bio.rodeo
ModelsOrganizationsProvidersLeaderboardAboutSign in
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

Categories
  • DNA & Gene
  • RNA
  • Protein
  • Small molecule
  • Single-cell
  • Spatial omics
  • Pathology
  • Imaging
  • Metabolomics
  • Biosignals
  • Language model
bio.rodeoModelsOrganizationsProvidersLeaderboardAboutFAQSubmit a modelContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / protein / idpfunnet
Protein
Shandong UniversityReleased May 2025

IDPFunNet

Intrinsically disordered region function prediction, scoring every residue for five binding subtypes plus disordered flexible linkers.

5Openness

Where to run it

No providers recorded yet. Browse all providers

IDPFunNetProteinShandong University

Roughly a third of eukaryotic proteome residues sit in intrinsically disordered regions — stretches with no fixed three-dimensional fold. They are not inert. IDRs mediate a large share of regulatory protein-protein interactions, bind nucleic acids, lipids, ions, and small molecules, and serve as flexible linkers that let folded domains move relative to one another. Because they have no stable structure, the entire apparatus of structure-based function annotation fails on them, and predicting which function a given disordered region performs has proven harder than predicting disorder itself.

IDPFunNet, from Zhenling Peng's group at Shandong University, addresses that second problem directly. It assigns each residue in a sequence a propensity for six distinct disorder functions: binding to proteins, nucleic acids, lipids, ions, or other small molecules, and acting as a disordered flexible linker (DFL). Existing tools cover subsets of this space — DeepDISOBind handles protein, DNA, and RNA binding, DisoLipPred specializes in lipid binding, DisoFLAG covers linkers and several binding types — but few produce all six from one model, and cross-dataset stability has been a persistent weakness.

The architectural bet is that binding and linker prediction are different enough to warrant separate pathways over a shared representation. Both paths read frozen ProtT5 embeddings, which the authors found outperform ESM-2-family embeddings and AlphaFold 2-derived structural features by at least 1.3% average AUC and 12.7% average APS on this task. The work was published in Briefings in Bioinformatics in 2026.

#Key Features

  • Six functional classes from one pass: A single forward pass yields per-residue scores for five binding subtypes and disordered flexible linkers, rather than requiring a separate specialized predictor per interaction type.
  • Dual-path decoupling: A hybrid CNN-BiLSTM path handles binding IDRs while a pure BiLSTM path handles linkers, preventing the two tasks from interfering while still sharing the underlying language model representation.
  • Multi-task gains: Training the five binding heads jointly improved protein-, lipid-, and small-molecule-binding AUC by 3.1-35.1% over single-task equivalents.
  • Blind-benchmark validation: Evaluated on the CAID2 and CAID3 community challenge datasets, which share less than 25% sequence identity with the training set and were never seen during development.
  • Public web server: A hosted predictor accepts sequences directly, alongside a standalone package on GitHub for local batch use.

#Technical Details

Each residue is encoded as a 1024-dimensional ProtT5 vector. The binding path applies a two-layer CNN with kernel size 3 for local motif detection followed by a bidirectional LSTM for long-range context, with each submodule emitting a 512-dimensional feature vector; the linker path uses BiLSTM alone. Six residue-level classifiers — five parallel two-layer residual MLP blocks for the binding subtypes plus one for linkers — convert features to sigmoid propensity scores. Training data derive from DisProt 9.4: 1,073 annotated sequences, clustered at 25% identity by CD-HIT into 814 clusters and split into TR552 (training), VA227 (validation), and TE210 (test). Five further independent sets are used for evaluation: TE83, comprising 83 sequences deposited after DisProt 9.4 but before 9.6, and the four official CAID blind sets (CAID2_Bind with 78 proteins, CAID2_DFL with 40, CAID3_Bind with 51, CAID3_DFL with 20). Protein-binding AUC ranges 0.832-0.866 across benchmarks, at least 1.5% above prior methods with p < 0.05; lipid binding on TE210 reached 0.896 AUC and 0.436 APS, exceeding DisoFLAG, DeepDISOBind, and DisoLipPred by at least 3.6% AUC. On CAID3_Bind it outperformed the top five challenge entries by 0.03-0.12 AUC.

#Applications

The tool serves anyone annotating an uncharacterized proteome or prioritizing regions of a disordered protein for experimental follow-up. Identifying which stretch of a disordered tail is likely a protein-binding motif versus a passive linker guides construct design for pull-downs, peptide array screens, and NMR titrations, and complements structure- and ensemble-oriented IDR models such as Disobind, which predicts inter-protein contacts for disordered complexes, and AI-IDP, which generates conformational ensembles. Because the predictor runs from sequence alone with no MSA or structure requirement, it scales to whole proteomes and to designed or synthetic sequences with no evolutionary neighbours.

#Impact

IDPFunNet's most durable contribution may be its ablation evidence rather than its headline scores: it establishes on a controlled benchmark that ProtT5 embeddings carry more IDR-function signal than ESM-family embeddings or AlphaFold 2 structural features, which is a useful negative result for a field that has increasingly reached for structure predictors by default. The model is candid about where it falls short. Nucleic-acid binding trails DisoFLAG on TE210 (0.775 versus 0.864 AUC), ion binding suffers from severe class imbalance — TE83 contains fewer than ten positive examples — and DFL prediction fluctuates across datasets, a pattern the authors attribute partly to incomplete linker annotation in DisProt itself. The GitHub repository commits both trained networks and the four annotated benchmark datasets but carries no license file, which leaves reuse terms unstated, and the preprint is distributed under CC BY-NC-ND.

At a glance

Released
May 2025
Category
Protein
Organization
Shandong University

Links

GitHub RepositoryResearch PaperDemo

Tags

bilstmcnnfunction_predictionintrinsically_disordered_proteinsmulti_taskprotein_binding_predictiontransfer_learning

Something wrong?

Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.