Intrinsic protein disorder and conditional folding in AlphaFoldDB

Abstract

Intrinsically disordered regions (IDRs) defying the traditional protein structure-function paradigm have been difficult to analyze. The availability of accurate structure predictions on a large scale in AlphaFoldDB offers a fresh perspective on IDR prediction. Here, we establish three baselines for IDR prediction from AlphaFoldDB models based on the recent CAID dataset. Surprisingly, AlphaFoldDB is highly competitive for predicting both IDRs and conditionally folded binding regions, demonstrating the plasticity of the disorder to structure continuum.

Keywords: CAID; critical assessment; intrinsically disordered proteins; machine learning; protein structure; structural bioinformatics.

FIGURE 1

Example of intrinsically disordered regions and conditional folding predictions derived from AlphaFold and best predictors. The human Ephrin‐B2 protein (UniProt accession: P52799) is shown as a representative example to illustrate the overlap between AlphaFold predictions and various sequence features. Panel a, the structure of the protein predicted by AlphaFold and colored by pLDDT score (<50 orange, <70 yellow, <90 light blue, >90 dark blue). Residue labels indicate annotated region boundaries. Panel b, database annotations (DisProt DP01588, PDB, InterPro P52799) and predicted regions (AlphaFold‐pLDDT, AlphaFold‐RSA, AlphaFold‐Bind, fIDPnn, ANCHOR‐2 ¹⁹ ). PDB annotation is generated by combining observed residues in different PDB experiments. Best predictors were selected based on their performance against DisProt and DisProt‐binding references. Annotated regions are shown colored according to the legend on top of panel b (i.e., disorder in red, binding in gold, structure in blue, other features in gray, while white regions correspond to no annotation. Per‐residue AlphaFold predictions are provided in Figure S1

FIGURE 2

Results for AlphaFold on the three main CAID categories. The results for the DisProt‐PDB (n = 646 proteins, panels a,b), DisProt (n = 646 proteins, panels c, d), and DisProt‐binding (n = 646 proteins, panels e, f) references are shown. Performance of predictors expressed as maximum F1‐Score across all thresholds (F _max) (panels a, c, e) and AUC (panels b, d, f) for AlphaFold (colored), the top 10 best ranking methods (gray) and baselines (white symbols) are shown. The legend on the right of each panel shows the name of the method alongside its F _max or AUC score (f and a, respectively) and coverage (c). Notice how the latter is usually 1.0 for most predictors, but only 0.76 for AlphaFold as predictions for some targets are not available