SaMCL: a multi-task collaborative learning framework for peptide-protein interaction prediction based on structure-aware protein language models - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Brief Bioinform . 2026 Apr 19;27(2):bbag178. doi: 10.1093/bib/bbag178 Search in PMC Search in PubMed View in NLM Catalog Add to search SaMCL: a multi-task collaborative learning framework for peptide-protein interaction prediction based on structure-aware protein language models Siyi He Siyi He 1 School of Artificial Intelligence, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China 2 National Engineering Laboratory for Big Data System Computing Technology, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China Find articles by Siyi He 1, 2 , Dongzhen Tang Dongzhen Tang 3 School of Computer Science, Guangdong University of Technology, 100 Waihuan West Road, 510006 Guangzhou, China Find articles by Dongzhen Tang 3 , Tiantian Zhu Tiantian Zhu 4 School of Artificial Intelligence, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China 5 National Engineering Laboratory for Big Data System Computing Technology, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China Find articles by Tiantian Zhu 4, 5 , Zexuan Zhu Zexuan Zhu 6 School of Artificial Intelligence, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China 7 National Engineering Laboratory for Big Data System Computing Technology, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China Find articles by Zexuan Zhu 6, 7 , Yumeng Liu Yumeng Liu 8 School of Artificial Intelligence, Shenzhen Technology University, 3002 Lantian Road, 518118 Shenzhen, China Find articles by Yumeng Liu 8, ✉ , Jun Zhang Jun Zhang 9 School of Artificial Intelligence, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China 10 National Engineering Laboratory for Big Data System Computing Technology, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China Find articles by Jun Zhang 9, 10, ✉ Author information Article notes Copyright and License information 1 School of Artificial Intelligence, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China 2 National Engineering Laboratory for Big Data System Computing Technology, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China 3 School of Computer Science, Guangdong University of Technology, 100 Waihuan West Road, 510006 Guangzhou, China 4 School of Artificial Intelligence, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China 5 National Engineering Laboratory for Big Data System Computing Technology, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China 6 School of Artificial Intelligence, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China 7 National Engineering Laboratory for Big Data System Computing Technology, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China 8 School of Artificial Intelligence, Shenzhen Technology University, 3002 Lantian Road, 518118 Shenzhen, China 9 School of Artificial Intelligence, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China 10 National Engineering Laboratory for Big Data System Computing Technology, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China ✉ Corresponding authors. Yumeng Liu, E-mail: [email protected] ; Jun Zhang, E-mail: [email protected] Received 2025 Nov 28; Revised 2026 Feb 26; Accepted 2026 Mar 20; Collection date 2026 Mar. © The Author(s) 2026. Published by Oxford University Press. This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial License ( https://creativecommons.org/licenses/by-nc/4.0/ ), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited. For commercial re-use, please contact [email protected] for reprints and translation rights for reprints. All other permissions can be obtained through our RightsLink service via the Permissions link on the article page on our site—for further information please contact [email protected]. PMC Copyright notice PMCID: PMC13092273 PMID: 42001471 Abstract The peptide drug points to a promising new therapeutic. Precisely predicting the interaction between peptides and proteins is fundamental to the discovery and design of functional peptides. While various computational methods have been proposed for this purpose, constructing an accurate and robust prediction model remains a challenge. In this study, we introduce a structure-aware multi-task collaborative learning (SaMCL) framework for detecting the interaction between peptides and proteins. To the best of our knowledge, SaMCL is the first method capable of performing a multilevel, simultaneous prediction of binary interactions and binding domains in both peptides and proteins. Experimental results demonstrate that SaMCL outperforms several state-of-the-art methods in terms of both prediction accuracy and generalization. Provides a new paradigm for modeling biomolecular interactions. Keywords: protein–peptide interaction, structure-aware protein language model, collaborative learning framework Introduction Biomolecular interactions play a key role in biological processes such as cellular signaling, protein transport, programmed cell death, and gene expression regulation [ 1 ]. Identifying the interactions between peptides and proteins (PepPIs) enables the discovery and design of functional peptides targeting specific proteins, thus facilitating the development of peptide-based therapeutics [ 2 ]. Conventional wet-lab methods for detecting PepPIs are typically time-consuming, technically demanding, and costly, making them impractical for large-scale investigations [ 3 , 4 ]. In recent years, the rapid development of artificial intelligence (AI) has significantly accelerated the computational analysis of PepPIs, leading to numerous AI-based approaches [ 5 , 6 ]. For example, CAMP [ 7 ] employs the convolutional neural network (CNN) [ 8 ] to extract features from raw sequences to predict the interaction between peptides and proteins and the binding sites in the peptides. DrugBAN [ 9 ] introduces a bilinear attention network to capture detailed molecular correspondences and incorporates domain-adversarial learning to enhance generalization in diverse datasets. Recent methods such as DeepPepPI and IIDL-PepPI extend PepPI modeling to domain adaptation and multitask learning. DeepPepPI [ 10 ] integrates contextual and cross-dependency modules to improve prediction in plant-specific systems, while IIDL-PepPI [ 11 ] leverages bidirectional attention and progressive transfer learning to jointly model interactions and binding residues. Some methods focus on predicting peptide-binding domains in proteins. For example, PepBind [ 12 ] captures local spatial patterns on protein surfaces using CNNs to identify peptide-binding sites in proteins. PepNN [ 13 ] extends this approach by incorporating transformer-based contextual embeddings and transfer learning, achieving more accurate predictions even with limited data. PepBCL [ 14 ] employs a bi-channel framework to