Machine learning tested against three definitions of protein shape-shifting

The work

Technical · from the record

The authors assembled residue-level datasets of three kinds of protein conformational behaviour — ordered, disordered, and 'ambiguous' residues that either fold upon binding or switch secondary structure — and trained random forest classifiers on seven sequence-based biophysical predictions to separate these classes. Sequence-predicted disorder, early folding propensity and backbone dynamics were the most important features; the model trained on the folding-upon-binding set had its lowest F1 score for the disorder class, and the fold-switching model reached an F1 of 0.36 for converting residues, with a recall of 0.26. Merging the folding-upon-binding and fold-switching categories into one combined model lowered F1 performance for the merged ordered class relative to the separate ordered and same classes. Applying the combined model to a held-out MFIB set of mutually folding proteins predicted 79.6% of residues as ordered, 20% as ambiguous and under 1% as disordered, and comparisons with AlphaFold2 pLDDT values placed ambiguous residues between the ordered and disordered classes.

How AI was used

Residues were labelled from curated structural resources: disorder-to-order transitions from DisProt, conformationally stable residues from CoDNaS clusters filtered at a 2 A maximum pairwise RMSD, secondary-structure switches from DSSP annotations of fold-switcher structure pairs, and a redundancy-filtered MFIB set held out of training. For every sequence, seven per-residue features were predicted with the b2bTools suite — backbone and side-chain dynamics and conformational propensities from DynaMine, early folding propensity from EFoldMine, and disorder from DisoMine — and these features alone, with no amino acid codes, were the inputs to random forest classifiers built with scikit-learn: one on the DisProt/CoDNaS labels split 90/10 into train and test, one on the fold-switching labels, and one on the merged combined set split 70/30, each with hyperparameters chosen by 3-fold cross-validation. The stated aim was interpretability rather than best performance, so feature importances were extracted and each forest was additionally summarised by a surrogate rule model induced with the Weka implementation of the Ripper algorithm over the forest's predictions. The combined model was then run over the MFIB set, and the per-residue predictions for the human proteome were cross-tabulated against AlphaFold2 pLDDT values and DSSP categories from downloaded AlphaFold2 models, against PTM sites compiled from four databases, and against deleterious, benign, somatic and germline missense variants.