Per-residue intrinsic disorder prediction from sequence: four convolutional ensembles trained on PDB missing residues plus curated DisProt regions.
No providers recorded yet. Browse all providers
A disorder predictor learns what "disordered" means from whichever experimental record it is trained on. The first PredIDR learned it from residues missing from X-ray crystal structures in the PDB, an abundant signal but only one kind of disorder. It matched the best CAID2 methods on crystallographic disorder and fell well behind them on disorder seen by NMR, circular dichroism and other solution methods. PredIDR2 tackles that gap by widening the definition of a positive example. Its disordered training residues come from both PDB missing residues and the curated, literature-derived regions of the DisProt database, while its ordered residues still come from the structured parts of PDB chains.
The second change is the network. PredIDR's shallow two-layer convolutional network is replaced by a deeper stack of four convolutional blocks, each with batch normalisation. The authors credit this network and the DisProt-augmented training set for most of the gain over the first generation. PredIDR2 is released as a series of four methods whose variants differ in how each residue is encoded and in how the ordered (negative) training samples are produced.
PredIDR2 was developed by Kun-Sop Han and colleagues at the University of Sciences in Pyongyang, the Branch of Biotechnology of the State Academy of Sciences and the University of Science and Technology. It was published in the International Journal of Biological Macromolecules in March 2025. It was the group's entrant in the third round of the Critical Assessment of protein Intrinsic Disorder (CAID3), and it sits between PredIDR and PredIDR3 in the PredIDR line.
The four methods are PredIDR2-Prof-Art, PredIDR2-Prof-Rnd, PredIDR2-Seq-Art and PredIDR2-Seq-Rnd. Each residue is represented by a feature vector over a sliding window. The Prof variants use a sequence profile built from a PSI-BLAST search against UniRef50, together with predicted eight-state secondary structure and 20-class solvent accessibility. The network produces one disorder score for the residue at the centre of the window. The PDB portion of the training data reuses the 8,657 non-redundant chains (August 2019 snapshot, under 25% sequence identity) assembled for PredIDR. DisProt release 2023_06 supplies the additional 2,408 sequences of disordered regions, and the CAID2 targets were held out for testing. On the CAID2 Disorder-PDB reference set the series reached AUC_ROC 0.952, against 0.933 for PredIDR. It led the CAID2 field on AUC_ROC, AUC_PR and per-protein disorder-content error, and ranked among the top seven methods by MCC. On the separate, blind CAID3 Disorder-PDB set, three of the four variants placed in the top ten, with AUC_ROC values of 0.936, 0.936 and 0.934, as reported in the PredIDR3 paper. PredIDR2-Seq-Art was the strongest of the series there, ahead of PredIDR-long by 0.032 in MCC and 0.012 in AUC_ROC.
Per-residue disorder calls guide construct design before crystallisation or cryo-EM, the choice of regions for NMR and SAXS, the reading of gaps in solved structures, and estimates of how much of a proteome is disordered. Training on DisProt as well as the PDB exposes it to disorder evidenced beyond crystallography. The faster Seq variants suit larger batches. The CAID methods registry lists all four variants, and the authors name the CAID Prediction Portal as the way to run them, alongside other CAID methods such as Metapredict. In practice that route is not currently available: the portal's configuration marks its PredIDR2 entries as not runnable.
PredIDR2 showed that adding curated DisProt annotations to PDB-derived training data lifts a compact convolutional predictor that uses no protein language model above every CAID2 participant in AUC_ROC on the Disorder-PDB set, in a comparison made after that round closed. Its own limitation set the agenda for the next version: its negatives still came only from the PDB, and PredIDR3 found that drawing ordered residues from DisProt as well was one of the two largest sources of further improvement. The authors also name a Singularity container in the BioComputingUP shared directory as a download route, but that directory holds only the first-generation PredIDR container. No PredIDR2 source code or trained weights have been released separately.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.