WheatRPred: deep learning framework for predicting disease resistance protein in wheat and its wild relatives Authors: Bharati Pandey, Manbir Singh, Awdhesh Kumar Mishra Published: 2026 ABSTRACT AbstractWheat, a globally vital staple crop, is under threats from a variety of pathogens, highlighting the urgent need for efficient identification of disease resistance (R) genes. Traditional experimental methods are slow and resource-intensive. To address this, we present WheatRPred, a web-based deep learning framework for accurate classification of R proteins in wheat and its wild relatives. Our integrative approach combines CNN, BiLSTM, and Transformer models which capture spatial, sequential, and contextual features from protein sequences. Trained on a curated dataset of validated R and non-R proteins, the models achieved 95–97% accuracy, outperforming traditional machine learning classifiers. WheatRPred allows users to submit protein sequences and obtain predictions classifying each sequence as resistance or non-resistance, along with confidence scores, thereby accelerating disease resistance research and genomic breeding in wheat. This platform provides a practical resource for researchers and breeders, supporting rapid R gene discovery, marker-assisted selection, and the development of disease-resistant wheat varieties.Keywords: Wheat, Deep learning, Web server, CNN, BiLSTM Introduction Plant pathogens can disrupt the host immune system, facilitating their growth and proliferation within plant tissues. Their spread poses a significant threat to global food security, causing substantial agricultural and economic losses (Strange & Scott, 2005). In contrast, most intracellular receptors, known as nucleotide-binding site leucine-rich repeat (NBS-LRR or NLR) proteins, detect pathogen-derived effectors inside the host cell and trigger effector-triggered immunity4, 1, 3, 2. Plants and pathogens communicate with each other during infection. The plant needs to detect and defend against pathogens, while the pathogen tries to change the plant’s biology to support its own growth. Both have evolved special genes for this interaction. Plant defense works at two main levels. The first level is pathogen-associated molecular patterns (PAMP) or microbe-associated molecular patterns (MAMPs) (PAMP-triggered immunity (PTI), where the plant recognizes common molecules from microbes (PAMPs/MAMPs) and activates general defense responses5. Pathogens, however, can block PTI using special proteins called effectors. The second level is effector-triggered immunity (ETI), where plants use resistance (R) genes to specifically recognize pathogen effectors, also called avirulence (Avr) proteins. This recognition triggers a stronger defense response, often known as gene-for-gene resistance. Both PTI and ETI cause major changes in plant gene expression, leading to the activation of defense-related genes. These include genes for antimicrobial proteins, enzymes that produce protective secondary metabolites, and proteins linked to hormone signaling. A key feature of ETI is the hypersensitive response (HR), a form of programmed cell death at the infection site that helps restrict pathogen spread6, 7, 8. Plant resistance (R) genes play a vital role in plant defense mechanisms by enabling the recognition of pathogen-derived effectors and activating downstream immune responses. To defend against these threats, plants have evolved a diverse array of disease resistance (R) proteins that detect invading pathogens and activate immune responses. Broadly, R proteins are classified into two groups: membrane-bound pattern recognition receptors (PRRs) and intracellular nucleotide-binding and leucine-rich repeat receptors (NLRs)4, 1, 3, 2. PRRs typically contain extracellular domains such as lysin motifs (LysM), leucine-rich repeats (LRRs), and lectin domains that recognize conserved microbial signatures9, 10. It includes, Receptor-like kinases (RLKs) and receptor-like proteins (RLPs) are cell-surface immune receptors involved in pathogen perception and defense signalling11. RLKs possess an extracellular recognition domain, a transmembrane region, and an intracellular kinase domain, whereas RLPs lack the kinase domain but function through interaction with signaling partners12. Among the diverse classes of R genes, nucleotide-binding site-leucine-rich repeat (NBS-LRR) genes represent the most prominent group and encode intracellular immune receptors that are central to plant innate immunity13. These proteins, commonly referred to as NLRs, are typically composed of three major regions: a variable N-terminal domain of approximately 200 amino acids, a conserved nucleotide-binding (NB-ARC) domain of about 300 amino acids involved in signal transduction14, 15, and a highly polymorphic C-terminal leucine-rich repeat (LRR) region consisting of 10–40 short motifs that mediate pathogen recognition (Jones and Jones, 1997). Based on variation in the N-terminal domain, NLRs are further classified into Toll/interleukin-1 receptor-type NLRs (TNLs), which contain a TIR domain; coiled-coil type NLRs (CNLs), which possess a coiled-coil or leucine zipper motif; and non-TIR NLRs that lack a TIR domain. Upon activation by pathogen-secreted effector molecules, NLR proteins trigger defense signaling pathways that often culminate in a hypersensitive response (HR) and localized programmed cell death, thereby restricting pathogen spread16, 17. To date, advances in plant immunity research have significantly enhanced our understanding of resistance mechanisms and enabled the molecular cloning of more than 450 resistance (R) genes from diverse plant species, with the majority reported in wheat (20.3%)18.Powdery mildew (PM), caused by the biotrophic fungus Blumeria graminis f. sp. tritici (Bgt), is one of the most destructive diseases affecting wheat worldwide. Although more than 60 officially designated powdery mildew resistance genes (Pm1-Pm68) have been identified and genetically mapped in wheat and its wild relatives, only 12 genes have been successfully cloned so far (McIntosh et al., 2020). To overcome the challenges associated with resistance gene cloning in wheat, several advanced sequencing-based approaches have been developed. Steuernagel et al.19 introduced MutRenSeq, which combines EMS mutagenesis with resistance gene enrichment sequencing, and successfully cloned the stem rust resistance genes Sr22 