Clustering in structural biology and protein science

How it is used

Structural biology works with data that arrives in large, unlabelled heaps: millions of protein sequences, stacks of spectra from magnetic resonance experiments, and three-dimensional images of cells recorded by electron microscopes. Clustering takes such a heap and sorts it into groups whose members resemble one another, without anyone saying in advance what the groups should be. Related methods reduce the number of measurements describing each item, squeezing many numbers down to a few that keep the overall pattern. In protein work this is used to group sequences that are near-duplicates of each other, to collapse the many points of an image into a smaller set of representative summaries, and to tidy up feature descriptions before a further step.

Why researchers use it

Researchers reach for grouping when they have far more data than they can inspect, and no labels to guide them. Protein databases are heavily redundant: the same family appears again and again, so a model trained or tested on them can look accurate simply by memorising repeats. Grouping near-identical sequences and keeping one from each group is a standard guard against that. In image and spectrum work, grouping stands in for an expert's eye: it proposes which parts of a picture belong together before anything is identified by name. It also serves as compression, turning an unmanageable number of raw measurements into a handful of summaries that a later model can handle.

In the record

structural-biologyA transformer model segments cryo-electron tomograms from three viewing directions at onceA convolutional context encoder clustered its feature map into sixteen graph nodes, which then acted as attention queries for a transformer segmenting electron tomograms.structural-biologyLanguage model trained on protein sequences predicts shape and stability of disordered proteinsSequence redundancy among disordered protein sequences was checked with MMSeqs2 at a 0.8 similarity threshold before the sequences were passed to a language model.structural-biologyResearchers measure folding stability for hundreds of thousands of protein variantsPredicted structures supplied burial, contact-count and secondary-structure features, and low-contact terminal segments were trimmed before stability assays were designed.

Limits

Grouping has no notion of being right. It will return clusters from any data, including noise, and the groups it finds reflect whichever similarity measure and cut-off were chosen. Set a sequence similarity threshold loosely and near-duplicates survive; set it tightly and genuine diversity is thrown away. Collapsing many measurements into a few discards detail that may have mattered, and the compressed summaries are rarely interpretable in terms of chemistry or shape. Because no labels are involved, a grouping that looks clean can still cut across the biological distinctions a researcher cares about.

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