-grained simulations that approximate what an experiment would show. Several examples also optimise more than one property at once, since a variant that is active but falls apart, or expresses but never reaches the membrane, is of no use.

In the record

Limits

These models learn from the variants already measured, so they extrapolate poorly to sequences unlike anything in the training set, which is where the interesting designs often lie. Predicted uncertainty is only as good as the assumptions behind it, and a confident wrong prediction wastes a round of bench work. Simulations and alignment-based scores used in place of measurements carry their own errors into the design loop. Assays measure proxies, such as expression level or a single catalytic number, which may not track the property actually wanted, and a design optimised for one property can quietly fail on another.

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

Related

• Experimental design: the idea, across every field
• AI in structural biology and protein science

Every article