and Sr45 in common wheat. Since then, MutRenSeq has been widely used to clone additional resistance genes, including Yr5, Yr7, and YrSP, Pm1a, Sr26 and Sr61, Sr27, Lr13/Ne2m, and Sr920, 21, 22, 25, 24, 23. Another effective approach, MutChromSeq, was used to clone the powdery mildew resistance gene Pm2a (Sánchez-Martín et al., 2016). This method has since been applied to isolate several other resistance genes, including Pm4a, Lr14a, Sr43, and Yr87/Lr8528, 29, 27, 26. In addition, MutRNASeq, proposed by Yu et al.30, was successfully used to clone the stem rust resistance gene Sr62 and was later applied to isolate the powdery mildew resistance gene Pm6931. More recently, two closely related methods MutIsoSeq32and STAM (Sequencing Trait-Associated Mutations)33have been developed and applied to clone wheat disease resistance genes, including the leaf rust resistance genes Lr9/Lr58 and the stripe rust resistance gene YrNAM, respectively. Accurate identification of R genes is crucial for accelerating wheat improvement programs through molecular breeding and disease resistance research34. Although a large number of plant genomes and transcriptomes have been sequenced, discovering novel R genes particularly in landraces and wild relatives remains a challenging and labour-intensive process. One promising strategy for enhancing resistance durability involves stacking multiple R genes to pre-emptively counter evolving pathogen effectors. However, traditional approaches to R gene discovery, especially in complex genomes like wheat, are cumbersome and time-consuming. Several computational tools have been developed to address this need, categorized broadly into alignment-based and machine learning-based methods. Alignment-based pipelines, such as NLR-parser, RGAugury, and the method used by Montoya et al. (2020), rely on sequence similarity and conserved domain searches to identify candidate R proteins (Steuernagel et al., 2015). While effective in detecting well-conserved domains, these methods often struggle with low-similarity sequences and demand significant computational resources. Machine learning (ML) and Deep learning (DL)-based resistance gene prediction tool, including PRGminer35, NBSPred, RiceDRDLM, DRPPP, prPred, and prPred-DRLF and RLPredictiOme (Duhan et al., 2025;37, 36, 41, 38, 39, 40, have aimed to overcome some of these limitations by leveraging numerical representations of sequence features. Notably, NBSPred and DRPPP are no longer available as web servers, and prPred operates via command line, limiting accessibility. Furthermore, these tools depend on handcrafted features such as amino acid composition, dipeptide frequencies, and physicochemical properties, which require substantial domain expertise and may not capture complex sequence patterns effectively. Prediction of resistance proteins requires an understanding not only of individual amino acids but also of their contextual relationships within the sequence. These relationships describe how each residue is influenced by and interacts with other amino acids, both proximal and distal, across the protein chain42. Such dependencies are biologically significant because protein function is often governed by conserved motifs, structural domains, and long-range interactions that ultimately determine three-dimensional conformation and activity. By effectively modeling these sequence dependencies, deep learning approaches can uncover subtle patterns that distinguish resistance proteins from non-resistance proteins, even in cases where traditional alignment-based or handcrafted feature methods show limited performance. Different deep learning architectures contribute complementary strengths to sequence modeling. Convolutional Neural Networks (CNNs) capture local spatial motifs, Bidirectional Long Short-Term Memory networks (BiLSTMs) model bidirectional sequential dependencies, and attention mechanisms emphasize functionally important residues within the sequence43. Deep learning strategies, including concatenated feature extraction with modified LSTM architectures, have also demonstrated improved performance over conventional methods in related tasks such as crop disease classification44. Collectively, these approaches enable automatic extraction of high-level abstractions and complex features directly from raw sequence data, leading to improved generalization across bioinformatics applications. In wheat, reliable identification of R proteins is particularly important due to its complex hexaploid genome and its central role in global food security. Conventional experimental methods and existing computational pipelines are often labor-intensive, dependent on specialized expertise, or limited in accessibility. To overcome these challenges, we developed WheatRPred, a species-specific deep learning framework that integrates CNN, BiLSTM, and attention mechanisms to comprehensively capture spatial, sequential, and contextual sequence patterns. The proposed framework achieved superior predictive performance compared with traditional machine learning approaches and has been implemented as a user-friendly web server. Together, these advancements provide a valuable resource for plant scientists and breeders, facilitating R gene discovery and accelerating progress in disease resistance research, marker-assisted selection, and genomic breeding programs. Method Data collection Resistance protein sequences from Triticum aestivum, Triticum dicoccoides, Triticum monococcum subsp. monococcum, Triticum spelta, Triticum turgidum, Triticum turgidum subsp. dicoccoides, Triticum urartu, and Aegilops tauschii were downloaded in FASTA format from PRGdb 4.045 and NCBI, resulting in a total of 83,112 sequences that constituted the positive dataset (March 2025). Non-resistance (negative) protein sequences were retrieved from UniProt using the organism name[Triticum aestivum, Triticum dicoccoides, Triticum monococcum subsp. monococcum, Triticum spelta, Triticum turgidum, Triticum turgidum subsp. dicoccoides, Triticum urartu, and Aegilops tauschii],” while excluding sequences associated with kinases, plant defense, antiviral defense, host-virus interaction, resistance, and immunity-related functions (NOT Kinase [KW-0418]; NOT Serine/threonine-protein kinase [KW-0723]; NOT Tyrosine-protein kinase [KW-0829]; NOT Protein kinase inhibitor [KW-0649]; NOT Plant defense [KW-0611]; NOT Antiviral defense [KW-0051]; NOT