BAN-SDBPred: Improving Single-Stranded and Double-Stranded DNA-Binding Protein Prediction Using an Attention Network with Bilinear Convolution and Adaptive Sampling Strategy Authors: Kamran Arshad, Muhammad Arif, Apilak Worachartcheewan, Dong-Jun Yu Published: 2026 ABSTRACT DNA binding proteins play essential roles in numerous biological mechanisms. The DBPs can be either single-stranded binding (SSBs) or double-stranded binding (DSBs) to a DNA molecule. The in-depth identification of SSBs and DSBs has been a hot topic in bioinformatics and is involved in the drug discovery process. Traditional experimental methods failed to characterize the types of DBPs because of high cost and time constraints. While computational prediction of novel SSBs and DSBs has made significant progress, there are still challenges remaining in enhancing overall prediction performance. Methods: Here, we develop a novel BAN-SDBPred (Bilinear Attention Network for Single and Double Stranded DNA-Binding Protein Prediction) method. BAN-SDBPred leverages the evolutionary features by protein language model-based Evolutionary Scale Modeling 2 (ESM2), ProtT5, and a histogram of oriented gradient-based residue pairwise energy content matrix (RECM-HOG)-transformed energy estimation features from sequence alone. Then, the adaptive neighborhood-based sampling (ANBS) algorithm was adopted to solve the imbalance issue. Compared to other deep learning models, the bilinear attention network (BAN) learns the local and global enriched features from the sequences. Extensive experimental results anticipate that BAN-SDBPred outperforms the existing predictors in terms of all performance measures, such as Acc, F1, MCC, etc., on the training and independent test data. Our designed model has significant advantages in discriminating SSBs and DSBs from DBPs with an improved Acc of 2%, Precision of 4.5%, F1 of 21%, MCC of 9%, and area under curve (AUC) of 20%, respectively. We expect this research will help to predict large-scale novel SSBs and DSBs in particular and other binding problems in general. All data and models are available at 10.5281/zenodo.18718092. Introduction DNA-binding proteins (DBPs) mediate protein-DNA interactions and serve as essential mediators in cellular processes including gene expression regulation, DNA repair, replication, and recombination. DBPs reflect the combined effects of intermolecular forces i.e., electrostatic forces, van der Waals, hydrogen bonding, and hydrophobic effects with protein structural properties. − Past studies have proven that DBPs are closely associated with for example deadly diseases such as cancer (liver, lung, and breast cancer), chronic diseases (asthma, HIV AIDS), and antibiotics. DBPs are categorized into two functional groups from broad spectrum: double stranded DBP (DSBs) and single stranded DBP (SSBs). − The fundamental difference between these two classes of DBPs is due to their biological function, structural-based specificity of DNA, and interaction mechanism. DSBs are involved in numerous cellular-based functions like DNA repair, transcription factors, or maintaining genomic stability. In contrast, SSBs are essential for DNA repair, recombination, and replication processes and are capable of binding with high affinity to single-stranded DNA. , Therefore, understanding the binding activity of DBPs and their functional type is critical for advance precision medicine, antibody–antigen interactions, and drug discovery. Conventional-based experimental techniques such as X-ray crystallography, nuclear magnetic resonance (NMR), cryo-electron microscopy (cryo-EM), and filter binding assays have been used to investigate the binding affinity of DBPs as SSBs and DSBs. However, these laboratory-based methods require huge amounts of time, money, and high resources to identify the DBPs’ functions. − To overcome these shortcomings, computational techniques can provide an alternative solution to highlight as a powerful, reliable, efficient, and cost-effective method to predict SSBs and DSBs − as experimental techniques. Over the past decade, a number of researchers have made an effort to develop computational methods, including machine learning (ML) − and deep learning (DL) approaches, for predicting DBPs and their functional types. These predictors often work on inputs to sequence- and 3D structure-based methods. Applications of evolutionary sequence-based methods such as Wang et al., Sharma et al., Xiong et al., Xiong et al., Dey et al., Sharma et al., SDBP-Pred, and CNN-Pred only required amino acid protein sequences with various feature extraction techniques, including evolutionary information obtained from PSI-BLAST, and other methods use secondary structure, amino acid composition (AAC), quasi residue couple (QRC), and physicochemical properties − for better identification of SSBs. Other feature extraction methods are the pseudoposition-specific scoring matrix (PsePSSM), N-gram features (NGR), and texture descriptors (TD). Sang et al. used hidden Markov models (HMMs) and converted the obtained matrices into fixed-length feature vectors based on AAC, autocovariance transformation (ACT), and cross-covariance transformation (CCT). Thus, sequence-based computational methods are still required for the accurate and genome-wide prediction of SSBs. The feature extraction, classification algorithm, and selection of the best features are crucial to the accurate computational prediction of the SSBs. Recently, four accurate benchmarked predictors, Wang et al., proposed a prediction method using amino acid composition, dipeptide composition, and position-specific scoring matrices (PSSMs) as feature descriptors and with support vector machines (SVM) and random forests (RF) as the classification algorithms. Thereafter, SDBP-Pred proposed a feature extraction approach on the basis of PSSMs in aggregation with a K-segmentation method with the notion of a consensus sequence, prior to SVM-based classification of the obtained