Prediction and analysis of anti-aging peptides using data augmentation and machine learning algorithms Authors: Zhiyuan Zhang, Yuanyuan Chen, Shihao Wang, Guozhong Chen, Mingyang Wang, Yuanyuan Pan, Erguang Li Published: 2025 ABSTRACT AbstractBackgroundFor most species, Aging is an inevitable biological process that poses significant challenges to global healthcare due to age-related diseases. Recent advances in peptide therapy have highlighted anti-aging peptides (AAPs) as a promising therapeutic strategy, owing to their low immunogenicity and ease of synthesis. However, the lack of computational tools for AAP identification has hindered systematic research in this field.ResultsIn this study, we provided a benchmark dataset of anti-aging peptides (AAPs) based on their annotated biological functions and AgingBase database. Subsequently, we proposed three novel predictive models, including 1) Antiaging-FL, which integrates feature representation learning with machine learning algorithms; 2) ESM_GAN, which utilizes a generative adversarial network (GAN) for data augmentation to enhance prediction robustness; and 3) ESM_CNN, which combines a data augmentation strategy based on conservative amino acid substitutions with a convolutional neural network (CNN) architecture for improved feature extraction. Comprehensive evaluations demonstrated that all three models achieved high predictive performance. Specifically, Antiaging-FL achieved an AUC of 1.00 on the AAP400 dataset and 0.99 on an independent test set, while ESM_GAN yielded AUCs of 0.99 and 0.95, and ESM_CNN obtained AUCs of 0.96 and 0.94, respectively.ConclusionsOur study presents three high-performance models for AAP prediction, accelerating the discovery of novel anti-aging peptides. These models not only provide a valuable resource for researchers but also offer insights into the functional mechanisms of AAPs, paving the way for targeted drug development in aging-related therapeutics.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12915-025-02469-4.Keywords: Aging, Anti-aging peptide, Data augmentation, Machine learning, Deep learning Background Aging is a complex physiological process mediated by multiple biological and genetic pathways. It is directly associated with lifespan and serves as a major driving force behind all age-related diseases [1, 2]. Disorders arising from aging, such as cancer, cardiovascular diseases, neurodegenerative conditions, and metabolic dysfunctions, not only significantly contribute to increased mortality among the elderly but also pose substantial challenges to global healthcare systems [3–7]. Therefore, developing effective strategies to design anti-aging interventions is critical for extending human lifespan. In recent years, a variety of anti-aging strategies have been developed to target aging-related pathways [8]. Among them, anti-aging peptides (AAPs), which are collectively defined as those that can improve aging markers (such as SA-β-gal, inflammatory factors, telomere length, etc.), extend the lifespan of model organisms, and regulate classical aging pathways, have attracted considerable attention as a promising class of bioactive molecules for therapeutic intervention [9–11]. They are collectively defined as anti-aging peptides. These peptides can participate in intercellular communication, regulate enzymatic activities (e.g., collagenases), facilitate the transport and delivery of biomolecules, and inhibit the production of neurotransmitters such as acetylcholine, thereby contributing to the mitigation of aging-related processes [12–14]. Given their biological significance, AAPs hold great potential as novel therapeutic agents in anti-aging medicine. Consequently, the accurate identification and discovery of AAPs is of vital importance. However, due to the fragmented and limited availability of related data, few studies have focused on the computational identification of AAPs, despite some researchers conducting analysis related to aging [15, 16]. To date, no dedicated predictor exists to distinguish anti-aging peptides from non-anti-aging ones. Thus, building a comprehensive dataset and developing high-performance predictive models based on such data remains a significant challenge. To address these challenges, we provided a benchmark dataset of AAPs based on their annotated biological functions and AgingBase database (Table S1) and developed a novel predictive model, Antiaging-FL, which leverages machine learning techniques to accurately predict anti-aging peptides [17]. Experimental results demonstrated that Antiaging-FL achieved robust performance on both the AAP400 dataset and an independent test set. Furthermore, to overcome the limitations posed by the small sample size, which may lead to model overfitting and poor generalization, so we employed two distinct data augmentation strategies. The resulting augmented datasets enabled the construction of predictive models that also performed well on both datasets. Accordingly, we retained the three best-performing models for public use and developed a user-friendly web server, Antiaging-FL, accessible at http://124.222.115.164/# or http://www.stay67.com. We hope this tool will facilitate further research in this domain. An overview of the dataset construction and predictive model development is illustrated in Fig. 1. The dataset construction and predictive framework for anti-aging peptides. Peptide sequences were collected from public databases, including AagingBase, peptideDB and Uniprot, to construct benchmark dataset, more details of the procedures are provided in the Methods section.Two data augmentation strategies were employed, including one based on generative adversarial network (GAN), and another based on amino acid conservative substitution. After obtaining the augmented datasets, predictive models were constructed using the original AAP400 dataset and the two augmented datasets. Specifically, a feature representation learning strategy was proposed, where features extracted from 40 manually designed descriptors were used as inputs for building machine learning models. In addition, machine learning models were developed using GAN-augmented datasets combined with ESM-derived features, and deep learning models were constructed based on datasets generated through the amino acid conservative substitution strategy. Further details on dataset construction, data augmentation, feature extraction, feature selection, and model development can be found in the Methods section Results Analysis of anti-aging peptides using amino acids composition and machine learning AAP400 dataset is based on peptide’s annotated biological