Host-virus interaction [KW-0945]; NOT Nucleotide-binding [KW-0547]; NOT Leucine-rich repeat [KW-0433]; NOT Immunity [KW-0391]; NOT Innate immunity [KW-0399]; NOT Pathogenesis-related protein [KW-0568]). Only sequences with an annotation score ≥ 2 and a sequence length greater than 200 amino acids were retained, yielding 77,192 negative sequences from the same set of wheat and related species. To remove redundancy, all putative resistance gene sequences were clustered using CD-HIT46 at a 90% sequence identity threshold, reducing the positive dataset from 83,112 to 15,080 sequences. Furthermore, 91 experimentally validated R gene sequences from wheat and its related species were taken from studies and retrieved from NCBI48, 47; and incorporated into the positive dataset alongside the predicted R proteins. Similarly, the negative dataset was clustered from 77,192 to 19,036 sequences. To further validate the negative dataset, InterProScan (v5.67–99.0) was used to detect the presence of leucine-rich repeat (LRR) domains. Nine sequences containing LRR domains were identified and removed, resulting in 19,027 sequences. Finally, two FASTA files were prepared: one containing resistance protein sequences from wheat and related species (positive dataset: 15,171 sequences) and the other containing non-resistance plant protein sequences (negative dataset: 19,027 sequences). Reading the FASTA file The dataset comprised protein sequences stored in FASTA format, containing both positive (resistance genes) and negative (non-resistance genes) sequences. These sequences were read using BioPython's SeqIO module, which efficiently parsed the FASTA files and extracted the sequences as string representations. Each sequence was labeled accordingly; 1 for resistance genes and 0 for non-resistance genes and stored in a DataFrame for further processing. Encoding amino acid sequences Standard single-letter amino acid codes (A, C, D, E, F, G, H, I, K, L, M, N, P, Q, R, S, T, V, W, Y) were mapped to unique integer values ranging from 1 to 20. Unknown or non-standard amino acids were mapped to a special index value of 21 to ensure they are distinctly represented during model training. The maximum sequence length was determined from the dataset, and all sequences were encoded using the defined mapping. Sequences were truncated or padded with zeros to ensure uniform length across the dataset using TensorFlow’s pad_sequences utility. Each sequence was assigned a binary label: 1 for resistance genes and 0 for non-resistance genes. These labels were encoded into numeric format using scikit-learn's LabelEncoder to be used for supervised machine learning tasks. Train-test split with stratification Stratified sampling was used to ensure that the class distribution remains consistent across both sets, especially in classification tasks with imbalanced data. After balancing the dataset with an equal number of positive and negative samples, the data was split into a training set (80%) and a test set (20%) using stratified sampling. Stratification ensures that both the training and test sets contain approximately the same proportion of each class, thereby avoiding bias during training and ensuring a more reliable assessment of the model’s performance on unseen data. Handling class imbalance through random undersampling The imbalance between the positive and negative datasets is handled through random sampling. Initially, the indices of positive and negative sequences are separated based on their labels. Since the positive dataset is very large and the negative dataset is much smaller, the code randomly selects up to 10,000 positive and up to 10,000 negative samples, but if the number of available negatives is less than 10,000, it simply uses all available negative samples without replacement. After selection, the positive and negative indices are combined and shuffled to mix the two classes randomly. The corresponding sequences and labels are then extracted to form a balanced dataset. Finally, the dataset is split into training and testing sets with stratification, meaning the proportion of positive and negative samples remains consistent in both sets. Model development In this paper, we utilize three distinct methods for model development. Fig. 1 illustrates these approaches. Flow chart illustrating the development of the model for resistance protein classification. Hardware and software configuration All the models were developed using TensorFlow 2.0 along with the Keras high-level API, both of which are open-source software libraries extensively used for developing and deploying machine learning and deep learning models. The development environment was Jupyter Notebook, integrated within the Anaconda distribution. The entire experimentation was executed on a high-performance computational setup to ensure efficiency and scalability. The system was powered by an Intel Xeon® W7-2495X processor, a 24-core server-grade CPU designed for compute-intensive workloads. The processor was supported by 256 GB of DDR4 RAM, providing ample memory bandwidth and capacity to handle large datasets and model parameters. To significantly reduce training time and accelerate deep learning operations, an NVIDIA Quadro RTX A6000 GPU with a compute capability of 7.5 was utilized. This GPU is equipped with advanced tensor cores and RT cores, enabling enhanced support for AI workloads, particularly deep neural network training and inference. The CUDA (Compute Unified Device Architecture) parallel computing platform, along with the cuDNN (CUDA Deep Neural Network) library, was employed to leverage GPU acceleration. Convolutional neural network (CNN) architecture based model development Convolutional neural network (CNN) Architecture: Model I A lightweight one-dimensional convolutional neural network (1D-CNN) was designed to classify resistance and non-resistance protein sequences based on their encoded amino acid representation (Algorithm of this model is given in Supplementary Material 1). The architecture comprised the following key components: Input Layer: The input layer received padded amino acid sequences of length Tmax. Each amino acid was previously encoded into an integer as described earlier. Embedding Layer: An embedding layer converted the integer-encoded sequences into a. d-dimensional space (d=16), providing a compact representation of the input space. Convolutional Blocks: Four successive 