jointly encode peptide and protein features via cross-modal attention, effectively modeling interdependencies between binding partners. PepCA [ 15 ] further incorporates cross-attention to align peptide and protein residue representations and to capture fine-grained interaction contexts, thereby improving performance. Despite these advances, improving the accuracy and generalization of predictions for partner-specific binary interactions and binding sites remains a challenge. Moreover, AI-based models typically rely on large-scale labeled data as the primary driver of training, an unrealistic requirement for protein–peptide interactions, where experimentally verified datasets remain extremely scarce. Pretrained language models provide an effective solution to the problem of learning from limited samples. Protein language models (PLMs), inspired by the intrinsic similarity between natural language and biological sequences, where words are arranged to form meaningful sentences and amino acids are ordered to create proteins with specific structures and functions [ 16 ], are designed to capture rich contextual representations of protein sequences. By pretraining on large-scale protein databases, PLMs such as ESM [ 17 ] have demonstrated strong generalization in various protein-related tasks and have been adapted for the peptide–protein interaction [ 18–20 ]. However, current approaches still rely solely on sequence-based PLMs [ 11 ]. The interactions are fundamentally governed by 3D structural complementarity. Integrating structural information may further enhance prediction performance. Recent breakthroughs in protein structure prediction, exemplified by AlphaFold [ 21–23 ], have catalyzed the development of structure-aware PLMs. For example, SaProt [ 24 ] transforms 3D structural data into 1D structural sequences (3Di-tokens) via Foldseek [ 25 ] and constructs a structure-aware vocabulary by combining them with amino acid sequences through a Cartesian product, achieving remarkable results in various downstream tasks. Similarly, ProSST [ 26 ] models the protein language with quantized structure and dual-path architecture, achieving superior performance in zero-shot mutation prediction, thermostability assessment, and metal–ion binding analysis. Integrating structural PLM into the prediction of peptide–protein interaction is expected to improve performance. However, these PLMs are general-purpose protein models and are not specifically optimized for peptide–protein interactions. How to adapt them to this task effectively remains to be further explored. To address these challenges and improve prediction accuracy, we propose structure-aware multi-task collaborative learning (SaMCL), a structure-aware multi-task collaborative learning framework [ 27 ] that unifies interaction prediction and binding site identification for both peptides and proteins. The main innovations of SaMCL are as follows: Structure-aware protein representation. To transcend the limitations of sequence-only modeling, SaMCL integrates the structure-aware PLM SaProt. Furthermore, to adapt SaProt for interaction prediction, we perform task-specific fine-tuning [ 28 ], allowing the model to retain its structural representation power while enhancing its sensitivity to interaction pattern learning. This structural prior, combined with fine-tuning, enables SaProt to effectively capture spatial cues for both proteins and peptides, providing a biologically meaningful foundation for interaction and binding site prediction. Hierarchical cross-modal feature fusion. SaMCL adopts a dual-stream architecture with dedicated protein and peptide encoders. A cross-attention module [ 29 ] is then introduced to dynamically align residue-level contextual embeddings between the two molecular entities, achieving deep bidirectional information exchange. This surpasses naive concatenation by explicitly modeling interaction interfaces and resolving spatial dependency mismatches. Unified multi-scale prediction. The framework jointly performs sequence-level binary interaction prediction and residue-level binding site identification. Meanwhile, SaMCL employs an uncertainty-weighted loss-balancing mechanism to adaptively adjust each task’s contribution during the optimization process, ensuring stable, balanced multi-task learning. Collectively, SaMCL achieves both high accuracy and robust performance in protein–peptide interaction prediction and binding-domain identification tasks. By leveraging structure-aware PLM embeddings and a multi-task collaborative learning framework, the model effectively integrates local and global sequence-structure information from both peptide and protein modules to form robust feature representations. Its modular design progressively refines discriminative features across layers and aligns cross-modal representations at higher levels, resulting in superior performance in both interaction prediction and binding site identification. Methods Datasets construction To comprehensively and reliably evaluate the predictive performance of peptide–protein interactions, we constructed a specialized dataset that simultaneously integrates molecular-level and residue-level binding information. The overall workflow is illustrated in Fig. 1 . First, we retrieved all protein complexes from the RCSB PDB [ 30 , 31 ] database up to 2025. Subsequently, we employed PLIP [ 32 ] to identify non-covalent peptide–protein interactions within these complexes (defined as heavy-atom distances Å between peptide segments and protein residues), retaining only structures containing such interactions. To meet SaProt’s requirement for structure-aware inputs, we used Foldseek [ 33 ] to extract structural tokens from the complexes. Because Foldseek skips any peptide segment shorter than four residues when serializing structure, we retained only peptides with 4 amino acids. This filtered only of all sequences in the raw data ( Table S1 in the Supplementary Material ). Next, we parsed the complexes with PDB-BRE [ 34 ] to locate the exact binding sites and assigned an interaction label of 1 to every documented contact. The resulting records, including raw amino-acid sequences, interaction flags, and residue-level binding labels, were merged into a single dataset that provides instant joint access to sequence, structure, and annotation information. Figure 1. Open in a new tab Dataset construction and model architecture. (a) The construction pipeline of the peptide–protein interaction dataset is illustrated. The final, balanced collection unifies positive and negative samples and is equipped with consistent sequence, structure, and interaction annotations. (b) The SaMCL architecture is depicted, encompassing molecular representation, feature learning, cross-attention fusion, and multi-task prediction modules. The prediction heads output interaction probabilities for protein and peptide pairs as well as