ORIGINAL RESEARCH article
Front. Genet.
Sec. Computational Genomics
Volume 16 - 2025 | doi: 10.3389/fgene.2025.1616880
BioSemAF-BiLSTM: A Protein Sequence Feature Extraction Framework Based on Semantic and Evolutionary Information
Provisionally accepted- 1Ocean University of China, Qingdao, China
- 2School of Medicine, Nankai University, Tianjin, China
Select one of your emails
You have multiple emails registered with Frontiers:
Notify me on publication
Please enter your email address:
If you already have an account, please login
You don't have a Frontiers account ? You can register here
S-sulfenylation is a critical post-translational modification that plays an important role in regulating protein function, redox signaling, and maintaining cellular homeostasis. Accurate identification of S-sulfenylation sites is essential for understanding its biological significance and relevance to disease. However, the exclusive detection of S-sulfenylation sites through experimental methods remains challenging, as these approaches are often time-consuming and costly. Motivated by this issue, the present work proposed a deep learning-based computational framework, named BioSemAF-BiLSTM, which integrated evolutionary and semantic features to improve the prediction performance of S-sulfenylation sites. The framework employed fastText to generate subword-based sequence embeddings that captured local contextual information, and employed position-specific scoring matrices (PSSMs) to extract evolutionary conservation features. Importantly, we also quantitatively evaluated feature sufficiency at the protein sequence level using a sequence compression-based measure approximating Kolmogorov complexity, revealing an 11% information loss rate in predictive modeling using these features. These representations were subsequently fed into a bidirectional long short-term memory (BiLSTM) network to model long-range dependencies, and were further refined via an adaptive feature fusion module to enhance feature interaction. Experimental results on a benchmark dataset demonstrated that the model significantly outperformed conventional machine learning methods and current state-of-the-art deep learning approaches, achieving an accuracy of 89.32% on an independent test. It demonstrated improved sensitivity and specificity, effectively bridging the gap between bioinformatics and deep learning, and offered a robust computational tool for predicting post-translational modification sites.
Keywords: Adaptive fusion, bidirectional LSTM neural network, bioinformation encoding, information loss, Kolmogorov complexity, post-translational modification, protein sequence embedding, sulfenylation site
Received: 30 Apr 2025; Accepted: 01 Sep 2025.
Copyright: © 2025 Zhang and Wang. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
* Correspondence: Zihan Zhang, Ocean University of China, Qingdao, China
Disclaimer: All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.