functions and AgingBase database (Fig. 2A). The analysis of the lengths of anti-aging peptide sequences (Fig. 2B) reveals that the log2 values of the peptide sequence lengths for the positive and negative datasets are 3.0 and 4.8, respectively, with maximum lengths of 80 and 50, respectively. We utilized the number of amino acids in the peptide sequences as an indicator to ensure the acquisition of high-quality data. Subsequently, to discern if certain types of residues are preferred in anti-aging peptides, we compared the amino acid residues present in anti-aging peptides and non-anti-aging peptides, as illustrated in Fig. 2C. The data indicates that there are some significant differences in the content of certain amino acids. In anti-aging peptides, the content of R (Arginine), G (Glycine), P (Proline), and K (Lysine) is notably higher compared to the remaining amino acids. In non-anti-aging peptides, the content of N (Asparagine), I (Isoleucine), K (Lysine), S (Serine), and L (Leucine) is significantly higher than that of other amino acids. Furthermore, the content of R (Arginine), G (Glycine), and P (Proline) in anti-aging peptides is significantly higher than the corresponding amino acid content in non-anti-aging peptides, with a difference of 0.059 for R (Arginine). The content of N (Asparagine) and I (Isoleucine) is significantly lower in anti-aging peptides compared to the corresponding amino acid content in non-anti-aging peptides, with a difference of 0.076 for I (Isoleucine). These findings suggest that significantly different amino acids, such as R (Arginine) and I (Isoleucine), may play a crucial role in distinguishing between anti-aging peptides and non-anti-aging peptides. We employed machine learning methods to validate this hypothesis (Fig. 2D). By extracting the content of the 20 types of amino acids in the peptide sequences as data features to construct the model and outputting the feature importance scores, we found that the top five features were I (Isoleucine), N (Asparagine), Y (Tyrosine), L (Leucine), and R (Arginine), which is highly consistent with the aforementioned analysis results. Through these analysis, we have demonstrated the excellent ability of machine learning methods in assisting in the analysis of the amino acid importance of sequence. The sequence analysis of AAP400 dataset. A The negative dataset comprised three categories of peptides, including 1) inflammation-related peptides, which are known to promote aging; 2) bioactive peptide precursors; and 3) cytotoxic peptides. The latter two categories have not been reported in the literature to be associated with aging and were therefore considered unrelated to aging in this study. B The analysis of the lengths of anti-aging peptide sequences (Additional file 1: Table S2). C Amino acid compositional comparison of anti-aging peptides and non-anti-aging peptides. D Machine learning methods assists in the analysis of the amino acid importance of sequence Comparative analysis of sequence-based feature descriptors using different classifiers In this study, we employed nine sequence-based feature encoding algorithms including AAC, DPC, TPC, CTD, CTriad, CKSAAP, PSSM, DDE, and PAAC, to extract informative features from peptide sequences [18–25]. To ensure diversity, effectiveness, and uniqueness in feature extraction, we applied several parameter settings: the parameter for AAC was set to 1 or 2; for CTriad, the gap parameter was varied from 0 to 5; and for CKSAAP, the gap parameter was set from 2 to 6. Additionally, both frequency and occurrence counts were considered in EAAC, with its gap parameter ranging from 5 to 10. We also combined AAC with DPC and TPC to enrich the feature representation. Notably, EGAAC and GAAC were included as derived variants of AAC. In total, we generated 40 feature descriptors, as summarized in Table 1. Table 1The performance of 41 feature descriptors on the AAP400 datasetFeatureSensitivitySpecificityPrecisionF1-scoreAAC0.9350.9300.9350.933CKSAAP(k = 2)0.9700.8900.9700.933CKSAAP(k = 3)0.9650.9050.9650.937CKSAAP(k = 4)0.9500.9050.9500.930CKSAAP(k = 5)0.9400.9150.9400.928CKSAAP(k = 6)0.9250.9150.9250.920CTriad(k = 0)0.9050.8250.9050.869CTriad(k = 1)0.9050.8250.9050.869CTriad(k = 2)0.9050.8250.9050.869CTriad(k = 3)0.9050.8250.9050.869CTriad(k = 4)0.9050.8250.9050.869CTriad(k = 5)0.9050.8250.9050.869DDE0.9300.9350.9300.933DPC0.9650.8800.9650.927CTDC0.8550.8950.8550.872CTDD0.9600.8250.9600.900CTDT0.9150.8500.9150.887AAC_count0.9200.9550.9200.936PAAC0.9050.9450.9050.923AAC + DPC0.9550.9350.9550.946AAC + TPC0.9600.9200.9600.942CTD0.9550.8250.9550.898EAAC(k = 7)0.9050.9350.9050.919EAAC(k = 5)0.9150.9350.9150.924EAAC(k = 6)0.9100.9350.9100.921EAAC(k = 8)0.9000.9400.9000.918EAAC(k = 9)0.8800.9400.8800.906EAAC(k = 10)0.8700.9450.8700.903p_EAAC(k = 7)0.9500.9100.9500.932p_EAAC(k = 5)0.9250.9000.9250.914p_EAAC(k = 6)0.9350.9100.9350.924p_EAAC(k = 8)0.9350.9150.9350.926p_EAAC(k = 9)0.9250.9000.9250.915p_EAAC(k = 10)0.9200.9000.9200.912EGAAC(k = 5)0.9400.7700.9400.868GDPC0.9350.7800.9350.869GTPC0.9850.8050.9850.906TPC0.9650.8200.9650.901GAAC0.8750.7800.8750.835PSSM0.9250.9150.9000.915ESM0.90.770.90.849 The performance of 41 feature descriptors on the AAP400 dataset Using these 40 descriptors, Support Vector Machine (SVM) classifiers were trained on the AAP400 dataset and their performance was evaluated via fivefold cross-validation. As illustrated in Fig. 3A and B, the prediction accuracy of SVM models constructed with individual descriptors ranged from 0.8275 (GAAC) to 0.945 (AAC + DPC). The Matthews correlation coefficient (MCC), which provides a comprehensive measure of model performance, ranged from 0.66 (GAAC) to 0.89 (AAC + DPC). These results indicate that the model based on the combined AAC and DPC descriptors demonstrated superior discriminative capability in distinguishing anti-aging peptides from non-anti-aging peptides. From a biological perspective, the AAC and DPC features capture the global amino acid composition and local dipeptide sequence patterns, respectively. The AAC descriptor reflects the physicochemical properties and overall functional tendencies of the peptide, whereas DPC reveals sequence patterns linked to structural conformation and specific biological activity. These biologically relevant features contributed to the improved predictive performance of the model. Other performance metrics, including sensitivity (0.955), specificity (0.935), accuracy (0.955), and F1-score (0.946), further validated the robustness of the AAC + DPC-based model (Table 1). In addition, excellent performance was observed for models based on the DDE descriptor (accuracy = 0.933; MCC = 0.866) and CKSAAP with k = 2 (accuracy = 0.930; MCC = 0.865). Comparison of different feature extraction methods. A-B The performance of 40 handcrafted feature descriptors was evaluated based on accuracy (ACC) and Matthews correlation coefficient (MCC) (Additional file 1: Table S3). C Structural features of peptides were extracted and used to construct predictive models with five different machine learning algorithms. Model performance was comprehensively assessed using metrics such as sensitivity, specificity, accuracy, and MCC (Additional file 1: Table S4). D High-dimensional features were extracted using the ESM-1b model (650 million parameters), which captures evolutionary information, structural properties, and contextual semantics from peptide sequences. These features were then used for model construction and performance evaluation (Additional file 1: Table S4) Comparative analysis of structure-based feature descriptors using different classifiers In this study, we employed seven structure-based feature extraction methods including RSA, ASA, Q3, Q8, phi, psi, and disorder, to capture the structural features of peptides [26–29]. These features collectively provide a comprehensive assessment of peptide conformation, secondary structure, and residue-level interactions. The extracted structural information was subsequently used for predictive model construction. During data preprocessing, the structural features of each peptide were integrated into a 16-dimensional vector representing a single peptide sequence. Five classical machine learning algorithms were then applied to construct prediction models, including Support Vector Machine (SVM) [30], Random Forest (RF) [31], Logistic Regression (LR) [32], Multi-Layer Perceptron (MLP) [33], and XGBoost [34]. As shown by the experimental results (Fig. 3C), the XGBoost model outperformed the others in terms of Matthews correlation coefficient (MCC), achieving the highest value of 0.838, while the MLP model yielded the lowest MCC of 0.564. Interestingly, the LR model achieved the highest accuracy of 0.895, surpassing XGBoost (0.868), MLP (0.705), RF (0.858), and SVM (0.716). However, considering other key performance indicators such as specificity and sensitivity, the XGBoost model demonstrated more balanced and superior overall performance compared to LR model. It is important to note that some anti-aging peptides in our dataset consisted of fewer than 10 amino acids, which hindered accurate prediction of their three-dimensional structures and, consequently, compromised the reliability of the derived structural features [28]. To maintain data quality, we excluded these unreliable structural samples and proceeded with model training using only the remaining high-confidence data. This decision introduced several limitations, including 1) the reduced dataset increased the risk of model overfitting; 2) the structural features, derived from a partial dataset, could not be aligned with the full set of sequence-based features; and, more critically, 3) the performance of the structure-based models was consistently and significantly inferior to those based on sequence features across all evaluation metrics. Therefore, in subsequent analysis, we discontinued the exploration of structure-based features and focused exclusively on models constructed from sequence-derived features. The analysis of feature descriptors based on ESM model Evolutionary Scale Modeling (ESM) is an artificial intelligence-based protein modeling framework designed to learn and generate protein sequences, structures, and functions [35]. Numerous studies have demonstrated the superior performance of ESM in protein feature extraction. In this study, we employed the ESM-b1 model to extract peptide features from the AAP400 dataset. The resulting feature representations were subsequently used as input for predictive model construction based on the Support Vector Machine (SVM) algorithm. As shown in our results (Fig. 3D), the SVM model achieved an accuracy of 0.835 and a Matthews correlation coefficient (MCC) of 0.676, which were lower than those obtained using other feature descriptors, such as the DDE descriptor, which yielded an accuracy of 0.900 and an MCC of 0.810. We further compared the performance of five different machine learning algorithms using the ESM-derived features. As illustrated in Fig. S1, the differences in predictive performance among these models were relatively minor. Feature selection analysis We employed a total of 41 feature extraction methods, comprising 40 handcrafted descriptors and one representation derived from the ESM model, to independently construct predictive models. Each model produced a binary classification label (1 for anti-aging, 0 for non-anti-aging) when applied to the AAP400 dataset. The outputs of these 41 models were concatenated to form a new 41-dimensional feature vector for each sample. This vector served as a novel representation of the peptide and was subsequently used to construct a final classification model. This process effectively created a meta-feature space based on prediction outputs, capturing the diverse decision-making behaviors of different feature descriptors with respect to the classification task. To further optimize model performance, we applied the Recursive Feature Elimination (RFE) algorithm, a greedy search method, to select an optimal subset of features from the meta-representation [36]. At this stage, we developed five widely used machine learning models, including Support Vector Machine (SVM), Random Forest (RF), k-Nearest Neighbors (KNN), Logistic Regression (LR), and XGBoost. In the LR model, the optimal feature subset was determined to include four descriptors from our proposed feature representation learning system: CKSAAP (k = 2), CTriad (k = 4), DDE, and AAC + TPC (Fig. 4A). For the RF model, the optimal number of selected features was k = 31 (Fig. 4B). The optimal feature subset dimensionalities for the SVM (Fig. 4C), XGBoost (Fig. 4D), and KNN models (Fig. 4E) were determined to be k = 10, k = 8, and k = 5, respectively (Fig. 4F). Feature representation learning strategy and feature selection (Additional file 1: Table S5). A-E The optimal feature subsets for each model were determined using the recursive feature elimination (RFE) algorithm. F Visualization of the dimensionality of the optimal feature subsets selected for the predictive models Classifier comparative analysis based on the optimal features We compared the performance of machine learning models constructed using their respective optimal feature subsets. As shown in Fig. 5A, the XGBoost model achieved an