binding-site annotations, with key segments indicating the predicted binding residues. Since the initial dataset exclusively comprises positive interaction samples, we generated synthetic negative samples to construct a balanced training set for the binary classification task. The methodology for negative sample construction aligns with the dataset partitioning strategies, which are categorized into Random Re-Pairing and Cross-cluster Pairing. The former is used to evaluate the model’s baseline performance, while the latter assesses its generalization and robustness under lower sequence similarity, using CD-HIT [ 35 ]. To ensure data integrity, we enforced strict non-overlap at the protein–peptide pair level, ensuring no negative pair matches any positive sample in the dataset. Detailed procedures for generating negative samples are shown in Fig. S1 in the Supplementary Material . For all negative samples, the interaction labels were set to 0, and the corresponding binding-site labels were initialized to zeros, since binding-site prediction is performed only when an interaction is present. Statistical information on the datasets used in this study is provided in Fig. S2 and Tables S2–S4 in the Supplementary Material . General framework We propose SaMCL, a multi-task collaborative learning framework for simultaneously predicting protein–peptide interactions and their binding sites. The general framework of the model is illustrated in Fig. 1 . SaMCL uses SaProt, which was fine-tuned using the low-rank adaptation technique (LoRA) [ 36 ] for large language models, to extract 3D structural features from protein–peptide complexes. These are combined with empirical features to form a comprehensive multi-modal representation. Feature learning is performed through two parallel pathways: a multi-scale convolutional neural network (MSCNN) [ 37 ] that captures local dependencies, and a Bidirectional Long Short-Term Memory module [ 38 ] with an Attention mechanism [ 39 ] (BiLSTM-Attention) that models long-range context. The resulting features are fused into unified and context-aware representations. Then, a Cross-Attention module [ 29 ] enables the fusion of protein and peptide features, capturing interaction-specific dependencies. SaMCL includes three tasks: binary prediction for peptide–protein interaction, annotation of protein-binding sites in peptides, and identification of peptide-binding sites in proteins. We used a task uncertainty weighting mechanism [ 40 ] to balance losses across tasks, ensuring stable and efficient multi-task optimization [ 41 ]. Molecular representation SaMCL adopts multimodal representations that integrate sequence information, physicochemical properties, and structural patterns [ 42 , 43 ]. All features are linearly projected into a unified latent representation to ensure consistent feature distributions across modalities. One-hot encoding of the sequence: Represents 20 standard amino acids and a special token X (for unknown or non-standard amino acids). Physicochemical properties: Reflecting amino acid polarity and hydrophilicity combinations. Structure-aware embeddings: Derived from the pretrained SaProt model, integrating 3D structural information into sequence representations through structure-aware tokenization to form a residue-level semantic feature matrix. SaProt was primarily designed for general protein-related tasks and has not been explicitly optimized for interactions. To better align SaProt with the task distribution in this study, we conducted a dedicated pretraining stage in which SaProt, together with the multi-task prediction heads, was fine-tuned using peptides and proteins from our dataset via the LoRA method. During this stage, the remaining components of SaMCL were not involved in training. Specifically, the LoRA modules were inserted into SaProt’s attention projection layers while the original backbone parameters remained frozen. This parameter-efficient fine-tuning strategy enables effective task adaptation by applying low-rank decomposition to selected weight matrices within the pretrained model, such as the attention and feedforward layers, thereby reducing computational overhead while preserving the general knowledge of the backbone model. Feature learning To capture both local patterns and long-range dependencies, SaMCL employs two parallel feature learning branches for proteins and peptides. Local feature learning branch : The local branch uses an MSCNN module that applies convolutional kernels of varying sizes to sequential representations. MSCNN enables the extraction of short-range dependency features across multiple receptive fields, thereby enhancing the recognition of fine-grained local patterns. For a set of convolutional kernels , the output is the following: (1) (2) Global feature learning branch : The global branch combines a BiLSTM with a self-attention mechanism [ 44 ] to capture global semantics and long-range dependencies. BiLSTM encodes information from both directions of the feature sequence, enabling residue embeddings to incorporate preceding and subsequent context. The self-attention layer further strengthens inter-residue dependencies, reflecting the holistic conformation and potential functional relationships between proteins and peptides. (3) (4) where D is the dimension of the input embedding, and h is the latent dimension; , and represent the projection matrices for query, key, and value, respectively, and denotes the scaling factor. Local and global features are integrated via a linear fusion layer to produce a context-aware base representation. This unified representation preserves fine-grained local patterns while incorporating global semantics and long-range dependencies, providing high-quality input for subsequent modules. Multi-modal features fusion To explicitly model residue-level interaction dependencies between proteins and peptides, SaMCL introduces a bidirectional cross-attention module. This mechanism enables proteins and peptides to exchange contextual information and refine their residue representations, thereby capturing potential binding-site dependencies and interaction patterns. Assuming the input protein and peptide representations are denoted as and , respectively, the enhanced protein and peptide representations become and . The protein-to-peptide attention and updated protein representation are computed as: (5) (6) Here, , , and are learnable projection matrices, and the residual connection stabilizes training and preserves the original protein features. The same procedure is applied symmetrically for the peptide. Multi-task collaborative training To simultaneously optimize interaction prediction and binding site annotation, SaMCL designs a multi-task collaborative framework based on fused features. This framework comprises three parallel task branches: binary prediction for peptide–protein interaction, annotation of protein-binding sites