When talking about intrinsically disordered proteins (IDPs), a class of protein which lacks a stable tertiary structure, two terms that crop up are SLiMs and MoRFs. In [the supplement to] a paper earlier this year in the journal Bioinformatics, Malhis and Gsponer give a nice explanation of the acronyms (above).
The paper itself describes a computational approach to classify MoRF sequences into [somewhat arbitrary] groups based on sequence similarity, with a high 'noise' tolerance fit for the messy reality of sequence variations in biology. This act of grouping is 'machine learning' - statistically determining 'clusters' with similar sequence.
Key to [IDPs'] regulatory function is the binding of molecular recognition features (MoRFs) to globular protein domains in a process known as a disorder-to-order transition. Predicting the location of MoRFs in protein sequences with high accuracy remains an important computational challenge.
If the groups were pre-defined, for instance if there were only three types of MoRF, it would be "classification" rather than "clustering" which is used when there's just a relationship (in which cases the groups are arbitrary and probably not hugely meaningful). In this case the 'quality' of groups may be given, a probability based on strength of the relationship between things classified into the same group. See supervised vs. unsupervised learning.
The 'classifier' is a support vector machine, a supervised learning model, and the high noise tolerance provided by a "Sigmoid kernel and… Radial Basis Function (RBF) Gaussian kernel". It's online at chibi.ubc.ca/morf/
