PredIDR3

A sliding-window disorder predictor scores one residue at a time: crop a window of amino acids around the target, push the window through a network, read a single number off the output. That single number is the bottleneck. Widen the window and more context arrives at the same lone output neuron — under the conventional encoding, accuracy on the CAID3 reference set peaks at a window of 51 residues and falls away beyond it, which is why disorder predictors have generally settled on windows of 11 to 17 residues.

PredIDR3 changes what the window is asked to predict. Its final fully connected layer carries one neuron per window position, so a single forward pass emits a disorder call for every residue in the window rather than only for the one at its centre. Each residue is then covered by up to W overlapping windows, and its score is the average of those W calls. With the output widened this way, accuracy keeps climbing out to a window of 91 residues.

The second change is where the negative examples come from. Earlier versions drew ordered residues only from the structured parts of PDB chains; PredIDR3 also draws them from the non-annotated regions of DisProt, and a 50:50 mix of the two sources trains a measurably better model than either alone. The work comes from Kun-Sop Han and colleagues at the University of Sciences in Pyongyang, with co-authors at Pyongsong University of Education, the University of Science and Technology, and the University of Padova, and is the third generation of the PredIDR line.

#Key Features

• Window-wide output encoding: The output layer has one neuron per window position, so the network learns the whole window's disorder state and each residue's score pools every window covering it.
• A 91-residue receptive field: Six times the span of the 15-residue windows used by the previous version, which lets local patterns be read in the context of a long surrounding stretch.
• Negatives from two databases: Ordered training residues come from non-IDR regions of both the PDB and DisProt rather than the PDB alone, which supplies much of the gain over PredIDR2.
• Two ensembles for two error profiles: PredIDR3_BigMCC maximises Matthews correlation and recovers more disordered residues; PredIDR3_Weaker cuts the false-positive rate from 3.95% to 2.37% at essentially the same AUC.
• No language model in the loop: Inputs are a sequence profile plus predicted secondary structure and solvent accessibility, all from one SCRATCH 1.2 run, replacing the PSI-BLAST search against UniRef50 earlier versions needed.

#Technical Details

Each window becomes a 31-by-W input tensor — 20 sequence-profile values, an eight-state secondary-structure vector, one solvent-accessibility value and two flags marking positions past the sequence termini. Four blocks of 32 3x1 convolutional filters with ReLU activation and batch normalisation are followed by a 32-unit dense layer and a sigmoid output layer of W units, trained with Adam on a mean-squared-error loss at a learning rate of 0.01 and a batch size of 128. Thirty networks were trained and the ten best on validation averaged into each ensemble. Training used 7,299 chains — 5,344 from the PDB and 1,955 from DisProt release 2025_12 — clustered at 25% identity with CD-HIT, yielding a balanced set of 616,952 residues. The 271 CAID2 chains served as validation and the 319 CAID3 chains as a held-out test set. On the CAID3 Disorder-PDB reference set, PredIDR3_BigMCC reaches MCC 0.790, AUC_ROC 0.953, AUC_PR 0.925 and F-max 0.853, against 0.697 and 0.936 for PredIDR2, and its mean absolute error in per-protein disorder content, 5.896, is the lowest among the thirty top CAID3 methods the authors compared against.

#Applications

Per-residue disorder annotation is the entry point to most work on intrinsically disordered proteins: deciding which segments of a target will resist crystallisation, choosing constructs for NMR or SAXS, mapping the disordered fraction of a proteome, and finding the flexible regions implicated in aggregation diseases and in drug discovery against undruggable targets. The low disorder-content error matters most for proteome-scale surveys, where the quantity of interest is each protein's disordered fraction rather than exact region boundaries. It is the same annotation Metapredict produces from sequence.

#Impact

PredIDR3 argues that the output side of a predictor deserves as much attention as the input side: the same convolutional backbone, given a wider output and a wider window, matches or beats most of the protein-language-model methods in the CAID3 field without using embeddings at all, and reaches disorder-content accuracy none of them match. The work is a preprint and has not been peer reviewed. The authors designate the CAID Prediction Portal and a Singularity container in the BioComputingUP shared directory as the distribution channels; the portal's method list currently carries the PredIDR2 entries, while the shared directory holds only a PredIDR container dated March 2024, and no PredIDR3 source code or trained weights have been released separately.