in peptides, and identification of peptide-binding sites in proteins. Interaction prediction : This branch uses multi-layer perceptron (MLP) to determine whether interactions exist between proteins and peptides. (7) where is a linear concatenation of the enhanced protein and peptide representations ( and ). and are the trainable weight matrix and bias vector, respectively, and represents the sigmoid function. It employs binary cross-entropy [ 45 ] as the loss function: (8) where denotes the interaction label and represents the predicted binding probability. Binding site annotation : To identify binding sites at the residue level, SaMCL constructs two separate sequence-annotation tasks for proteins and peptides, respectively. Each branch projects the contextual embeddings of residues into label scores via an MLP predictor, then applies element-wise softmax to predict each residue independently: (9) (10) where is the length of the given protein and is the length of the given peptide. The corresponding loss function is the average cross-entropy over the residues: (11) (12) where and denote the true labels of the th protein residue and th peptide residue, respectively, and , represent their feature embeddings. Because the binding site prediction task is biologically meaningful only when genuine interactions exist, SaMCL calculates the loss of binding site annotation only for interaction-positive samples. Uncertainty-based task weighting : To enhance training stability and mitigate the risk of convergence to saddle points or suboptimal local minima, we adopt an uncertainty-weighted loss strategy that dynamically balances the contributions of different tasks by introducing a learnable uncertainty parameter . By adaptively scaling each task loss according to its estimated uncertainty, this strategy prevents tasks with larger loss magnitudes from dominating gradient updates and reduces gradient conflicts across branches, resulting in more stable and coordinated optimization. We define the total loss function as: (13) where denotes the loss term for task (including interaction prediction int , protein binding site identification pro , and peptide binding site annotation pep ), and represents its learnable uncertainty parameter. In this study, dropout technology and L2 regularization (built-in the AdamW optimizer) were used to constrain the parameter margin, stabilize training, and improve generalization. The overall learning curve of SaMCL is shown in Fig. S3 , and more details about the experimental setup and hyperparameter optimization ( Table S5 ) are provided in the Supplementary Material . Results Binary interaction prediction Benchmarking with other methods We evaluated its performance against several state-of-the-art methods using five-fold cross-validation [ 46 ] on randomly partitioned datasets in Table 1 , employing multiple metrics, including precision, recall, F1 score, accuracy (ACC), Matthews correlation coefficient (MCC), area under the receiver operating characteristic curve (AUC), and area under the precision–recall curve (AUPR) [ 47 ]. Table 1. Performance comparison of different methods on peptide–protein interaction prediction Methods Precision Recall MCC ACC F1 AUC AUPR DrugBAN 0.645 0.035 * * * 0.658 0.059 * * 0.298 0.048 * * * 0.650 0.025 * * * * 0.648 0.026 * * * 0.713 0.026 * * * * 0.695 0.020 * * * * DeepPepPI 0.732 0.016 * * * * 0.788 0.025 * * 0.518 0.048 * * 0.759 0.018 * * * 0.758 0.018 * * * 0.836 0.017 * * * * 0.828 0.021 * * CAMP 0.759 0.017 * * * * 0.768 0.028 * * 0.528 0.012 * * * * 0.768 0.015 * * * * 0.762 0.011 * * * 0.851 0.010 * * * 0.831 0.008 * * * IIDL-PepPI 0.785 0.018 * * * 0.798 0.012 * * * 0.605 0.016 * * * 0.788 0.007 * * * 0.795 0.013 * * 0.888 0.006 * * 0.875 0.009 * SaMCL 0.832 0.014 0.852 0.010 0.665 0.014 0.832 0.011 0.835 0.012 0.904 0.010 0.882 0.010 Open in a new tab Note: 1. Results are reported as mean standard deviation over five-fold cross-validation. 2. Statistical significance was evaluated using a paired two-sided t-test on the fold-wise performance, where * .05, * * .01, * * * .001, * * * * .0001. 3. The values in bold indicate the best performance for each metric. As demonstrated by the results, SaMCL exhibits outstanding performance across all metrics, fully showcasing its robust modeling capabilities in predicting peptide–protein interactions. This indicates that an integrated, structure-aware PLM captures complementary sequence-structure information more effectively, thereby enhancing the ability to identify interaction patterns. Generalization and robustness analysis To better evaluate the model’s generalization capability, we further designed three evaluation settings based on clustering protein and peptide sequences at different levels of sequence similarity to compare the seven metrics mentioned above with baseline methods [ 7 ]. As shown in Fig. 2 , SaMCL consistently achieves the highest performance across all settings, demonstrating superior generalization. The Novel Protein (c and d) shows the steepest decline in performance across methods, suggesting that generalizing to unseen proteins poses the most significant challenge. Interestingly, higher performance was observed under the Novel Peptide (a and b) compared with the Novel Protein (c and d) and Novel Pair (e and f), likely because the greater diversity of peptide sequences mitigates the impact of distribution shifts. Figure 2. Open in a new tab Performance comparison under different generalization scenarios across varying sequence similarity thresholds. (a and b) Novel Peptide: evaluation on peptides with sequence similarity below the specified thresholds relative to the training set. (c and d) Novel Protein: evaluation on proteins with sequence similarity below the specified thresholds relative to the training set. (e and f) Novel Pair: evaluation on peptide–protein pairs where both components are dissimilar to those in the training set. AUC (a, c, e) and AUPR (b, d, f) are reported for each scenario. NA denotes no clustering, while 0.9–0.6 represents the maximum allowed sequence similarity thresholds. Additionally, we evaluated these methods on an independent test set, as shown in Fig. 3 . Similarly, SaMCL achieved the highest performance. These results demonstrate that SaMCL is more robust and generalizes better across diverse evaluation conditions, further validating its effectiveness in protein–peptide interaction prediction. Figure 3. Open in a new tab The receiver operating characteristic (ROC) and precision–recall (PR) curves for SaMCL and the other advanced methods on the independent test set. Annotation of binding sites As shown in Table 2 , SaMCL outperforms other methods in both protein and peptide binding site prediction tasks, demonstrating the advantages of the proposed framework. Table 2. Performance comparison of different methods in identifying binding sites in peptides and proteins (a) Peptide binding site prediction Method Precision Recall MCC ACC F1 AUC AUPR CAMP 0.901 0.010 * * * 0.223 0.011 * * * * 0.230 0.010 * * * * 0.476 0.005 * * * * 0.357 0.070 * * * * 0.755 0.005 * * * * 0.834 0.005 * * * * IIDL-PepPI 0.800 0.009 * * * * 0.881 0.008 * * * * 0.504 0.013 * * * * 0.785 0.006 * * * * 0.843 0.007 * * * * 0.849 0.009 * * * * 0.908 0.010 * * SaMCL 0.940 0.008 0.966 0.007 0.824 0.012 0.931 0.006 0.953 0.006 0.935 0.005 0.921 0.003 (b) Protein binding site prediction Method Precision Recall MCC ACC F1 AUC AUPR PepBind 0.556 0.015 * * * * 0.049 0.010 * * * * 0.159 0.014 * * * * 0.949 0.007 * * 0.095 0.011 * * * * 0.643 0.005 * * * * 0.182 0.007 * * * * PepNN 0.157 0.013 * * * * 0.521 0.021 * * * 0.215 0.009 * * * * 0.831 0.004 * * * * 0.240 0.007 * * * * 0.754 0.003 * * * * 0.193 0.005 * * * * PepBCL 0.520 0.015 * * * * 0.261 0.010 * * * * 0.345 0.014 * * * * 0.949 0.017 0.345 0.011 * * * * 0.627 0.005 * * * * 0.403 0.007 * * * * PepCA 0.777 0.012 * * 0.330 0.014 * * * * 0.495 0.013 * * * * 0.961 0.007 0.462 0.007 * * * * 0.892 0.007 * * * * 0.548 0.010 * * * * IIDL-PepPI 0.717 0.014 * * * 0.425 0.018 * * * * 0.539 0.009 * * * * 0.965 0.004 * * * 0.533 0.004 * * * * 0.880 0.003 * * * * 0.569 0.007 * * * * SaMCL 0.807 0.012 0.625 0.012 0.687 0.009 0.955 0.004 0.704 0.007 0.946 0.003 0.789 0.005 Open in a new tab Note: 1. Results are reported as mean standard deviation over five-fold cross-validation. 2. Statistical significance was evaluated using a paired two-sided t-test on the fold-wise performance, where * .05, * * .01, * * * .001, * * * * .0001. 3. The values in bold indicate the best performance for each metric. In protein binding-site prediction, SaMCL shows a slightly lower ACC than PepCA and IIDL-PepPI. However, this difference is primarily attributable to the severe class imbalance inherent in protein sequences, in which binding residues typically constitute only a small fraction of the sequence. Under such conditions, accuracy can be dominated by the large number of non-binding residues, and models that adopt conservative strategies—predicting most residues as non-binding—may achieve artificially inflated ACC while failing to identify true binding sites effectively. On these threshold-independent metrics, such as AUC and AUPR, SaMCL consistently outperforms competing methods, indicating superior sensitivity and robustness in recognizing true binding residues. In addition, we selected two peptide–protein complexes to visually illustrate SaMCL’s effectiveness in predicting binding residues, as shown in Fig. 4 . These results suggest that SaMCL can better focus on biologically critical binding regions rather than adopting conservative predictions, thereby providing more precise binding-site localization. Figure 4. Open in a new tab Performance comparison of SaMCL and IIDL-PepPI on peptide- and protein-binding residue prediction using two representative complexes. (a) For complex 3D25, SaMCL’s peptide-binding residue prediction ranks within the top 1%, while its protein-binding residue prediction ranks around the top 25%. (b) For complex 5XBO, SaMCL’s protein-binding residue prediction ranks within the top 1%, and its peptide-binding residue prediction ranks around the top 25%. Ablation study on model architecture To investigate the impact of different module designs on the method performance, we conducted an ablation analysis on SaMCL. The results are shown in Fig. 5 . Figure 5. Open in a new tab Ablation study on feature components and architectural modules of SaMCL. (a,c,e) showing AUC performance when removing individual feature types: One-hot encoding, physicochemical properties (Phys), or SaProt embeddings (SaProt). (b,d,f) showing AUC performance when removing MSCNN, BiLSTM-Att, or Cross Att. Results are presented for three prediction tasks: peptide–protein interaction (PepPI), peptide binding sites (Peptide), and protein binding sites (Protein). Statistical significance uses t-test: * .05, * * .01. In feature ablation, removing SaProt embeddings causes the largest performance degradation across all three tasks, making it the most critical feature component. Removing One-hot encoding or physicochemical features results in modest but statistically significant decreases, confirming their complementary functions. In architectural ablation, we evaluated three variant models in which key modules were selectively removed: w/o MSCNN, w/o BiLSTM Att, and w/o Cross Att. In the w/o MSCNN, sequence representations are directly fed into the subsequent BiLSTM and Cross Attention modules. In the w/o BiLSTM Att, the outputs of MSCNN are projected via a simple linear layer to match feature dimensions before entering the Cross Attention module. In the w/o Cross Att, protein and peptide branches are processed independently and then simply concatenated for the final prediction. All ablation models use the same data splits, training protocols, and hyperparameters as the full model. The results show that the cross-attention module has the most significant impact on interaction prediction, while MSCNN and BiLSTM Att also contribute to performance across downstream tasks. Impact of protein language models The architecture of the proposed method determines that the choice of PLMs significantly impacts SaMCL’s performance. Therefore, we evaluated SaMCL with different PLMs (including ESM-2, SaProt, and ProtSST) and their LoRA-fine-tuned variants across three tasks. Table 3 presents a comparison of AUC. The ACC values are shown in Table S6 in the Supplementary Materials . SaProt achieved the best results among three PLMs, with and without LoRA fine-tuning, likely due to its structure-aware pretraining, which captures 3D protein information. Table 3. Performance comparison of different pretrained PLMs and their LoRA fine-tuned versions across multiple tasks Model PepPI Peptide Protein ESM2 0.864 0.005 0.807 0.006 0.856 0.005 ProtSST 0.852 0.006 0.795 0.006 0.846 0.006 SaProt 0.878 0.004 0.870 0.005 0.862 0.006 ESM2_LoRA 0.882 0.004 0.878 0.004 0.928 0.004 ProtSST_LoRA 0.864 0.005 0.848 0.005 0.903 0.005 SaProt _ LoRA 0.904 0.003 0.974 0.003 0.964 0.004 Open in a new tab The values in bold indicate the best performance for each metric. In Fig. 6a , the “performance improvement” is defined as the relative performance gain achieved by LoRA fine-tuning compared with the same language model without fine-tuning, measured in terms of ACC and AUC. It shows that all tasks exhibit improved performance after LoRA fine-tuning, with the most significant gains observed in the site prediction task. Compared with SaMCL, ESM-2 and ProtSST demonstrate smaller improvements and benefit less from interaction prediction, reflecting their stronger reliance on sequence patterns. Figure 6b radar chart directly compares SaProt with SaProt-LoRA, more explicitly revealing the effectiveness of LoRA fine-tuning. Figure 6. Open in a new tab (a) Heatmap