accuracy of 0.963, while the LR, SVM, RF, and KNN models all attained an accuracy of 0.975. Another key performance metric, the area under the ROC curve (AUC), indicated that the XGBoost model reached a score of 0.98, whereas the KNN model had the lowest AUC at 0.96. The remaining models each achieved an AUC of 0.97. Taking into account the dimensionality of the optimal feature subsets in each model, we determined that the Logistic Regression (LR) model exhibited the best overall performance. Performance of the optimal feature descriptor (Additional file 1: Table S5). A Comparison of the predictive performance of the optimal feature vector across five different machine learning algorithms. B In the logistic regression (LR) model, the optimal feature subset was derived from four descriptors within the feature representation learning framework: DDE, AAC + TPC, CTriad (k = 4), and CKSAAP (k = 2). The importance scores of these selected features were also summarized. C Summary of the optimal feature vectors identified in the five predictive models. D Overlapping features shared among the optimal feature subsets of the five models The optimal feature subset in the LR model was composed of four descriptors derived from the proposed feature representation learning system: DDE, AAC + TPC, CTriad (k = 4), and CKSAAP (k = 2). Feature importance analysis revealed that the features derived from the DDE descriptor contributed more significantly to the model than those from the other descriptors (Fig. 5B). Notably, analysis of the optimal feature subsets across the five machine learning models indicated that CKSAAP (k = 2) was included in the selected subsets of four models, namely LR, KNN, RF, and XGBoost (Fig. 5C). Additionally, the DDE descriptor was found to be a common component of the optimal feature subset in all classification models (Fig. 5D), highlighting its critical role in the identification of anti-aging peptides. The DDE descriptor characterizes sequence-level preferences by quantifying the deviation between observed and expected frequencies of dipeptides in peptide sequences, based on a reference background database. This allows for the identification of potential functional associations within the peptide sequence. Comparison of data augmentation results with different methods and proportions Two data augmentation strategies were proposed to expand the dataset and mitigate the risk of model overfitting due to limited training data. The first strategy involved the use of a WGAN-GP model, a variant of the Generative Adversarial Network (GAN), which incorporates a gradient penalty to improve training stability and model performance [37, 38] (Fig. 6A). During training, after approximately 10,000 iterations, three key indicators, including Discriminator Loss, Generator Loss, and Wasserstein Distance, converged to near-zero values (negative for the discriminator, positive for the generator and Wasserstein distance), indicating that the model had reached a stable state (Fig. 6B). The augmented data were subsequently used to construct predictive model. The results showed that the Logistic Regression (LR), XGBoost, Multi-Layer Perceptron (MLP), and Random Forest (RF) models all achieved accuracies above 0.9, specifically 0.929, 0.946, 0.946, and 0.937, respectively (Fig. 6C). In terms of MCC, the values were 0.862 for LR, 0.898 for XGBoost, 0.895 for MLP, and 0.877 for RF. Compared to models trained without data augmentation, the average accuracy and MCC across the five models improved by 0.085 and 0.166, respectively. Among all models, XGBoost exhibited the most favorable overall performance based on a comprehensive comparison of sensitivity, specificity, accuracy, MCC, recall, and F1-score (Fig. 6D). Performance comparison of predictive models based on GAN-augmented data. A The overall architecture and training workflow of the Wasserstein GAN with Gradient Penalty (WGAN-GP). B Training loss curves of WGAN-GP, including the discriminator loss, generator loss, and the Wasserstein distance, which quantifies the divergence between the distribution of generated samples and that of real data. C Predictive models were constructed using GAN-augmented datasets. The performance of five machine learning algorithms was evaluated and compared based on metrics including sensitivity, specificity, accuracy, and Matthews correlation coefficient (MCC) (Additional file 1: Table S5). D Visualization of the XGBoost model, which achieved the best overall performance among the evaluated models (Additional file 1: Table S6) In accordance with the strategies described in References [39–41], we also implemented a second data augmentation approach based on conservative amino acid substitutions. In this method, amino acids were grouped by physicochemical properties, such as acidity and polarity, enabling biologically equivalent replacements in peptide sequences (Fig. 7A). Unlike other augmentation strategies, this method avoids insertions or deletions and preserves the overall biochemical characteristics of the sequence [39]. Additionally, it significantly altered the latent distribution of the data during deep learning modeling, as illustrated in Fig. S2. Subsequently, we constructed predictive models using both the handcrafted feature descriptors and ESM-based representations at various augmentation levels (Fig. 7B). When using the 40 handcrafted descriptors, model performance improved most noticeably at the 3 × augmentation level. As the dataset size increased further, model performance plateaued, reaching optimal values at the 9 × augmentation level, with average accuracy and MCC reaching 0.973 and 0.947, respectively. In models built using ESM-derived features, the SVM model’s accuracy increased from 0.835 to 0.914, and its MCC from 0.676 to 0.830, as the size of the augmented dataset increased. Other metrics also showed notable improvements: sensitivity rose from 0.900 to 0.944, specificity from 0.770 to 0.884, and F1-score from 0.849 to 0.916. We further compared the performance of five machine learning models including SVM, MLP, RF, XGBoost, and LR, across varying levels of data augmentation (Fig. 7B). At the 1 × augmentation level, the LR model demonstrated the best performance, with an accuracy of 0.914 and an MCC of 0.833, outperforming the other four models. However, as the degree of data augmentation increased, XGBoost began to outperform the others, achieving the highest accuracy and MCC of 0.993 and 0.986, respectively, at the 9 × augmentation level. Nevertheless, it is important to note that when using the conservatively substituted peptide dataset for traditional machine learning