1D convolutional layers (kernel size k = 3, ReLU activation, “same” padding), each followed by batch normalization and dropout (p = 0.3). •The first block consisted of 32 filters, followed by batch normalization and dropout (rate = 0.3). The first block consisted of 32 filters, followed by batch normalization and dropout (rate = 0.3). •The second block had 64 filters, batch normalization, and dropout. The second block had 64 filters, batch normalization, and dropout. •The third and fourth blocks each contained 128 filters, with batch normalization and dropout after each convolution. The third and fourth blocks each contained 128 filters, with batch normalization and dropout after each convolution. •All convolutional layers used a kernel size of 3 and the ReLU activation function with ‘same’ padding. All convolutional layers used a kernel size of 3 and the ReLU activation function with ‘same’ padding. Global Max Pooling: A GlobalMaxPooling1D layer was applied to capture the most salient features across the entire sequence. Fully Connected Layer: A dense layer with 128 units, ReLU activation, and L2 regularization (λ = 0.01) was used, followed by a dropout layer (rate = 0.5). Output Layer: The final output layer used a single neuron with a sigmoid activation function to predict the binary class label (resistant vs. non-resistant). Model Compilation and Optimization The model was compiled using the binary cross-entropy loss function and optimized using the Adam optimizer. Model performance was tracked using classification accuracy. To prevent overfitting and enhance model generalization, the following training strategies were used: Early Stopping: Training was halted if the validation loss did not improve for 10 consecutive epochs, and the best weights were restored. Model Checkpointing: The model state corresponding to the lowest validation loss was saved. Learning Rate Scheduling: The learning rate was automatically reduced when the validation loss plateaued using the ReduceLROnPlateau call-back. The learning rate was halved if no improvement was observed for three epochs, with a minimum threshold set to 1e-6. Equation (binary cross-entropy loss): Training and Evaluation The model was trained on the training dataset, with 20% of the samples randomly withheld as a validation set using the validation_split strategy implemented in Keras. Training was conducted over multiple epochs with a fixed batch size, and callback mechanisms were applied to monitor validation loss and control overfitting through early stopping. Model evaluation was performed on the held-out test set using accuracy and binary cross-entropy loss as performance metrics. Memory used by the trained model was explicitly cleared post-evaluation using TensorFlow and Python garbage collection utilities to ensure efficient resource management. CNN, BiLSTM, and attention Mechanism-Based model II Convolutional neural network (CNN) block The CNN block is designed to extract local patterns, such as conserved motifs, from the embedded sequences. It consists of three sequential Conv1D layers with increasing filter sizes 32, 64, and 128 each using a kernel size of 3. These layers are interleaved with batch normalization to stabilize learning and dropout to prevent overfitting. By stacking these convolutional layers, the model can detect hierarchical features from the sequence data. The final output of this block is passed through a global max pooling layer to retain only the most dominant features across the entire sequence length, reducing the dimensionality of the data while preserving important information (Algorithm of this model is given in Supplementary Material 2). Bidirectional LSTM Block. In parallel with the CNN block, the model includes a recurrent pathway built using Bidirectional LSTM (BiLSTM) layers. The first BiLSTM layer consists of 128 units and returns sequences so that another BiLSTM layer with 64 units can be applied. These layers allow the model to learn long-range dependencies in both forward and backward directions along the sequence, which is critical for biological data where amino acids influence each other in complex ways. Batch normalization is applied after each BiLSTM layer to improve training stability and convergence. The output of this block captures a rich temporal representation of the sequence. Forward hidden state and backward hidden state at time t Self-Attention Mechanism. To further enhance the sequential understanding, a custom self-attention mechanism is employed on the output of the BiLSTM block. This layer calculates an attention score for each position in the sequence using trainable weights and applies a softmax function to derive attention weights. These weights determine which positions in the sequence should be emphasized more. A weighted sum (context vector) is then computed using these attention weights, summarizing the most relevant features from the sequence. The resulting context vector is normalized using a layer normalization step, ensuring more stable and faster convergence during training. W is a trainable weight matrix, h_t) is the hidden state at time t, score_t is the attention score for position t. α_t is the attention weight at position t, context is the weighted sum of hidden states h_t . Feature Combination After extracting features from both the CNN and BiLSTM-attention branches, the outputs are merged. The global max pooled output from the CNN block and the attention-based context vector from the BiLSTM block are concatenated. This fusion brings together spatial features (from the CNN) and contextual, position-aware features (from the BiLSTM-attention mechanism). By combining both types of features, the model gains a holistic understanding of the sequence, which can significantly enhance its discriminative power in classification tasks. Dense Layers and Output. The concatenated feature vector is fed into a fully connected dense layer with 512 neurons and ReLU activation. This layer is regularized using L2 regularization to avoid overfitting and is followed by a dropout layer with a rate of 0.5 for additional regularization. Finally, the model concludes with a dense output layer consisting of a single neuron with sigmoid activation, which outputs the probability of the sequence being an R gene. This final probability is used to perform binary classification distinguishing resistance genes from non-resistance