showing AUC and ACC performance improvements (%) after LoRA fine-tuning across three PLMs for peptide–protein interaction (PepPI), peptide binding site (Peptide), and protein binding site (Protein) predictions. (b) Radar plot comparing SaProt and SaProt-LoRA across six performance metrics. Effectiveness of multi-task learning framework To validate the effectiveness of multi-task collaborative learning, we designed two sets of comparative experiments: single-task learning, in which models are trained independently for each task, and multi-task collaborative Learning. As shown in Fig. 7 , the multi-task collaborative learning framework consistently outperformed single-task learning models across all three subtasks. The results highlight the benefit of joint optimization. Figure 7. Open in a new tab Impact of multi-task learning on SaMCL performance. The right represents multi-task learning, and the left represents single-task models. Statistical significance uses t-test: * .05, * * .01, * * * .001. Efficiency analysis Beyond its performance advantages, SaMCL is more efficient than competing methods (e.g. CAMP and IIDL-PepPI) that rely on precomputed feature extraction using tools such as PSI-BLAST [ 48 ], IUpred2A [ 49 ], and SCRATCH [ 50 ]. These tools introduce substantial computational overhead (minutes per sample) that outweighs SaMCL’s runtime. To further evaluate the trade-off between model complexity and efficiency, we additionally compared SaMCL with lightweight baseline algorithms, including random forest (RF), MLP, and k-nearest neighbors (KNN), using the same multi-modal input features. The results are shown in Table 4 . Although these baselines exhibit lower raw inference time, their predictive performance is notably inferior, particularly in modeling complex peptide–protein interactions and residue-level binding sites. In contrast, SaMCL consistently achieves superior prediction performance across tasks, indicating that the combination of fine-tuned SaProt embeddings, cross-attention mechanisms, and multi-task learning more effectively captures intricate inter-molecular relationships. Overall, these results suggest that SaMCL attains a favorable balance between accuracy and efficiency, providing substantial performance gains while maintaining practical inference costs. Table 4. Comprehensive comparison of predictive performance and inference efficiency across different tasks Task Model ACC Precision Recall F1 MCC AUC AUPR Runtime Interaction RF 0.5632 0.6377 0.4632 0.5366 0.1493 0.6721 0.7028 0.001210 MLP 0.5632 0.6377 0.4632 0.5366 0.1493 0.6721 0.7028 0.000829 KNN 0.5805 0.6250 0.5789 0.6011 0.1606 0.6685 0.7090 0.001215 SaMCL 0.8321 0.8324 0.8520 0.8351 0.6654 0.9040 0.8821 0.161703 Peptide RF 0.7411 0.7617 0.9536 0.8469 0.0978 0.7275 0.8913 0.000005 MLP 0.7297 0.8394 0.7917 0.8148 0.3183 0.7620 0.8951 0.011666 KNN 0.7952 0.7940 0.9820 0.8781 0.3591 0.7583 0.8953 0.007567 SaMCL 0.9316 0.9408 0.9660 0.9536 0.8242 0.9355 0.9213 0.161703 Protein RF 0.8869 0.3089 0.3914 0.3453 0.2867 0.7466 0.3821 0.000050 MLP 0.9310 0.5714 0.3908 0.4642 0.4375 0.7915 0.4860 0.011666 KNN 0.9403 0.7617 0.3146 0.4452 0.4657 0.7805 0.4678 0.081159 SaMCL 0.9554 0.8071 0.6252 0.7047 0.6879 0.9463 0.7895 0.161703 Open in a new tab Interaction, peptide–protein interaction prediction; Pep, peptide binding-site identification; Pro, protein binding-site identification. Runtime is reported in microsecond/sample. The values in bold indicate the best performance for each metric. Feature analysis To further investigate the contributions of the different components to the proposed model, we visualized the learned embeddings at various stages of the interaction prediction pipeline using t-SNE [ 51 ], as shown in Fig. 8 . At the input stage, the embeddings of interacting and non-interacting pairs are largely mixed; after feature learning, slight grouping tendencies emerge. The cross-attention further enhances class separability, as reflected by more distinct clusters between interaction and non-interaction samples after feature fusion. Finally, the clusters become more compact, and the boundary between the two classes becomes clearer when using the final latent features, consistent with the increase in the silhouette coefficient. This progressive improvement demonstrates that the model successfully refines latent representations layer by layer, transforming initial heterogeneous embeddings into a more structured and task-relevant space. Figure 8. Open in a new tab The t-SNE visualization of feature representations at different stages of SaMCL for interaction prediction. Input (initial feature concatenation), Feature Fusion (after MSCNN and BiLSTM-Attention), Cross Attention (after inter-molecular interaction modeling), and Output (final prediction layer). Silhouette scores quantify the quality of cluster separation, with higher values indicating better discrimination. Conclusion This study proposes SaMCL, a structure-aware multi-task collaborative learning framework for protein–peptide interaction prediction. Experiment results show that SaMCL consistently outperforms existing methods across interaction prediction and residue-level binding-site identification. LoRA fine-tuning of SaProt effectively transfers structural knowledge to downstream tasks, while the multi-task learning strategy enables shared representations that enhance generalization. Although the dual-branch optimization setup is challenging, experimental results across multiple random seeds have demonstrated the model’s stability (see Table S7 in the Supplementary Material ). These results demonstrate the feasibility and effectiveness of the proposed method. Despite these advantages, several limitations remain. The framework relies on high-quality structural data, and its performance is therefore affected by the availability and quality of experimentally resolved structures. Although tools such as AlphaFold3 can predict protein and peptide structures, samples for which high-quality predicted structures remain inaccessible may still be affected (see Figs S5–S7 in the Supplementary Material ). The model is trained on structure-derived training data, which may favor stable, crystallizable interactions and underperform on dynamic, transient, or intrinsically disordered interactions due to limited supervision. Moreover, the use of synthetic negative samples may not fully reflect real biological distributions, potentially introducing noise into the model’s learning. Future work will explore incorporating more diverse structural representations and more realistic negative sampling strategies to improve robustness and generalization further. Key points Present a multi-task learning framework capable of jointly predicting peptide–protein binary interactions, peptide-binding residues, and protein-binding residues. Integrate the structure-aware protein language model SaProt and apply task-specific