features’ performance. Likewise, Sharma et al. employed hidden Markov models (HMMs) together with PSSM profiles to construct feature representations for the same prediction task, using monogram and bigram schemes for feature generation, and then applied SVM, RF, and K-nearest neighbors (KNN) classifiers to predict SSBs and DSBs. More recently, CNN-Pred was introduced, which extracted evolutionary features through monogram and bigram profiles derived from PSSM and implemented 1D convolutional neural network (CNN) classifiers to identify SSBs and DSBs. Although the above traditional state-of-the-art predictors and the CNN-based predictor CNN-Pred have achieved substantial performance improvement, most of them still rely on either hand-crafted sequence descriptors or single-view evolutionary PLM-based embeddings, which limit their ability to capture discriminative global or local dependencies and to remain robust under severe class imbalance. To overcome these limitations, we propose BAN-SDBPred, a multimodal deep learning framework that robustly discriminates SSBs from DSBs using sequence information. BAN-SDBPred integrates complementary global and local representations by combining evolutionary-scale embedding (ESM2) with a residue energy contact matrix profile encoded via the histogram of oriented gradients (RECM-HOG), as visualized in Figure . These heterogeneous features are fused through a bilinear attention network (BAN), which facilitates local as fine-grained and global as long-range fusion among multiple modalities, which enables BAN-SDBPred to capture both global dependencies and local binding motifs. Furthermore, the ANBS strategy, along with focal loss, is explicitly applied to balance the classes between DSBs and SSBs, which improves minority-class detection while maintaining overall calibration and generalization. Experimental performance highlights that BAN-SDBPred attains new higher performance by outperforming four existing SSB methods on independent test data sets, as illustrated via Figure . Extensive ablation studies demonstrate that the strong performance is due to the fusion between ESM2 global embeddings and RECM-HOG local descriptors, with ANBS class balancing, making BAN-SDBPred a generalizable protocol for predicting the binding function of DBPs. BAN-SDBPred outperforms the existing PLM-based predictors in three aspects:By utilizing a bilinear attention network, global evolutionary embeddings (ESM2) are integrated with local structural features (RECM-HOG) to capture the long-range sequence dependencies and local residue interactions simultaneously.The bilinear attention fusion layer biologically magnifies the cross-modal interaction between conserved residues and structural motifs, enabling the layer to learn more interpretable feature insights to bind DNA–protein.By integrating ANBS to address the impact of extreme class imbalance, BAN-SDBPred consistently improves generalization and achieves higher performance over several multimodal protein predictors while retaining interpretability via SHAP analysis. This incorporation of established techniques in a synergistic framework signifies a conceptual advance beyond previous PLM-based SSB/DSB predictors, offering both an improved accuracy and biological interpretability. By utilizing a bilinear attention network, global evolutionary embeddings (ESM2) are integrated with local structural features (RECM-HOG) to capture the long-range sequence dependencies and local residue interactions simultaneously. By utilizing a bilinear attention network, global evolutionary embeddings (ESM2) are integrated with local structural features (RECM-HOG) to capture the long-range sequence dependencies and local residue interactions simultaneously. The bilinear attention fusion layer biologically magnifies the cross-modal interaction between conserved residues and structural motifs, enabling the layer to learn more interpretable feature insights to bind DNA–protein. The bilinear attention fusion layer biologically magnifies the cross-modal interaction between conserved residues and structural motifs, enabling the layer to learn more interpretable feature insights to bind DNA–protein. By integrating ANBS to address the impact of extreme class imbalance, BAN-SDBPred consistently improves generalization and achieves higher performance over several multimodal protein predictors while retaining interpretability via SHAP analysis. This incorporation of established techniques in a synergistic framework signifies a conceptual advance beyond previous PLM-based SSB/DSB predictors, offering both an improved accuracy and biological interpretability. By integrating ANBS to address the impact of extreme class imbalance, BAN-SDBPred consistently improves generalization and achieves higher performance over several multimodal protein predictors while retaining interpretability via SHAP analysis. This incorporation of established techniques in a synergistic framework signifies a conceptual advance beyond previous PLM-based SSB/DSB predictors, offering both an improved accuracy and biological interpretability. Illustrates framework BAN-SDBPred that includes a global evolutionary representation of the ESM2 encoder, a local descriptor representation of RECM, the HOG transformation on RECM, and the BAN with multimodal dual fusion, which features interaction and pooling layers. Materials and Methods Benchmark Data Sets The collection of a data set that is standardized and widely recognized plays a key part in the process of comparing, developing, evaluating, and benchmarking the effectiveness of models in predicting single-stranded binding (SSBs) and double-stranded DNA-binding proteins (DSBs). Standard benchmark data sets are used to assess the quality of research and are key to making reliable predictors. In order to be comparable with previous work, the data sets followed the recent standard in the literature. Two data sets were used: Uniprot1065, which was employed to train the model, and