genes. Compilation and Training. The model is compiled using the binary crossentropy loss function, suitable for binary classification tasks. The Adam optimizer is used with a decaying learning rate to ensure gradual learning and to prevent overshooting minima during optimization. To prevent overfitting and improve generalization, several callbacks are used: early stopping halts training when validation performance stops improving, model checkpointing saves the best performing model, and a learning rate scheduler gradually reduces the learning rate after a set number of epochs. This training strategy ensures robust performance on unseen test data. CNN, BiLSTM, and attention Mechanism-Based model III This hybrid model is a simplified and more regularized version of the second model, the second model reduces the size of the LSTM layers from 128/64 units to 64/32 units, and the Dense layer before the output is made smaller, with 128 units instead of 512. Additionally, stronger L2 regularization (0.01 compared to 0.001) is applied to control overfitting more effectively. A key improvement is the replacement of the manual learning rate scheduling with the more adaptive ReduceLROnPlateau callback, which automatically lowers the learning rate when the validation loss stops improving. The optimizer settings are also kept simpler by directly using the Adam optimizer without manual decay parameters. These changes make the third model lighter, faster to train, and more robust, particularly beneficial when working with smaller or imbalanced biological sequence datasets. Model deployment on web The trained models were successfully deployed using Hugging Face, a widely used platform for hosting and sharing machine learning models. After training, all three models were saved in .h5 format. A repository was then created on Hugging Face to store the models, making them publicly accessible. To enable web-based interaction through WheatRPred, the models were integrated using the Hugging Face Inference API, allowing users to submit protein sequences via the web interface. The platform processes the input sequences, runs predictions using the selected model(s), and returns outputs classifying each sequence as a resistance or non-resistance protein. This setup ensures that users can access state-of-the-art deep learning predictions without needing to download or run models locally. Results Model development All three developed models were evaluated using precision, recall, and F1-score. In addition, the learning dynamics of each model were analyzed by examining the training and validation loss and accuracy curves throughout the training process. CNN-base model I We implemented a lightweight CNN model for binary classification of protein sequences. The model begins with an embedding layer that transforms input sequences into dense vectors, followed by four convolutional layers with progressively increasing filter sizes (32, 64, and 128). Each convolutional block is accompanied by batch normalization and dropout for regularization, which helps prevent overfitting. The model was compiled with binary cross-entropy loss and the Adam optimizer. The training and validation loss curves (Fig. 2A) show a sharp decline during the initial epochs, indicating rapid convergence and effective optimization of model parameters. After approximately 10–15 epochs, both curves gradually stabilize at a low loss value (∼0.18–0.20), suggesting that the model has successfully learned meaningful feature representations. Similarly, the model achieves accuracy above 95%, with both curves converging closely toward the end of training (Fig. 2B). The minimal gap between training and validation accuracy further supports that the model does not suffer from significant overfitting. Performance evaluation of the CNN-Base Model I. (A) Training and validation loss decrease steadily across epochs, indicating efficient learning and minimal overfitting. (B) Training and validation accuracy increase consistently and converge above 95%, demonstrating strong generalization and model stability. (C) The ROC curve (AUC = 0.985) reflects outstanding discriminative performance in distinguishing R proteins from non-R proteins. (D) The confusion matrix indicates balanced classification with very few false positives and false negatives, confirming prediction reliability. (E) The precision-recall curve remains near optimal across. Furthermore, the Receiver Operating Characteristic (ROC) curve (Fig. 2C) further highlights the excellent discriminative power of the model, with an Area Under the Curve (AUC) of 0.985. An AUC value close to 1.0 indicates outstanding classification capability, meaning the model can effectively distinguish between positive and negative classes across different threshold settings. The ROC curve remains near the top-left corner, reflecting high sensitivity and specificity. The confusion matrix (Fig. 2D) provides a detailed breakdown of prediction performance. The model correctly classified 1942 negative samples (True Negatives) and 1860 positive samples (True Positives), while misclassifying only 58 negatives as positives (False Positives) and 140 positives as negatives (False Negatives). These results demonstrate balanced predictive performance with low misclassification rates. The relatively small number of false predictions indicates strong reliability in both classes. Furthermore, the Precision-Recall (PR) curve (Fig. 2E) shows consistently high precision across a wide range of recall values, remaining near 1.0 for most thresholds. This indicates that the model maintains a low false-positive rate while achieving high sensitivity. Collectively, the integrated evaluation metrics confirm that the proposed CNN model achieves high accuracy, excellent discrimination ability (AUC = 0.985), balanced sensitivity and specificity, and minimal overfitting. These findings demonstrate that the model is robust, reliable, and well-suited for accurate classification tasks in biological sequence prediction. CNN-BiLSTM-Attention based model II This model combines convolutional, recurrent, and self-attention mechanisms for sequence classification, specifically aimed at predicting binary outcomes. The training and validation loss curves (Fig. 3A) show a rapid decline during the initial epochs, indicating efficient optimization and fast convergence of the model. Both curves gradually stabilize at a very