modeling, there remains a risk of model overfitting. Performance comparison of predictive models based on amino acid conservative substitution data augmentation. A Rules of amino acid conservative substitution. New peptide sequences were generated by substituting amino acids in the original sequences according to conservation rules, thereby achieving data augmentation. B The performance of predictive models constructed under different levels of augmentation was compared. Additionally, models using ESM-based feature extraction were compared with those using handcrafted features derived from 40 traditional descriptors (Additional file 1: Table S7) Performance analysis of classifier constructed based on deep learning Deep learning models typically require large-scale datasets for effective training [42, 43]. Therefore, we utilized the 10 × augmented dataset as input to train a deep learning model. The model architecture consisted of an embedding layer based on ESM-derived sequence features, followed by two convolutional neural network (CNN) layers and a multi-head attention mechanism [44]. During model training, the number of epochs was set to 5000, and dropout as well as L2 regularization were applied to mitigate overfitting. According to the results of fivefold cross-validation, the accuracy curve exhibited a rapid increase in the early training stages and began to stabilize after approximately 2000 epochs. Eventually, the validation accuracy converged to above 90%, indicating successful model convergence (Fig. 8A-E). The final trained model achieved an AUC of 0.96, demonstrating strong predictive performance (Fig. 8F). Additionally, we explored an alternative embedding strategy by integrating features derived from DDE, AAC_TPC, CKSAAP (k = 2), and CTriad (k = 4). A deep learning model was constructed based on this fused feature embedding. However, the experimental results indicated that the final model's accuracy remained below 80%, with an AUC of 0.90 (Fig. S3), which was inferior to the performance of the model using ESM-based representations. Construction of deep learning models based on data augmentation. A-E These subplots A-E show the validation accuracy across training epochs for each fold during cross-validation. The accuracy tends to stabilize after approximately 2000 epochs, ultimately converging to a validation accuracy above 90% in all folds. F The overall receiver operating characteristic (ROC) curve, which evaluates the model’s classification performance across various thresholds. The area under the curve (AUC) is 0.96, indicating excellent overall discriminative ability of the model Prediction performance comparison with other models In this study, we compared the performance of the optimal feature vector-based model with four individual feature descriptor-based models, as well as its predictive performance against the ESM_9x, ESM_GAN, and CNN_9x models on an independent test dataset. The results of fivefold cross-validation on both the AAP400 dataset and the independent test set are summarized in Table S8. As shown (Fig. 9A), the optimal feature vector model outperformed all four descriptor-based models across all performance metrics, including accuracy (ACC), sensitivity (SE), specificity (SP), Matthews correlation coefficient (MCC), recall, and F1-score. Specifically, the optimal model achieved the highest ACC and MCC values of 99.7% and 0.995, respectively, which were significantly higher than those of CKSAAP (k = 2) (ACC = 88%, MCC = 0.77), CTriad (k = 4) (ACC = 84%, MCC = 0.68), AAC + TPC (ACC = 86%, MCC = 0.72), and DDE (ACC = 90%, MCC = 0.81). As illustrated in Fig. 9A, the area under the ROC curve (AUC), a commonly used metric for evaluating overall classification performance, was also higher for the optimal feature vector model, indicating superior predictive ability. This suggests that the optimal features provide more accurate discrimination between true anti-aging and non-anti-aging peptides than their original descriptors. This improvement may be attributed to the feature representation learning framework, which integrates the outputs of multiple heterogeneous models to reconstruct features, thus implicitly achieving feature fusion and enhancement. Similarly, results on the independent test set confirmed that the optimal feature descriptor yielded better performance than the individual descriptors, achieving an AUC of 0.99 (Fig. 9B). Importantly, the dimensionality of the optimal feature vector was only 4, significantly lower than that of the original descriptors (Fig. 9C). Additionally, as shown in Table S10, the learned 4-dimensional feature vector exhibited superior performance across five out of six major evaluation metrics, including SN, MCC, recall, F1-score, and ACC, compared to the original descriptors, with SP being the only exception. Performance and comparison of prediction models on AAP400 and independent test datasets (Additional file 1: Table S8). A Comparing the performance of the optimal feature vector with the original four descriptors using fivefold cross-validation on the AAP400 dataset. B Comparing the performance of the optimal feature vector with the original four descriptors using fivefold cross-validation on the test dataset. C The dimensions of the optimal feature vector and the original four descriptors. D The area under the receiver operating characteristic curve (AUC) was employed as the primary evaluation metric to systematically compare the performance of four models, including Antiaging-FL, ESM_9x, ESM_GAN, and CNN_9x, in the task of anti-aging peptide identification. A higher AUC value indicates a stronger discriminative ability of the model in distinguishing between positive and negative samples We further compared the performance of the optimal feature vector-based model with that of the ESM_9x, ESM_GAN, and CNN_9x models on the independent test dataset. The ESM_9x model was built using a 9 × augmented dataset generated by the conservative amino acid substitution method, with ESM features used for peptide representation. The ESM_GAN model was constructed from a GAN-augmented dataset (2 × the original size), also using ESM-based features. The CNN_9x model employed the same 9 × amino acid substitution-based augmented dataset, using ESM features as input to a deep learning model with an embedding layer. The results showed that the Antiaging-FL model achieved the highest AUC score of 0.99, outperforming ESM_9x (0.95), ESM_GAN (0.95), and CNN_9x (0.94) (Fig. 9D). These findings indicate that the Antiaging-FL model possesses superior predictive capability compared to other constructed models. However, it is worth