fine-tuning to transfer structural information and improve prediction performance effectively. Employ a dual-stream architecture with dedicated encoders for proteins and peptides; a cross-attention module aligns residue-level contextual embeddings between the two molecules, enabling deep bidirectional information exchange. Achieve state-of-the-art performance in three tasks, confirming the critical role of structure-aware embeddings and key architectural components. Supplementary Material Supplementary_Material_bbag178 supplementary_material_bbag178.pdf (2.6MB, pdf) Contributor Information Siyi He, School of Artificial Intelligence, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China; National Engineering Laboratory for Big Data System Computing Technology, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China. Dongzhen Tang, School of Computer Science, Guangdong University of Technology, 100 Waihuan West Road, 510006 Guangzhou, China. Tiantian Zhu, School of Artificial Intelligence, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China; National Engineering Laboratory for Big Data System Computing Technology, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China. Zexuan Zhu, School of Artificial Intelligence, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China; National Engineering Laboratory for Big Data System Computing Technology, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China. Yumeng Liu, School of Artificial Intelligence, Shenzhen Technology University, 3002 Lantian Road, 518118 Shenzhen, China. Jun Zhang, School of Artificial Intelligence, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China; National Engineering Laboratory for Big Data System Computing Technology, Shenzhen University, 3688 Nanhai Avenue, 518060 Shenzhen, China. Conflicts of interest None declared. Funding This work is supported by the National Natural Science Foundation of China (grant nos 62302311, 62302316, 62406199, and 62471310), the National Key Research and Development Program of China (grant no. 2022YFF1202104), the Guangdong Basic and Applied Basic Research Foundation (grant no. 2024A1515011681), Shenzhen Research Initiation Program for High-Caliber Critical Talent (grant no. 827-000932), and the Internal Fund of National Engineering Laboratory for Big Data System Computing Technology (grant nos SZU-BDSC-IF2024-01 and SZU-BDSC-IF2024-10). Data availability The datasets and the code of the proposed method are available at https://github.com/xh20011225/SaMCL . References 1. Fosgerau K, Hoffmann T. Peptide therapeutics: current status and future directions. Drug Discov Today 2015;20:122–8. 10.1016/j.drudis.2014.10.003 [ DOI ] [ PubMed ] [ Google Scholar ] 2. Konc J, Janežič D. Protein binding sites for drug design. Biophys Rev 2022;14:1413–21. 10.1007/s12551-022-01028-3 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Kortemme T, Joachimiak LA, Bullock AN et al. Computational redesign of protein-protein interaction specificity. Nat Struct Mol Biol 2004;11:371–9. 10.1038/nsmb749 [ DOI ] [ PubMed ] [ Google Scholar ] 4. Horie Y, Suzuki T, Inoue J et al. Molecular basis for the disruption of Keap1-Nrf2 interaction via Hinge &Latch mechanism. Commun Biol 2021;4:576. 10.1038/s42003-021-02100-6 [ DOI ] [ Google Scholar ] 5. Zhang X, Wei L, Ye X et al. SiameseCPP: a sequence-based siamese network to predict cell-penetrating peptides by contrastive learning. Brief Bioinform 2023;24:bbac545. 10.1093/bib/bbac545 [ DOI ] [ Google Scholar ] 6. Hasan MM, Alam MA, Shoombuatong W et al. NeuroPred-FRL: an interpretable prediction model for identifying neuropeptide using feature representation learning. Brief Bioinform 2021;22:bbab167. 10.1093/bib/bbab167 [ DOI ] [ Google Scholar ] 7. Lei Y, Li S, Liu Z et al. A deep-learning framework for multi-level peptide–protein interaction prediction. Nat Commun 2021;12:5465. 10.1038/s41467-021-25772-4 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. LeCun Y, Bottou L, Bengio Y et al. Gradient-based learning applied to document recognition. Proc IEEE 2002;86:2278–324. 10.1109/5.726791 [ DOI ] [ Google Scholar ] 9. Bai P, Miljković F, John B et al. Interpretable bilinear attention network with domain adaptation improves drug–target prediction. Nat Mach Intell 2023;5:126–36. 10.1038/s42256-022-00605-1 [ DOI ] [ Google Scholar ] 10. Wang Z, Meng J, Dai Q et al. DeepPepPI: a deep cross-dependent framework with information sharing mechanism for predicting plant peptide-protein interactions. Expert Syst Appl 2024;252:124168. 10.1016/j.eswa.2024.124168 [ DOI ] [ Google Scholar ] 11. Chen S, Yan K, Li X et al. Protein language pragmatic analysis and progressive transfer learning for profiling peptide–protein interactions. IEEE Trans Neural Netw Learn Syst 2025;36:15385–99. 10.1109/TNNLS.2025.3540291 [ DOI ] [ PubMed ] [ Google Scholar ] 12. Das AA, Sharma OP, Kumar MS et al. PepBind: a comprehensive database and computational tool for analysis of protein–peptide interactions. Genomics Proteomics Bioinform 2013;11:241–6. 10.1016/j.gpb.2013.03.002 [ DOI ] [ Google Scholar ] 13. Abdin O, Nim S, Wen H et al. PepNN: a deep attention model for the identification of peptide binding sites. Commun Biol 2022;5:503. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Wang R, Jin J, Zou Q et al. Predicting protein–peptide binding residues via interpretable deep learning. Bioinformatics 2022;38:3351–60. 10.1093/bioinformatics/btac352 [ DOI ] [ PubMed ] [ Google Scholar ] 15. Huang J, Li W, Xiao B et al. PepCA: unveiling protein-peptide interaction sites with a multi-input neural network model. Iscience 2024;27:110850. 10.1016/j.isci.2024.110850 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Li H, Liu B. BioSeq-Diabolo:: biological sequence similarity analysis using Diabolo. PLoS Comput Biol 2023;19:e1011214. 10.1371/journal.pcbi.1011214 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Rives A, Meier J, Sercu T et al. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. Proc Natl Acad Sci USA 2021;118:e2016239118. 10.1073/pnas.2016239118 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Zhou K, Lei C, Zheng J et al. Pre-trained protein language model sheds new light on the prediction of arabidopsis protein–protein interactions. Plant Methods 2023;19:141. 10.1186/s13007-023-01119-6 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Elnaggar A, Heinzinger M, Dallago C et al. ProtTrans: toward understanding the language of life through self-supervised learning. IEEE Trans Pattern Anal Mach Intell 2021;44:7112–27. [ Google Scholar ] 20. Tran DT, Pham NT, Nguyen NDH et al. HyPepTox-Fuse: an interpretable hybrid framework for accurate peptide toxicity prediction fusing protein language model-based embeddings with conventional descriptors. J Pharma Anal 2025;15:101410. 10.1016/j.jpha.2025.101410 [ DOI ] [ Google Scholar ] 21. Senior AW, Evans R, Jumper J et al. Improved protein structure prediction using potentials from deep learning. Nature 2020;577:706–10. 