ProtBert-BFD token classification improves intrinsically disordered protein region prediction

Abstract Background Intrinsically disordered regions (IDRs) lack stable tertiary structures yet are crucial to transcriptional regulation, signal transduction, and molecular recognition. Experimental annotation of IDRs remains limited—only ~25,000 disordered proteins are cataloged compared with hundreds of millions of known sequences—highlighting the need for scalable computational prediction. We present a systematic evaluation of classical, deep, and transformer-based models for IDR prediction and introduce a fine-tuned ProtBert-BFD token classification framework that achieves high accuracy and generalizability in data-limited settings. Methods Using manually curated annotations from the DisProt database, redundant sequences were filtered with CD-HIT (<30% similarity) and restricted to ≤526 residues, producing ~700 high-quality sequences. The dataset was partitioned (70:15:15) into training, validation, and test sets. We compared multiple architectures: (1) logistic regression and multilayer perceptron as classical baselines; (2) bidirectional and Seq2Seq LSTMs capturing sequential dependencies; and (3) ProtBert-BFD models leveraging pretrained protein-language embeddings for residue-level classification. Results Classical models showed limited predictive power (AUC 0.56–0.69), while LSTM variants improved recall but overfitted due to data imbalance. ProtBert-BFD fine-tuning substantially enhanced accuracy (75.6%), recall (68.4%), and F1-score (64.0%). The token classification variant achieved precision 0.816, recall 0.823, and F1 0.815—surpassing all internal baselines and approaching the state-of-the-art PROFbval model (recall 0.835). Average inference time was only 1.8 s per 530-residue sequence, enabling proteome-scale deployment. Conclusion These findings demonstrate that pretrained transformers effectively capture disorder-related sequence contexts and outperform conventional deep learning in low-data regimes. By integrating transfer learning with efficient inference, the ProtBert-BFD token classification model provides a robust framework for large-scale IDR annotation. Future work will expand datasets, incorporate physicochemical descriptors, and analyze error distributions to refine understanding of protein disorder in health and disease. References Hu, G., Katuwawala, A., Wang, K. et al. flDPnn: Accurate intrinsic disorder prediction with putative propensities of disorder functions. Nat Commun 12, 4438 (2021).

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.