noting that the Antiaging-FL model was trained on a relatively small dataset, which may lead to overfitting. Therefore, we retained both the ESM_GAN and CNN_9x models as additional reference models for users. Prediction of potential anti-aging peptide using public data Peptides intervene in the aging process through various mechanisms, including anti-cancer, antimicrobial, anti-inflammatory, antioxidant, and immunomodulatory pathways [45–47]. Therefore, we aimed to investigate whether peptides with functions such as anticancer and antimicrobial properties also possess anti-aging capabilities. We collected peptides with corresponding functions, including antimicrobial peptides, anticancer peptides, IL-13 inducing peptides, antimalarial peptides, antioxidant peptides, and anti-inflammatory peptides, to analyze their potential anti-aging function. The antimicrobial peptides were primarily sourced from the Antimicrobial Peptide Database, which includes 3146 natural antimicrobial peptides [48]. We analyzed them using the constructed prediction model, revealing that 23% (729/3167) of the antimicrobial peptides were predicted to be potential anti-aging peptides (Fig. 10A). This is understandable, as antimicrobial peptides possess broad-spectrum antibacterial activity, capable of combating pathogenic microorganisms by not only directly killing bacteria but also modulating the host’s immune response. As aging progresses, the immune system’s functionality declines, increasing infection likelihood. Antimicrobial peptides may exert anti-aging effects by enhancing the immune defense system. Additionally, some antimicrobial peptides are multifunctional, possessing anti-inflammatory and antioxidant properties, thereby potentially slowing aging, ameliorating various age-related diseases. IL-13 inducing peptides generally function within the immune system. We downloaded IL-13 inducing peptide sequences from the iIL13Pred web server [49] and input them into the model for prediction, showing that 32.9% (103/313) of the sequences were classified as anti-aging peptides (Fig. 10B). Similarly, we obtained anticancer peptides from the ACPred-FL web server [50], antimalarial peptides, antioxidant peptides, and anti-inflammatory peptides from the Antimicrobial Peptide Database, and analyzed them using the prediction model. The results indicated that 23.4% (88/376) of anticancer peptides (Fig. 10C), 8.8% (3/34) of antimalarial peptides (Fig. 10D), 14.7% (5/34) of antioxidant peptides (Fig. 10E), and 29.7% (11/37) of anti-inflammatory peptides (Fig. 10F) demonstrated potential anti-aging properties. This information highlights the biological diversity of peptides, enabling them to perform multiple roles in various physiological and pathological contexts. Prediction of potential anti-aging peptide (Additional file 1: Table S9). A-F We analyzed the potential anti-aging function of antimicrobial peptides, anticancer peptides, IL-13 inducing peptides, antimalarial peptides, antioxidant peptides, and anti-inflammatory peptides Discussion Anti-aging peptides have shown potential in delaying aging, enhancing neurological function, and boosting immune responses, making them highly promising in medical applications [51]. However, the development and application of anti-aging peptides face significant challenges. Bioinformatics, machine learning, and molecular dynamics modeling provide new avenues for addressing these challenges. In this study, we successfully provided a benchmark dataset and developed these models to predict anti-aging peptides. Prior to this research, the lack of a non-anti-aging peptide dataset hindered the establishment of predictive models. By analyzing the biological functions and mechanisms of anti-aging peptides, we selected pro-aging peptides and peptides unrelated to aging as the negative dataset. This strategy for negative dataset construction is worth exploring. Nonetheless, potential issues remain, for peptides considered unrelated to aging, we cannot confirm with certainty that they are truly unrelated in the objective biological context, and this requires further experimental validation. Another challenge is the dataset size. After filtering with CD-HIT, only 220 anti-aging peptide sequences remained, matched with 220 non-anti-aging peptide sequences, with 40 sequences assigned to the independent test set. Overfitting is inevitable during model training due to the small sample size. Thus, data augmentation strategies were employed to mitigate overfitting [52]. In this study, we used two data augmentation methods including GAN-based data augmentation and amino acid conservative substitution. The GAN-based approach increased the dataset size from 400 to 800. However, further attempts to enlarge the dataset using GAN failed to converge, suggesting limitations of this method. The amino acid substitution strategy allowed augmentation from 1 × to 9x, increasing the dataset size from 800 to 4000. Notably, when using traditional machine learning algorithms, feature distributions did not change significantly, indicating limited improvement in data diversity. In contrast, deep learning models showed substantial changes in feature distribution, resulting in more robust models. In addition, we applied five different machine learning algorithms to compare 41 feature descriptors (including nine sequence-based encoding methods and the ESM model) for anti-aging peptide prediction. Among the 41 machine learning models, the model trained with DDE features and an SVM classifier achieved the best performance. To build an efficient and accurate model, we developed a feature representation learning and selection scheme to extract the most discriminative features. Results on the AAP400 and test datasets demonstrated that this method outperformed existing sequence feature descriptors. In this study, we initially explored the incorporation of structure-based features to enhance peptide prediction performance. Specifically, seven structure-derived descriptors including RSA, ASA, Q3, Q8, phi, psi, and disorder, were applied to capture peptide conformational and secondary structure characteristics. These features were integrated into a 16-dimensional representation and used to train predictive models with five classical machine learning algorithms, including SVM, RF, LR, MLP, and XGBoost. However, we observed that a substantial proportion of anti-aging peptides in our dataset contained fewer than ten amino acids, leading to inaccurate or unreliable three-dimensional structure predictions. As a result, the corresponding structural features were of low confidence and could not reliably represent peptide