10.1038/s41586-019-1923-7 [ DOI ] [ PubMed ] [ Google Scholar ] 22. Jumper J, Evans R, Pritzel A et al. Highly accurate protein structure prediction with alphafold. Nature 2021;596:583–9. 10.1038/s41586-021-03819-2 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Abramson J, Adler J, Dunger J et al. Accurate structure prediction of biomolecular interactions with alphafold 3. Nature 2024;630:493–500. 10.1038/s41586-024-07487-w [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Jin S, Han C, Zhou Y et al. SaProt: protein language modeling with structure-aware vocabulary. Nat Methods 2024; 21 :1234–1242. 10.1038/s41592-024-02264-6 [ DOI ] [ Google Scholar ] 25. Van Kempen M, Kim SS, Tumescheit C et al. Fast and accurate protein structure search with foldseek. Nat Biotechnol 2024;42:243–6. 10.1038/s41587-023-01773-0 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Li M, Tan Y, Ma X et al. ProSST: protein language modeling with quantized structure and disentangled attention. Adv Neural Inf Process Syst 2024;37:35700–26. [ Google Scholar ] 27. Caruana R. Multitask learning. Multitask learning Machine learning 1997;28:41–75. 10.1023/A:1007379606734 [ DOI ] [ Google Scholar ] 28. Devlin J, Chang M-W, Lee K et al. BERT: pre-training of deep bidirectional transformers for language understanding. In: Burstein J, Doran C, Solorio T, (eds.), Proc. 2019 Conf. North Am. Chap. Assoc. Comput , Vol. 1 (Long Short Papers) , pp. 4171–4186. Linguist: Hum. Lang. Technol, 2019. [ Google Scholar ] 29. Lin H, Cheng X, Xiangyu W et al. CAT: cross attention in vision transformer. In: Huet B, Gravier G, Chang S-F, et al. , (eds.), Proc. IEEE Int. Conf. Multimedia Expo (ICME) , pp. 1–6. Taipei, Taiwan: IEEE, 2022. [ Google Scholar ] 30. Berman HM, Westbrook J, Feng Z et al. The protein data bank. Nucleic Acids Res 2000;28:235–42. 10.1093/nar/28.1.235 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Burley SK, Berman HM, Bhikadiya C et al. RCSB Protein Data Bank: biological macromolecular structures enabling research and education in fundamental biology, biomedicine, biotechnology and energy. Nucleic Acids Res 2019;47:D464–74. 10.1093/nar/gky1004 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Salentin S, Sven Schreiber V, Haupt J et al. PLIP: fully automated protein–ligand interaction profiler. Nucleic Acids Res 2015;43:W443–7. 10.1093/nar/gkv315 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. van Kempen M, Kim SS, Tumescheit C et al. Fast and accurate protein structure search with Foldseek. Nat Biotechnol 2024; 42 :243–246. 10.1038/s41587-023-01773-0 [ DOI ] [ Google Scholar ] 34. Chen S, Yan K, Liu B. PDB-BRE: a ligand–protein interaction binding residue extractor based on protein data bank. Proteins Struct Funct Bioinform 2024;92:145–53. 10.1002/prot.26596 [ DOI ] [ Google Scholar ] 35. Limin F, Niu B, Zhu Z et al. CD-HIT: accelerated for clustering the next-generation sequencing data. Bioinformatics 2012;28:3150–2. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 36. Hu EJ, Shen Y, Wallis P et al. LoRA: low-rank adaptation of large language models. Int Conf Learn Represent 2022;1:3. [ Google Scholar ] 37. Cai Z, Fan Q, Feris RS et al. A unified multi-scale deep convolutional neural network for fast object detection. In: Leibe B, Matas J, Sebe N, Welling M, (eds.), Eur. Conf. Comput. Vis (ECCV), Lecture Notes in Computer Science, vol 9908. Amsterdam, The Netherlands: Springer; 2016: 354–370. 10.1007/978-3-319-46493-0_22. [ DOI ] [ Google Scholar ] 38. Hochreiter S, Schmidhuber J. Long short-term memory. Neural Comput 1997;9:1735–80. 10.1162/neco.1997.9.8.1735 [ DOI ] [ PubMed ] [ Google Scholar ] 39. Vaswani A, Shazeer N, Parmar N et al. Attention is all you need. Adv Neural Inf Process Syst 2017;30:5998–6008. [ Google Scholar ] 40. Kendall A, Gal Y, Cipolla R. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In: Forsyth D, Torr P, Zisserman A, (eds.), Proc. IEEE Conf. Comput. Vis. Pattern Recognit (CVPR) . Salt Lake City, UT, USA: IEEE; pp. 7482–91, 2018. 41. Loshchilov I, Hutter F. Decoupled weight decay regularization. In: 7th Int. Conf. Learn. Represent. (ICLR 2019). New Orleans, LA, USA: International Conference on Learning Representations; 2019. arXiv:1711.05101. 2017. 42. Sangaraju VK, Pham NT, Wei L et al. mACPpred 2.0: stacked deep learning for anticancer peptide prediction with integrated spatial and probabilistic feature representations. J Mol Biol 2024;436:168687. 10.1016/j.jmb.2024.168687 [ DOI ] [ PubMed ] [ Google Scholar ] 43. Basith S, Sangaraju VK, Manavalan B et al. mHPpred: accurate identification of peptide hormones using multi-view feature learning. Comput Biol Med 2024;183:109297. 10.1016/j.compbiomed.2024.109297 [ DOI ] [ PubMed ] [ Google Scholar ] 44. Shaw P, Uszkoreit J, Vaswani A. Self-attention with relative position representations. In: Proc. 2018 Conf. North Am. Chap. Assoc. Comput. Linguist.: Human Lang. Technol., Volume 2 (Short Papers), pages 464–468. Association for Computational Linguistics; 2018. 10.18653/v1/N18-2074 [ DOI ] 45. Hastie T, Tibshirani R, Friedman J. The Elements of Statistical Learning: Data Mining, Inference, and Prediction, 2nd ed. New York, NY, USA: Springer; 2009. 10.1007/978-0-387-84858-7. [ DOI ] [ Google Scholar ] 46. Ruppert D. The Elements of Statistical Learning: Data Mining, Inference, and Prediction, [review of the book by Hastie T, Tibshirani R, Friedman J]. J Am Stat Assoc 2004;99:567. 10.1198/jasa.2004.s339. [ DOI ] [ Google Scholar ] 47. Powers DMW. Evaluation: from precision, recall and F‑measure to ROC, informedness, markedness and correlation. J Mach Learn Technol 2011;2:37–63. 10.48550/arXiv.2010.16061 [ DOI ] [ Google Scholar ] 48. Altschul SF, Madden TL, Schäffer AA et al. Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. Nucleic Acids Res 1997;25:3389–402. 10.1093/nar/25.17.3389 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 49. Dosztányi Z, Csizmok V, Tompa P et al. IUPred: web server for the prediction of intrinsically unstructured regions of proteins based on estimated energy content. Bioinformatics 2005;21:3433–4. 10.1093/bioinformatics/bti541 [ DOI ] [ PubMed ] [ Google Scholar ] 50. Resnick M, Maloney J, Monroy-Hernández A et al. Scratch: programming for all. Commun ACM 2009;52:60–7. 10.1145/1592761.1592779 [ DOI ] [ Google Scholar ] 51. van der Maaten L, Hinton G. Visualizing data using t-SNE. J Mach Learn Res 2008;9:2579–605. [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary_Material_bbag178 supplementary_material_bbag178.pdf (2.6MB, pdf) Data Availability Statement The datasets and the code of the proposed method are available at https://github.com/xh20011225/SaMCL . Articles from Briefings in Bioinformatics are provided here courtesy of Oxford University Press ACTIONS View on publisher site PDF (3.8 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top