low loss value (∼0.08–0.10), reflecting strong feature learning and minimal residual error. The accuracy curves (Fig. 3B) reveal a sharp increase in both training and validation accuracy during the early epochs, followed by gradual stabilization above 97%. The training accuracy slightly exceeds validation accuracy toward the final epochs, but the difference remains minimal, confirming stable learning dynamics. The ROC curve (Fig. 3C) highlights outstanding discriminative performance, achieving an AUC of 0.995. An AUC value this close to 1.0 signifies near-perfect classification capability. The curve remains tightly aligned with the top-left corner, demonstrating extremely high true positive rates at very low false positive rates across multiple thresholds. This indicates good sensitivity and specificity of the model. Comprehensive performance evaluation of the proposed Model II. (A) Training and validation loss curves showing rapid decline and stable convergence, indicating effective learning with minimal overfitting. (B) Training and validation accuracy curves demonstrating consistent improvement and high final accuracy (>97%). (C) ROC curve with an AUC of 0.995, reflecting excellent discriminative capability. (D) Confusion matrix showing high true positive (1950) and true negative (1967) predictions with very few misclassifications. (E) Precision-Recall curve illustrating near-perfect precision across most recall values, confirming robust and reliable classification performance. The confusion matrix (Fig. 3D) showed that the model correctly classified 1967 negative samples (true negatives) and 1950 positive samples (true positives), while only 33 negatives were misclassified as positives (false positives) and 50 positives were misclassified as negatives (false negatives). These results correspond to an overall accuracy of approximately 97.9%. The precision is extremely high due to the very low number of false positives, while recall (sensitivity) is also strong given the small number of false negatives. The specificity is similarly excellent, demonstrating balanced performance across both classes. The Precision-Recall curve (Fig. 3E) further confirms the robustness of the classifier. Precision remains nearly 1.0 across most recall values, with only a slight drop when recall approaches its maximum. This behavior indicates that the model maintains high reliability in positive predictions while capturing the majority of true positive instances. The strong PR performance is particularly important in biological sequence classification tasks, where minimizing false discoveries is critical. The high performance can be attributed to the hybrid model architecture, which integrates convolutional layers, Bidirectional LSTMs, and a Self-Attention mechanism. The convolutional layers effectively capture local sequence motifs, while the Bidirectional LSTM layers model long-range sequential dependencies in both directions, allowing the model to understand broader contextual relationships. The Self-Attention mechanism further enhances this capability by dynamically focusing on the most informative parts of the sequence, thereby improving the model’s ability to identify critical patterns and reduce the influence of irrelevant or noisy features. Additionally, learning rate scheduling and training callbacks, including early stopping and model checkpointing, optimize training efficiency, prevent overfitting, and ensure that the most performant model is selected. CNN-BiLSTM-Attention based model III In Model III, the training and validation loss curves show a rapid decrease during the initial epochs, indicating efficient parameter optimization and effective feature learning. After approximately 10–15 epochs, both curves gradually stabilize at a low loss value (∼0.10–0.12), reflecting convergence toward an optimal solution (Fig. 4A). The training and validation accuracy curves, which show a sharp increase in the early epochs followed by steady improvement and stabilization above 97%. The validation accuracy closely follows the training accuracy, with only a marginal difference between them, further confirming the robustness and stability of the model (Fig. 4B). The ROC curve highlights the excellent discriminative ability of the classifier, achieving an AUC of approximately 0.99 (Fig. 4C). The curve remains very close to the top-left corner of the plot, demonstrating a high true positive rate even at very low false positive rates. This confirms that the model effectively distinguishes between positive and negative classes across different classification thresholds. Performance evaluation of the proposed Model III. (A) Training and validation loss curves showing rapid convergence and minimal overfitting across epochs. (B) Training and validation accuracy curves demonstrating stable improvement and high final accuracy (>97%). (C) ROC curve with an AUC of ∼ 0.99, indicating excellent discriminative ability. (D) Confusion matrix summarizing classification performance with high true positives and true negatives and minimal misclassifications. (E) Precision-Recall curve illustrating consistently high precision across a wide range of recall values, confirming robust predictive performance. The confusion matrix provides detailed insight into class-wise performance. The model correctly classified 1961 negative samples (true negatives) and 1928 positive samples (true positives), while misclassifying only 39 negatives as positives (false positives) and 72 positives as negatives (false negatives). These results correspond to an overall accuracy of approximately 97.2%. The precision is very high due to the low number of false positives, while recall (sensitivity) remains strong despite the slightly higher number of false negatives. The specificity is also excellent, indicating balanced and reliable performance across both classes (Fig. 4D). The Precision-Recall curve further supports the robustness of the model. Precision remains close to 1.0 across a broad range of recall values, declining slightly only when recall approaches its maximum (Fig. 4E). Building upon these results, Model III represents a streamlined and more regularized version of the original architecture. In this enhanced design, the recurrent and dense layers are reduced in size to lower the parameter count, while stronger regularization is applied to further control overfitting. Manual learning-rate