ProtAttn-QuadNet: An attention-based deep learning framework for protein–protein interaction prediction using ProtBERT embeddings

RESEARCH ARTICLE ProtAttn-QuadNet: An attention-based deep learning framework for protein–protein interaction prediction using ProtBERT embeddings Md. Shahidul Islam *, Md. Muhtasim Rahman Mim , Md. Raihan Kabir Department of Computer Science and Engineering, University of Asia Pacific, Dhaka‌‌, Bangladesh * Abstract OPEN ACCESS Citation: Islam MS, Mim MMR, Kabir MR (2026) ProtAttn-QuadNet: An attention-based deep learning framework for protein–protein interaction prediction using ProtBERT embeddings. PLoS One 21(6): e0349433. https://doi.org/10.1371/journal.pone.0349433 Editor: Musa Aydin, Samsun University: Samsun Universitesi, TÜRKIYE Received: November 26, 2025 Accepted: April 30, 2026 Published: June 2, 2026 Peer Review History: PLOS recognizes the benefits of transparency in the peer review process; therefore, we enable the publication of all of the content of peer review and author responses alongside final, published articles. The editorial history of this article is available here: https://doi.org/10.1371/journal. pone.0349433 Copyright: © 2026 Islam et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, Protein–protein interactions (PPIs) form the backbone of most cellular processes, governing signal transduction, gene regulation, and metabolic control. However, experimental approaches to identifying PPIs remain expensive, laborious, and often incomplete. Recent advances in protein language models (PLMs) have transformed sequence-based PPI prediction by enabling deep contextual encoding of biochemical and structural information directly from amino acid sequences. Building upon this progress, we present ProtAttn-QuadNet, an attention-based deep learning framework that leverages ProtBERT embeddings to model reciprocal dependencies between protein pairs. The proposed model employs a quad-stream attention mechanism that integrates individual protein features, synergistic interactions, and complementary differences through multi-level self- and cross-attention layers. This architecture enables the discovery of fine-grained relational patterns while ensuring balanced bidirectional modeling of interacting proteins. Evaluated on the independent test set of a large-scale dataset from UniProt, ProtAttn-QuadNet achieves 97.16% accuracy (AUC-ROC 99.00%) on balanced data and 99.19% accuracy (AUC-ROC 99.76%) on oversampled datasets, surpassing several recent state-of-the-art PPI prediction methods. Statistical validation using the Chi-square and Wilcoxon signed-rank tests confirms the model’s predictive significance and reliability. ProtAttn-QuadNet offers a powerful computational framework for large-scale PPI prediction. Introduction Protein–protein interactions (PPIs) are fundamental to almost all cellular processes, including signal transduction, gene expression regulation, metabolic control, and immune responses [1–3]. Understanding the complex network of PPIs provides valuable insights into cellular functions and disease mechanisms [4,5]. Although numerous experimental techniques, such as yeast two-hybrid screening, PLOS One | https://doi.org/10.1371/journal.pone.0349433 June 2, 2026 1 / 16 and reproduction in any medium, provided the original author and source are credited. Data availability statement: The primary data are available from UniProt (https://www. uniprot.org/uniprotkb?query=reviewed:true). All reviewed (Swiss-Prot) entries from UniProtKB were used in this study, comprising 573,661 protein sequences. The processed data and code supporting this study are publicly available on Figshare and can be accessed through the following link: https://doi.org/10.6084/ m9.figshare.30637145. Funding: The author(s) received no specific funding for this work. Competing interests: The authors have declared that no competing interests exist. co-immunoprecipitation, and affinity purification coupled with mass spectrometry, have been developed to detect PPIs, these methods remain time-consuming, costly, and often limited in coverage [6]. Consequently, computational prediction methods have become indispensable for large-scale PPI analysis. Early computational approaches primarily relied on handcrafted sequence features, including amino acid composition, evolutionary profiles, and physicochemical descriptors. Classical machine learning algorithms such as Support Vector Machines (SVM), Random Forests (RF), and Bayesian classifiers were employed to classify interacting protein pairs based on these features [7–12]. While these models demonstrated moderate success, their dependence on manually engineered descriptors and incomplete structural data limited their generalization capabilities, particularly across species and diverse protein families [11,13–15]. The increasing availability of large-scale protein sequence databases has encouraged sequence-based prediction methods that rely less on structural information. Deep learning has substantially advanced this field by enabling hierarchical feature extraction and representation learning. DeepPPI [16] used a fully connected neural network to model complex non-linear relationships between protein features, whereas DPPI [17] applied a Siamese-like convolutional architecture to learn symmetric relationships between interacting proteins. Similarly, PIPR [18] introduced a residual recurrent convolutional neural network (RCNN) to capture both local motifs and long-range dependencies, while Wu et al. proposed DL-PPI [19], a graph neural network–based model that integrates multi-scale features and attention mechanisms to enhance relational reasoning among proteins. These architectures collectively improved predictive performance but often struggled with interpretability, data imbalance, and computational efficiency. Recent advances in transformer architectures and PLMs have transformed sequence-based PPI prediction by learning contextualized residue representations through self-attention mechanisms. Pretrained models such as ProtTrans [5], ProtBERT [20], and ESM-2 [21] encode rich biochemical and evolutionary information from massive unlabeled protein corpora, effectively capturing secondary and tertiary structure tendencies directly from primary sequences. Several recent studies have leveraged these embeddings for PPI prediction using hybrid deep architectures. For example, xCAPT5 [22] integrated ProtTrans embeddings with a multi-kernel convolutional network to capture local and global dependencies, while TUnA [23] incorporated uncertainty modeling within a transformer framework to improve robustness. PPI-Graphomer [24] combined pretrained language models with graph transformers to integrate sequence and structural representations, achieving high performance across benchmark datasets. Despite these advances, existing frameworks often treat protein pairs asymmetrically and fail to explicitly model the reciprocal dependencies inhe (...truncated)