Document
1. Datasets
MultiDisoPred is a sequence-based method for residue-level prediction of intrinsically disordered regions (IDRs) and disorder-related binding functions. For IDR prediction, the model was trained on the benchmark dataset constructed by Liu et al.[1], comprising 4,845 protein chains, including 4,229 intrinsically disordered proteins and 616 ordered proteins. For disorder-related binding function prediction, the TR552 training set was adopted from the benchmark constructed by Liang et al. [2], which was derived from experimentally annotated binding information in DisProt [3]; after CD-HIT clustering at 25% sequence identity, 552 proteins were retained. Generalization was evaluated on five independent IDR datasets—MXD494, SL329, DISORDER723, CAID2-Disorder PDB, and CAID3-Disorder PDB—and two binding datasets, CAID2 binding and CAID3 binding. To avoid information leakage, proteins sharing more than 25% sequence identity with any test protein were removed from the training data using CD-HIT [4].
Training Dataset: DM4845_training.fasta TR552_binding.fasta
Independent Dataset: MXD494.fasta SL329.fasta DISORDER723.fasta CAID2_disorder_pdb.fasta CAID3_disorder_pdb.fasta CAID2_binding.fasta CAID3_binding.fasta
2. Feature Extraction Tools
In MultiDisoPred, we employed three types of residue feature extraction methods:
(1) Evolutionary features: Position-Specific Scoring Matrix (PSSM) obtained by three iterations of PSI-BLAST [4] search against the non-redundant NR90 sequence database [5].
(2) ESM-2 protein language model: 1,280-dimensional per-residue embeddings extracted from the ESM-2 model with 650 million parameters [6].
(3) DR-BERT protein language model: 768-dimensional per-residue embeddings extracted from a disorder-focused protein language model pre-trained for disordered-region annotation [7].
3. References
[1] Liu, Y., Wang, X. and Liu, B. RFPR-IDP: reduce the false positive rates for intrinsically disordered protein and region prediction by incorporating both fully ordered proteins and disordered proteins. Briefings in bioinformatics 2021;22(2):2000–2011. [2] Liang, J., et al. Hybrid deep learning with protein language models and dual-path architecture for predicting IDP functions. Briefings in Bioinformatics 2026;27(2). [3] Aspromonte, M.C., et al. DisProt in 2024: improving function annotation of intrinsically disordered proteins. Nucleic Acids Research 2024;52(D1):D434–D441. [4] Li, W. and Godzik, A. Cd-hit: a fast program for clustering and comparing large sets of protein or nucleotide sequences. Bioinformatics 2006;22(13):1658–1659. [5] Holm L, Sander C. Removing near-neighbour redundancy from large protein sequence collections[J]. Bioinformatics, 1998, 14(5): 423-429. [6] Lin Z, Akin H, Rao R, et al. Evolutionary-scale prediction of atomic-level protein structure with a language model[J]. Science, 2023, 379(6637): 1123-1130. [7] Nambiar A, Forsyth J M, Liu S, Maslov S. DR-BERT: A protein language model to annotate disordered regions[J]. Protein Science, 2024, 33(4): e4970.