scheduling is replaced with an adaptive callback that automatically adjusts learning rates when the validation loss plateaus, and the optimizer settings are simplified to improve training stability and reproducibility. Collectively, these changes make Model III lighter, faster, and more robust particularly beneficial for smaller or imbalanced biological sequence datasets. By balancing complexity and regularization, Model III achieves smoother convergence, improved generalization, and more stable predictions. As a result, the model becomes computationally efficient, reliably captures essential sequence patterns, and is better suited for real-world bioinformatics applications where consistency, scalability, and resource efficiency are critical. Web implementation of the developed model To facilitate reproducibility and ensure broad accessibility, the three deep learning models developed in this study have been deployed on the Hugging Face Model Hub. The models include: (i) a convolutional neural network (CNN) designed specifically for deployment in resource-constrained environments (Low) (ii) a lightweight variant of the LSTM model optimized for performance-efficiency (Balanced), and (iii) a high-capacity self-attention-augmented bidirectional LSTM (High GPU). These models are made available for direct download and inference through the Hugging Face ecosystem, supporting standard interfaces such as Transformers, Keras, and TensorFlow. A comprehensive requirements.txt file is provided to replicate the development environment. To enhance user accessibility, we have developed an interactive web interface where users can input protein sequences and select the appropriate model from a dropdown menu. This interface provides a user-friendly environment for making predictions without requiring a local setup. The web interface is accessible via the following link: https://huggingface.co/spaces/BharatiPandey/DLPSM (Fig. 5). All models, along with detailed documentation and usage instructions, are publicly available for research and application purposes. Currently, the web server accepts 10 protein sequence in fasta format at a time. The output provides a general prediction of the protein as either resistant (positive) or non-resistant (negative). Submission form of the Wheat resistance protein prediction server and its output. Discussion Over the past decade, several computational tools such as DRAGO3(Calle García et al., 2022), PRGminer (Duhan & Kaundal, 2025), RiceDRDLM36, ResCap49, prPred38, PrPred-DRLF39, and RLPredictiOme40have been developed for resistance protein prediction. Although these approaches have advanced the field, many rely on domain-based detection, handcrafted features, or conventional machine-learning algorithms, which may limit their predictive robustness and generalization across diverse datasets. In the present study, we developed three progressively enhanced deep learning architectures integrating convolutional, recurrent, and attention mechanisms to capture both local sequence motifs and long-range dependencies. As shown in Table 1, the CNN-based Model I achieved a precision of 96.9%, recall of 93.0%, F1-score of 94.9%, test accuracy of 95.0%, validation accuracy of 94.7%, and an AUC of 98.5%. Building upon this framework, the hybrid CNN-BiLSTM-Attention Model II significantly improved performance, achieving precision of 98.3%, recall of 97.5%, F1-score of 97.9%, test accuracy of 97.93%, validation accuracy of 97.41%, and the highest AUC of 99.5%. Similarly, Model III achieved precision of 98.0%, recall of 96.4%, F1-score of 97.2%, test accuracy of 97.22%, validation accuracy of 96.50%, and an AUC of 99.0%. These results clearly demonstrate the advantage of integrating CNNs for motif extraction, BiLSTMs for contextual sequence modeling, and attention mechanisms for capturing global feature relationships. Table 1Comparative statistical score between developed model and previous models.ModelsSoftwarePrecisionRecallF1-scoreTestAccuracyValidationAccuracyAUCCNN-based Model ICNN96.993.094.995.094.798.5CNN-BiLSTM-Attention Based Model IICNN-BiLSTM98.397.597.997.9397.4199.5CNN-BiLSTM-Attention Based Model IIICNN-BiLSTM98.096.497.297.2296.5099.0DRAGO3Domain detection using HMM1007585−82−PRGminerCNN99.0298.650.9998.7595.72∼95–98RiceDRDLMMulti-layer Perceptr (MLP)95.6592.593.7791.59−97.84ResCapSupport vector machine−−−76.64−−prPredSupport vector machine10080.689.393.5−94.8PrPred-DRLFBiLSTM, UniRep, and TAPE-BERT−−93.395.60.926 to 0.95999.7RLPredictiOmeMachine Learning−−−79−−*Bold one are model developed in the study. Comparative statistical score between developed model and previous models. *Bold one are model developed in the study. When compared with previously reported tools, the superiority of the proposed models becomes evident (Table 1). DRAGO3, which is based on HMM domain detection, reported 100% precision but only 75% recall and an F1-score of 85%, indicating limited sensitivity. RiceDRDLM36), using a multilayer perceptron architecture, achieved 91.59% accuracy and a 93.77% F1-score, both lower than those of the proposed hybrid models. Similarly, prPred38, based on support vector machine classification, achieved 93.5% accuracy with 80.6% recall, whereas Model II substantially improves recall to 97.5% while maintaining high precision (98.3%). PrPred-DRLF39, which integrates BiLSTM, UniRep, and TAPE-BERT embeddings with multiple classifiers, reported an F1-score of 93.3% and validation accuracy ranging from 0.926 to 0.959; in contrast, Model II achieves higher validation accuracy (97.41%) and comparable or superior AUC (99.5%). RLPredictiOme40, based on Bayesian inference and KNN classification, reported only 79% test accuracy, which is considerably lower than the performance of all three proposed models. ResCap, using SVM, achieved 76.64% accuracy, further highlighting the advantage of deep learning-based architectures. PRGminer (Duhan & Kaundal, 2025), which relies on dipeptide composition features and a CNN architecture, reported 98.75% test accuracy; however, Model II demonstrates competitive performance with higher AUC and stronger validation stability. Collectively, these comparisons confirm that deep learning architectures specifically tailored to biological sequence data significantly enhance resistance protein prediction compared with traditional machine-learning and domain-based methods. Among the developed