properties. To ensure data quality, we excluded these unreliable samples and trained models only on the remaining high-confidence data. This approach, however, introduced several limitations: the reduced dataset increased the risk of overfitting, and the structure-based models consistently underperformed compared to those based on sequence features across all evaluation metrics. Given that short peptide sequences often yield inaccurate structural predictions, which undermine the reliability of extracted features, we ultimately decided not to incorporate structural features into the final model training and focused exclusively on sequence-derived representations in subsequent analyses. Additionally, omics data related to aging phenotypes could serve as a valuable resource for identifying anti-aging peptides in future studies [53]. Peptides possess diverse biological functions. We investigated whether antimicrobial, anticancer, anti-inflammatory, and antioxidant peptides exhibit anti-aging properties. Our results showed that 23.4% (88/376) of anticancer peptides were predicted as anti-aging peptides, possibly because senescent cells can transform into cancer cells, and anticancer peptides may eliminate these precancerous cells or improve the cellular environment. Similarly, 23% (729/3167) of antimicrobial peptides were predicted to have anti-aging potential, likely by modulating host immune responses. 29.7% (11/37) of anti-inflammatory peptides also exhibited anti-aging properties, consistent with the known role of inflammation in aging and age-related diseases. Additionally, Oxidative stress has long been proposed as a key contributor to aging, though its precise causal role remains debated [54], with evidence supporting both damaging and signaling (hormetic) functions of ROS [55]. In this study, 14.7% (5/34) of antioxidant peptides showed anti-aging potential by reducing oxidative stress and enhancing cellular vitality. Among IL-13-inducing peptides, 32.9% (103/313) were classified as anti-aging, potentially through immune enhancement. These analysis suggest that peptides intervene in the aging process through multiple mechanisms, including anticancer, antimicrobial, anti-inflammatory, antioxidant, and immunomodulatory pathways. Different types of peptides may have overlapping or synergistic functions in mitigating or reversing aging-related physiological changes. Further research is needed to elucidate the specific roles of peptides in complex biological systems. Conclusions In conclusion, our study presents three high-performance models for AAP prediction, accelerating the discovery of novel anti-aging peptides. These models not only provide a valuable resource for researchers but also offer insights into the functional mechanisms of AAPs, paving the way for targeted drug development in aging-related therapeutics. Methods Framework for dataset construction, feature selection, and predictive model development In this study, we proposed a comprehensive framework for anti-aging peptide (AAP) dataset construction and augmentation, feature selection, and predictive model development. First, we constructed the AAP400 dataset for model training and evaluation. It must be emphasized that these positive samples are derived from the AgingBase database [56], which has an undeniable contribution to the construction of the dataset in this research, and the negative samples included pro-inflammatory peptides, which may promote aging, as well as cytotoxic peptides and precursor peptide, both of which are considered unrelated to aging. During the model training phase, we implemented a feature representation learning strategy to extract informative feature from 40 sequence-based descriptors, including amino acid composition (AAC), dipeptide composition (DPC), tripeptide composition (TPC), composition-transition-distribution (CTD), composition of triads (CTriad), composition of k-spaced amino acid pairs (CKSAAP), position-specific scoring matrix (PSSM), distance-based descriptors (DDE), pseudo-amino acid composition (PAAC), and high-dimensional embeddings derived from the ESM model. These outputs were integrated into a novel 41-dimensional feature representation vector for each peptide. To obtain an optimal subset of features, we employed the recursive feature elimination (RFE) strategy. During feature selection, five widely used machine learning algorithms were utilized to construct predictive models: k-nearest neighbors (KNN), support vector machine (SVM), random forest (RF), XGBoost, and logistic regression (LR). To further enhance data diversity and improve model robustness, we adopted two data augmentation strategies, including 1) a GAN-based data augmentation approach, and 2) a conservative amino acid substitution-based augmentation method. For the former, predictive models were developed using GAN-augmented datasets and ESM-derived features via traditional machine learning algorithms. For the latter, a deep learning model was constructed using the substitution-augmented dataset. This architecture included an ESM embedding layer, followed by convolutional neural network (CNN) layers for local feature extraction and a multi-head attention mechanism to capture global contextual dependencies. Finally, we compared the performance of all constructed models and selected the top three most effective models for downstream application and public use by researchers. Construction of datasets In this study, we provided a benchmark dataset for the construction of predictive models for anti-aging peptides (AAPs). The positive dataset was derived from AagingBase [56], consisting of 282 experimentally validated anti-aging peptides. To ensure model reliability and mitigate overfitting, we applied strict constraints on peptide sequence length and similarity, removing sequences shorter than four residues and those with pairwise identity greater than 90%. As a result, 220 high-quality anti-aging peptide sequences were retained for further analysis. Obtaining a reliable negative dataset proved to be more challenging. Searches in PubMed and UniProt [57] yielded limited information on experimentally confirmed non-anti-aging peptides, highlighting the relative lack of research on anti-aging peptides compared to antimicrobial and signal peptides [58, 59]. To ensure the biological validity of the negative dataset, we deliberately avoided the commonly used strategy of randomly generating synthetic peptides based on homology. Additionally, peptides retrieved from databases such as UniProt and PeptideDB (http://www.peptides.be/) with non-anti-aging annotations, such as antibacterial or signal peptides, were