Per-residue intrinsic disorder prediction from sequence, using a convolutional ensemble trained on missing residues in PDB X-ray structures.
No providers recorded yet. Browse all providers
Residues that are present in a crystallised protein but leave no electron density are the most abundant experimental record of intrinsic disorder: the PDB holds thousands of such stretches, far more than the few thousand disordered regions curated by other means. PredIDR is a convolutional network trained to reproduce exactly that signal. Given a protein sequence, it scores each residue for its propensity to be disordered in the X-ray sense, missing from the solved structure.
The harder problem is building a fair training set from such data. Disordered residues make up under 8% of the PDB chains used, and disordered stretches sit in characteristic places — at the N-terminus, the C-terminus, or in the middle of a chain. Undersampling the ordered residues at random would teach the network where regions tend to fall rather than what they look like. PredIDR instead sorts every disordered region by its position and pairs it with an ordered segment from the same position class, cut to the same length, building extra N-terminal ordered segments from fully structured chains where there were too few. The result is a perfectly balanced set whose positives and negatives share length and position distributions.
PredIDR was developed by Kun-Sop Han and colleagues at the University of Sciences in Pyongyang and the Branch of Biotechnology of the State Academy of Sciences, with Alessio Del Conte and Damiano Piovesan of the University of Padova, who organise the CAID assessment. It was entered in the second round of CAID in 2022, posted as a preprint in July 2024 and published in the International Journal of Biological Macromolecules in 2025. It is the first generation of a line that continued with PredIDR2 and PredIDR3.
Training data came from 8,657 non-redundant, high-resolution PDB chains (August 2019 snapshot, under 25% identity by CD-HIT, longer than 51 residues). Of these, 5,400 chains with disorder plus 2,660 fully structured chains built the training set of 114,876 disordered and 114,876 ordered residues, and 597 held-out chains formed the Val597 validation set. Each residue is represented by a 32-by-15 tensor over a 15-residue sliding window: 21 profile frequencies, the SS8 vector, one accessibility value and two terminus flags. The network has two convolutional layers of 32 and 64 3x1 filters with ReLU, 2x1 max-pooling, a 128-unit dense layer and a single output neuron, trained with Adam on a mean-squared-error loss from a learning rate of 0.005. Twenty networks were trained and the ten best on Val597 averaged. In the feature ablation, predicted solvent accessibility alone reached AUC_ROC 0.921 on Val597, and the three features together 0.940. On the CAID2 Disorder-PDB set, PredIDR_long reached AUC_ROC 0.933 and AUC_PR 0.870, ranking first by AUC_ROC among 28 methods that returned predictions for all 348 sequences. On an X-ray-only subset the authors built from CAID2, the two outputs placed first and second; on the complementary subset of non-X-ray disorder annotations, PredIDR_long fell to seventh.
PredIDR serves structural biologists who need to know which segments of a target are likely to be missing from a crystal structure: trimming constructs before crystallisation trials, interpreting gaps in solved structures, and flagging flexible termini and loops before expression. The short and long outputs let users favour sensitivity to brief disordered loops or to extended disordered domains. It can be run alongside other CAID methods such as Metapredict through the CAID Prediction Portal, or locally from the Singularity container in the BioComputingUP shared directory, which the authors name as the two distribution channels.
PredIDR showed that a compact convolutional network on classical profile and predicted-structure features could match the top of the CAID2 field when judged on the Disorder-PDB reference. Its authors are also candid about what that training signal cannot capture: on disorder evidenced by NMR, circular dichroism and other non-crystallographic methods it drops well behind the best predictors, and they attribute this to training on X-ray missing residues alone. That diagnosis shaped its successors, which add DisProt annotations to the training data. No standalone source-code repository or separate weight release exists; the model is distributed only as the prebuilt container and the portal service.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.