Random forests in structural biology and protein science

How it is used

Structural biology and protein science ask what shape a protein takes, which parts of it touch other molecules, and how changes in its sequence alter what it does. A random forest is a set of decision trees, each tree a chain of yes or no questions about measured quantities, each grown on a random part of the data and a random subset of those quantities; the trees then vote. In this field the quantities are almost always numbers computed from an amino acid sequence or a structure: net charge, how water-repelling a stretch is, how often each residue type appears, how strongly a segment tends to coil into a helix.

The examples follow a common recipe. A set of known positive and negative proteins is gathered from public databases, checked against a reference sequence archive, and thinned so that near-identical sequences do not appear twice. Each sequence, or each short window within it, is then converted into a fixed list of numbers. Some studies use classical descriptors counting residues, pairs and triples, or physical indices; others use vectors read out of large pre-trained models of protein language, used with their weights frozen. Tree ensembles usually sit alongside other classifiers in a stack, where a second model learns how to combine the first ones.

Why researchers use it

The data in this field arrive as tables: one row per protein or per candidate site, one column per computed descriptor. Tree ensembles suit that shape. They need no scaling of the columns, cope with descriptors measured in unrelated units, tolerate columns that carry no signal, and work on the few thousand labelled examples a curated protein set typically yields, where larger models would simply memorise. They also report which descriptors the trees leaned on, which lets a biologist check that the answer rests on something sensible.

What they stand in for is experiment. Determining a structure, or assaying every candidate binding segment or modification site in a proteome, is slow and costly, so a classifier is used to rank which candidates are worth the bench work. They also replace hand-written rules: an expert scan can propose candidate segments, but deciding which proposals are real is left to the learned model.

In the record

structural-biologyMachine learning tool flags protein helices likely to bind DNA or RNAScanned sequences for helix-like windows rich in certain residues, then computed twenty sequence-derived numbers per segment, including charge, water-repellence and helix, sheet and turn propensities.structural-biologyMachine learning sorts plant protein sequences into disease-resistance proteins or notTurned plant resistance protein sequences into numbers using a pairwise residue energy matrix with wavelet and cosine transforms, then trained a stacked classifier on deduplicated sequences.structural-biologyLanguage models trained on protein sequences used to predict sugar-attachment sitesEncoded candidate glycosylation sites as embeddings from three frozen protein language models plus averaged physical parameters, splitting the training data between base learners and a combining learner.structural-biologyProtein language model embeddings tested against classical descriptors for spotting resistance genesEncoded resistance-linked proteins as residue, dipeptide and tripeptide descriptors, as fixed vectors from a pretrained ESM-2 transformer, and as the two concatenated.

Limits

A forest knows only the numbers it is handed. Sequence descriptors flatten three-dimensional geometry, so a model can label a site without any sense of the fold around it. Public protein databases are redundant and skewed towards well-studied families, so accuracy measured on them often overstates how the model behaves on an unfamiliar protein; clustering out similar sequences reduces this but does not remove it. True positives are scarce relative to negatives, which flatters naive scores. And a vote gives a label, not a mechanism, so the result is a hypothesis for experiment rather than evidence.

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