EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

DPANet: Dual-Encoder Protein Attention Network for Secondary Structure Prediction

Sina Abbaskhani, Masoud Kargar, Ali Bayani, Alireza Assadzadeh

IEEE Access · 2026 · 14 · 105424-105441

DOI: 10.1109/access.2026.3704132 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Sina Abbaskhani und 3 weiteren Autor:innen. Die Arbeit nennt 100 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Protein secondary structure prediction is essential for understanding protein function and informing downstream structural modeling. We introduce DPANet, a dual-encoder protein attention network that integrates primary amino acid sequences and evolutionary profiles (PSSMs) to achieve high-accuracy secondary structure prediction. DPANet employs parallel encoders combining multi-head self-attention with LSTM layers, enabling the model to capture both local residue interactions and long-range contextual dependencies. The encoder outputs are fused and processed by a multi-level attentive decoder, facilitating precise Q3 and Q8 secondary structure classification. The model was extensively evaluated on three benchmark datasets: TS115, CB513, and PDB. On TS115, DPANet achieved 97.07% SOV, 0.964 Accuracy, and 0.936 Cohen’s Kappa for Q3 classification, and 93.98% SOV and 0.927 Accuracy for Q8. For CB513, it reached 99.01% SOV and 0.988 Accuracy in Q3, with 96.77% SOV and 0.970 Accuracy for Q8. On the large-scale PDB dataset, the model attained 98.21% SOV, 0.984 Accuracy, and 0.976 Cohen’s Kappa for Q3, and 98.97% SOV, 0.993 Accuracy, and 0.990 Cohen’s Kappa for Q8. These results underscore the effectiveness of the dual-encoder design and attention mechanisms in integrating sequential and evolutionary information, enhancing discrimination among secondary structure classes, and preserving structural continuity across entire sequences. DPANet also provides residue-level interpretability, robust performance in low-homology regions, and computational efficiency suitable for large proteomic datasets, making it a powerful tool for proteome-wide secondary structure annotation.

Zitieren

APA

Abbaskhani, S., Kargar, M., Bayani, A., & Assadzadeh, A. (2026). DPANet: Dual-Encoder Protein Attention Network for Secondary Structure Prediction. IEEE Access, 14, 105424-105441. https://doi.org/10.1109/access.2026.3704132

Vancouver

Abbaskhani S, Kargar M, Bayani A, Assadzadeh A. DPANet: Dual-Encoder Protein Attention Network for Secondary Structure Prediction. IEEE Access. 2026;14:105424-105441. doi:10.1109/access.2026.3704132

BibTeX
@article{Abbaskhani2026,
  author = {Abbaskhani, Sina and Kargar, Masoud and Bayani, Ali and Assadzadeh, Alireza},
  title = {DPANet: Dual-Encoder Protein Attention Network for Secondary Structure Prediction},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {105424--105441},
  doi = {10.1109/access.2026.3704132}
}
Kostenlos starten →