PDB401, which was considered an independent test set to assess the generalization capability of our method. The training data set was taken from UniProtKB/Swiss-Prot and contained, initially, 2,136 DSB and 339 SSB sequences. The CD-HIT was used to remove the sequence redundancy, with the sequence identity threshold of 0.7, to obtain the final training set with 873 of DSBs and 183 of SSBs. An independent test data set (PDB401) was obtained from the Protein Data Bank (https://www.rcsb.org/) by using PISCES (http://dunbrack.fccc.edu/Guoli/PISCES.php), a web server that selects NMR and X-ray crystal structures with resolution > 3 Å. Sequence similarity between training and test sets was limited to <30%, and minimum sequence length was 40 amino acids. The CD-HIT toolkit was used to identify the nonredundant protein sequences, thereby obtaining a final independent test set of 124 DSBs and 41 SSBs as represented with an imbalance ratio in Table . In this research, the DSBs are taken as the positive class and SSBs as the negative class while computing the evaluation measures. To overcome the class imbalance problem in the training data set, a neighborhood selection and adaptive synthesis-based sampling technique called ANBS was employed iteratively to augment the minority class. 1Benchmark Data Set Statistical Composition Data set Total Sequence (DSBs, SSBs) Ratio Train1,056(873, 183)∼4.77:1Test165(124, 41)∼3.02:1aRepresents the total number of DSBs and SSBs, respectively.bRepresents the imbalance ratio of DSBs:SSBs. Benchmark Data Set Statistical Composition Represents the total number of DSBs and SSBs, respectively. Represents the imbalance ratio of DSBs:SSBs. Architectural Overview of the Proposed Model BAN-SDBPred utilizes a dual-encoder-based design combining ESM2 evolutionary embeddings and RECM-HOG structural feature encoders. The RECM-HOG captures local interaction patterns, while ESM2 captures long-range global evolutionary dependencies. A bilinear attention network fuses these complementary modalities through high-order cross-modal interactions, and pooling is applied for a unified representation. To handle class imbalance, we applied ANBS during training. Finally, fully connected layers perform the binary classification of SSBs and DSBs. The graphical representation of the BAN-SDBPred framework is illustrated in Figure . Feature Encoding In the study, the multiview feature extraction method employed feature fusion on local handcrafted descriptors to extract semantic and global protein language model-based embeddings and evolutionary information, respectively, for the identification of DSBs and SSBs. Local Handcrafted Features In this study, two types of feature extraction techniques, the composition-based feature extraction method (CPSR) and RECM-HOG transformation, were employed on local handcrafted descriptors to extract semantics. CPRS is described in the Supporting Information: Note S1, Section (a). RECM-HOG Transformation Residue Energy Contact Matrix (RECM) Protein sequence stability quantification requires analysis of the interaction between amino acid residues at two levels, intra- and inter-residue contacts. Although energy-based potentials calculated from 3D structures reflect these interaction energies, they cannot be used for proteins without resolved structures, such as intrinsically disordered proteins or sequences without defined tertiary folds. To overcome this limitation, predictive methods have been proposed, which estimate pairwise amino acid interaction energies from primary sequences alone, without any structural constraints, but their performances for fully disordered proteins are suboptimal. In this work, we employ the Residue Energy Contact Matrix (RECM) to encode amino acid interaction propensities, utilizing predicted contact energy parameters, as shown in Table S2 of the Supporting Information: Note S3. The RECM parameters were derived through least-squares optimization, fitting pairwise residue interaction energies from 785 tertiary-structured proteins against 674 corresponding primary sequences. This square matrix shows the self-contact energies on its diagonal and the predicted contact energies between different standard amino acids in the other positions. For a target protein sequence of length L, RECM creates a feature representation that is 20 rows by L columns, changing the input sequence into a set of 20-dimensional vectors that show how different amino acids interact with each other. Formally, the RECM representation can be expressed in eq RECM=[E1,1E1,2⋯E1,20E2,1E2,2⋯E2,20⋮⋮⋮⋮EL,1EL,2⋯EL,20]L×20 1 where the matrix element E(k, L) denotes the energy contribution at position L for each of the 20 standard amino acid types, and k represents the residue position in the protein sequence. For a protein sequence of length L, the RECM matrix R∈RL×20 is generated as described in eq . To capture local residue interaction patterns, we calculate gradient magnitudes and orientations and then divide them into a 3 × 3 spatial grids for HOG transformation. Each cell contains a 9-bin orientation histogram, and concatenating all cells constitutes a fixed length 81-dimensions RECM-HOG feature vector. This method preserves local energy interaction patterns yet generates a uniform feature size for the downstream classifiers. Transforming RECM Using the HOG Descriptor The Histogram of Oriented Gradients (HOG) descriptor is a popular feature extraction technique in the fields of pattern recognition, image processing, and computer vision, especially for object detection. Empirical results show that HOG extracts more effective information from images than the traditional wavelet-based methods. , The HOG algorithm transforms R into a fixed-dimensional feature vector (RECM-HOG) through gradient analysis. First, horizontal gradient G x and vertical gradients G y at matrix position (i, j) were computed using discrete derivative filters represented in eqs and . Gx(i,j)={R(i+1,j)−0i=1R(i+1,j)−R(i−1,j)1