Disobind

Intrinsically disordered proteins and regions (IDPs/IDRs) make up a large fraction of eukaryotic proteomes and mediate many interactions without adopting a single fixed structure. Their binding modes span coupled folding-and-binding, fuzzy binding, and fully disordered binding, and the resulting complexes are hard to characterize experimentally and harder still to predict. Structure predictors are the default tool, yet one predicted conformation poorly describes an ensemble, and their confidence metrics degrade precisely where disorder is highest.

Disobind, from the Viswanath lab at the National Centre for Biological Sciences, reframes the problem to sidestep that mismatch. Instead of predicting a structure, it takes two sequences — at least one an IDR — and predicts a binary inter-protein contact map and the interface residues on each partner. It is built on frozen embeddings from ProtT5, part of the ProtTrans suite, and needs no multiple sequence alignment and no structural input. First posted to bioRxiv in December 2024, it was published in Cell Systems in 2026.

The distinguishing property is partner dependence. Disorder-focused interface predictors such as AIUPred, MORFchibi, and DeepDISOBind score an IDR in isolation; Disobind conditions on the specific partner, which is what makes its output usable as a distance restraint in integrative modelling of large assemblies — the authors' stated motivation.

#Key Features

• Partner-dependent interface prediction: The prediction is made for a specific IDR–partner pair rather than a partner-agnostic propensity score along the IDR.
• No MSAs, no structures: Sequence embeddings alone drive the model, so it applies to regions with few homologs and no experimental coordinates.
• Coarse-grained output resolutions: Predicting contacts over blocks of 1, 5, or 10 contiguous residues reduces output sparsity and class imbalance, and matches the modular, motif-driven character of IDR interfaces.
• Merged contact-map targets: Contacts from all available structures of a given sequence pair are combined with a logical OR, giving one target that spans multiple bound conformations instead of assuming a single binding mode.
• Composable with structure predictors: A logical-OR combination with AlphaFold-Multimer predictions beats either method alone, and the same recipe extends to any predictor that reports pLDDT and PAE.

#Technical Details

The architecture is deliberately shallow, which is what frozen language-model embeddings make possible. Input embeddings are projected to a lower dimension, then an interaction block computes the outer product and outer difference of the two projected sequences and concatenates them into an interaction tensor. An MLP maps that tensor to a contact map; for interface residues, the tensor is first reduced by row-wise and column-wise averaging. Outputs are element-wise sigmoid scores thresholded at 0.5. Training uses the singularity-enhanced loss, a class-imbalance-adapted variant of binary cross-entropy, with AdamW. Global embeddings — computed over the full protein, then sliced to the fragment — outperform fragment-local ones, and ProtT5 outperformed ProtBERT and ProSE.

The dataset draws on DIBS, MFIB, FuzDB, PDBtot, PDBcdr, DisProt, IDEAL, and MobiDB. Chains are split at missing residues and cropped to 100 residues for memory during training; this is a training crop only, and inference imposes no length limit. An out-of-distribution test set of 52 merged binary complexes shares under 20% identity with both the Disobind data and the AlphaFold2 PDB70 set, with 297 in-distribution complexes held out separately. At single-residue resolution Disobind reaches F1 0.57 (in-distribution) and 0.33 (out-of-distribution) for contact maps, and 0.70 and 0.48 for interface residues; coarse-graining raises these substantially. It beats AlphaFold-Multimer and AlphaFold 3 at ipTM cutoffs of 0.75, 0.4, and 0.0, with AlphaFold-Multimer the stronger of the two — only 8 of 52 AlphaFold-Multimer and 2 of 52 AlphaFold 3 out-of-distribution predictions were high confidence.

#Applications

The intended downstream use is integrative structural modelling, where IDRs such as the FG nucleoporins of the nuclear pore complex, the MBD3 IDR in the NuRD complex, and the Plakophilin-1 N-terminus are the regions that make models imprecise. Disobind or Disobind+AlphaFold-Multimer contacts can enter IMP, HADDOCK, or Assembline as inter-protein distance restraints, or seed molecular dynamics. Proteome-scale sweeps for IDR-mediated interactions and motif discovery are a second use, and predicted interfaces can suggest mutations for perturbing an interaction.

#Impact

Disobind establishes that a shallow head on frozen protein-language-model embeddings can beat general-purpose structure predictors on a task those predictors were never well suited to, and that combining the two beats either. The authors are explicit about the boundaries: it handles binary complexes only, so larger assemblies require combining pairwise predictions; it assumes the pair binds, and cannot separate binders from non-binders — as a protein–protein interaction classifier on 2,355 pairs it reaches AUROC 0.48–0.50, no better than chance; and it cannot model post-translational modifications, since ProtT5 does not distinguish modified residues. Code, trained weights, and a Colab notebook running Disobind with AlphaFold2 are released under GPL-3.0, with datasets and AlphaFold predictions on Zenodo.