models, Model II consistently achieves the best balance of precision, recall, F1-score, validation accuracy, and AUC, indicating superior robustness and generalization capability. To further assess biological plausibility, a qualitative validation was performed. From the test dataset, sequences labeled as positive (R proteins) and negative (non-R proteins) were identified, and five sequences were randomly selected from each group using a fixed random seed to ensure reproducibility. After selection, model-wise prediction was performed using the three trained deep learning models. For each model, the predicted class (R protein or non-R protein) and the corresponding probability score were obtained (Table S1). In total, 12 sequences were used for validation. The predicted sequences and their associated conserved domains were further cross-validated using the DRAGO3 server to assess consistency with an independent R-gene prediction platform. All three models produced identical classification outcomes for the selected sequences with comparable probability scores, indicating strong agreement among different model architectures. Moreover, the predictions were largely consistent with DRAGO3 results and with the presence of resistance-related conserved domains, supporting the biological plausibility of the computational predictions. This qualitative validation demonstrates the robustness and reliability of the proposed framework for R-gene identification. Although WheatRPred demonstrates high accuracy and robust generalization on wheat and related species, its current training and validation are primarily based on wheat R-gene datasets. Therefore, its direct applicability to other crops, particularly non-wheat monocots and dicots, requires careful consideration. Monocot genomes, including wheat, are predominantly enriched in CC-NBS-LRR (CNL) class R genes, whereas dicot species such as Arabidopsis thaliana contain both CNL and TIR-NBS-LRR (TNL) types50, 51, with TNLs being largely absent in most monocots52. These lineage-specific structural features may influence the sequence patterns learned by deep learning models trained on wheat-derived data.To achieve broader applicability across plant lineages, retraining or fine-tuning the model using representative R-gene datasets from diverse monocot and dicot species will be necessary. Future work will therefore focus on systematic evaluation of WheatRPred on additional crops and on developing lineage-adaptive models to improve its utility in comparative genomics and crop improvement research. Conclusion The developed deep learning framework for classifying resistance (R) genes and non-resistance genes demonstrated exceptional performance across multiple evaluation metrics, confirming its robustness and applicability to real-world bioinformatics analyses. The steady decline in both training and validation loss curves, along with the convergence of accuracy values toward near-perfect levels, indicates efficient learning and strong generalization capability without evident overfitting. This stability can be attributed to the incorporation of regularization strategies such as Dropout, Batch Normalization, and L2 regularization. The ROC curve, with an AUC of 0.99, highlights the model’s outstanding discriminative power in distinguishing R genes from non-R genes. Additionally, the confusion matrix revealed high sensitivity and specificity, demonstrating the model’s effectiveness in correctly identifying true positives while minimizing false positives. The precision-recall curve further supports its reliability, maintaining consistently high precision across varying recall thresholds, an essential feature for imbalanced biological datasets where false negatives can be particularly detrimental. Collectively, these results establish the model as a highly accurate, stable, and dependable tool for resistance gene prediction and related bioinformatics applications. Limitations Random undersampling was applied to balance the dataset, which may have resulted in the loss of potentially informative sequences and could affect model generalization to real-world genomic data. Additionally, redundancy reduction using CD-HIT at a 90% sequence identity threshold may retain relatively similar sequences; thus, future studies should evaluate performance at more stringent thresholds (e.g., 70% or 60%) to better assess generalization. The use of a filtered negative dataset, excluding defense-related and R-like proteins, may create a less challenging classification task and potentially lead to an overestimation of model performance. Future work will incorporate more diverse and challenging negative samples to better reflect real-world conditions. The hybrid CNN-BiLSTM-attention architecture requires substantial computational resources, which may limit accessibility. Although the attention mechanism highlights informative sequence regions, explicit biological interpretation (e.g., motif- or domain-level attribution) was not performed. Future studies will integrate explainable AI approaches to improve interpretability. Finally, WheatRPred currently functions as a binary classifier and does not provide subtype-level classification or explicit domain- or structure-based annotations. Extending the framework to support R gene subtype classification (e.g., NBS-LRR, RLK, and RLP) and integrating domain or structural features (e.g., Pfam or InterPro) represent important directions for future development. CRediT authorship contribution statement Bharati Pandey: Writing – review & editing, Writing – original draft, Visualization, Validation, Supervision, Software, Resources, Project administration, Methodology, Formal analysis, Data curation, Conceptualization. Manbir Singh: Writing – review & editing, Writing – original draft, Software, Methodology. Awdhesh Kumar Mishra: Writing – review & editing, Writing – original draft. Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgements Authors are thankful to the Anusandhan National Research Foundation (ANRF) (Erstwhile Science and Engineering Research Board (SERB), Govt. of India under the grant SRG/2022/001472. Footnotes Supplementary data to this article can be found online at https://doi.org/10.1016/j.jgeb.2026.100700. Supplementary data The following are the Supplementary data to this article: Supplementary Data 1 References Associated Data Supplementary Materials Supplementary Data 1