excluded due to literature evidence suggesting they may indirectly influence aging processes [60–62]. To construct a biologically relevant and reliable negative dataset, we prioritized peptides that exhibit functions opposite to those of anti-aging peptides. For instance, previous studies have reported that the secretion of senescence-associated secretory phenotype (SASP) factors can induce cellular senescence [63, 64], implying that pro-inflammatory peptides promoting the sustained release of pro-inflammatory cytokines such as TNF-α and IL-6 may contribute to aging. This rationale supports our selection of pro-inflammatory peptides as the primary source for the negative class. Furthermore, we included cytotoxic peptides and precursor peptide in the negative dataset, as there is currently no evidence in the literature indicating their involvement in aging processes. These peptides were therefore assumed to be unrelated to aging and appropriate for inclusion in the negative class. Following this strategy, we constructed a benchmark dataset comprising 440 peptide sequences. A stratified random sampling approach was then employed to select 40 peptide sequences for use as an independent test set. The remaining 400 sequences, referred to as the AAP400 dataset, were used for training and model development. Feature representation Adopting efficient feature extraction methods is a key step in building high-performance predictors. A comprehensive review of existing sequence-based feature-encoding algorithms for peptide feature representation was conducted, including AAC, which reflects the overall compositional features of the peptide and provides basic insights into properties such as solubility, structural stability, and hydrophobicity/hydrophilicity, DPC, which captures local sequence information and interactions between adjacent residues, aiding in the identification of structural conformations and functional motifs, TPC, which offers higher-order local sequence structure information compared to DPC and enhances the ability to recognize functional domains and structural patterns, CTD, which represents the trends and spatial distribution patterns of physicochemical properties within the sequence, CTriad, which captures the combinatorial patterns of amino acid triplets based on physicochemical classifications, CKSAAP, which reflects the long-range spatial and structural dependencies between amino acid residues, thereby aiding in the prediction of tertiary structures and functional sites, PSSM, which provides evolutionary conservation information and highlights conserved regions and potential functional domains, playing a crucial role in functional site identification and protein family classification, DDE, which integrates genetic code bias and sequence composition to reflect evolutionary preferences, expression regulation, and functional constraints and PAAC, which incorporates both compositional and sequence-order information, offering a more holistic representation of peptide structure and function. Amino Acid Composition (AAC) To include the compositional information, we computed the frequencies and the occurrences of 20 different amino acids in the peptide sequence, respectively. The feature vector for the AAC descriptor is represented below: where Va denotes the value of the amino acid type a, and k(k = 1,2) denotes the method type; for example, 1 denotes the frequency while 2 denotes the occurrence number. Composition of k-spaced Amino Acid Pairs (CKSAAP) The CKSAAP feature encoding calculates the frequency of amino acid pairs separated by any k residues (k = 0, 1, 2, …, 5. The default maximum value of k is 5). Taking k = 0 as an example, there are 400 0-spaced residue pairs (i.e., AA, AC, AD,…, YY.). Then, a feature vector can be defined as: The value of each descriptor denotes the composition of the corresponding residue pair in the protein or peptide sequence. For instance, if the residue pair AA appears m times in the protein, the composition of the residue pair AA is equal to m divided by the total number of 0-spaced residue pairs (Ntotal) in the protein. For k = 0, 1, 2, 3, 4 and 5, the value of Ntotal is P – 1, P – 2, P – 3, P – 4, P – 5 and P – 6 for a protein of length P, respectively. Tri-Peptide Composition (TPC) The Tripeptide Composition (TPC) gives 8000 descriptors, defined as: where Nrst is the number of tripeptides represented by amino acid types r, s and t. Dipeptide Deviation from Expected Mean (DDE) Where Nrs is the number of dipeptides represented by amino acid types r and s. The Dipeptide Deviation from Expected Mean feature vector is constructed by computing three parameters, i.e. dipeptide composition (Dc), theoretical mean (Tm), and theoretical variance (Tv). The above three parameters and the DDE are computed as follows. Dc(r,s), the dipeptide composition measure for the dipeptide ‘rs’, is given as where Nrs is the number of dipeptides represented by amino acid types r and s and N is the length of the protein or peptide. Tm(r,s), the theoretical mean, is given by: where Cr is the number of codons that code for the first amino acid and Cs is the number of codons that code for the second amino acid in the given dipeptide ‘rs’. CN is the total number of possible codons, excluding the three stop codons (i.e., 61). Tv (r,s), the theoretical variance of the dipeptide ‘rs’, is given by: Finally, DDE(r,s) is calculated as: Conjoint Triad (CTriad) The Conjoint Triad descriptor (CTriad) considers the properties of one amino acid and its vicinal amino acids by regarding any three continuous amino acids as a single unit. First, the protein sequence is represented by a binary space (V, F), where V denotes the vector space of the sequence features, and each feature (Vi) represents a sort of triad type; F is the number vector corresponding to V, where fi, the value of the i-th dimension of F, is the number of type Vi appearing in the protein sequence. For the amino acids that have been catalogued into seven classes, the size of V should be equal to 7ⅹ7ⅹ7 = 343. Accordingly, i = 1, 2, 3, …, 343. In principle, the longer a protein sequence, the higher the probability to have larger values of fi, confounding the comparison of proteins with different lengths. Thus, we define a new parameter, di, by normalizing fi with the following equation: Di-Peptide Composition (DPC) The Dipeptide Composition gives 400 descriptors. It is defined as: Composition/Transition/Distribution (CTD) The Composition, Transition and Distribution (CTD) features represent the amino acid distribution patterns of a specific structural or physicochemical property in a protein or peptide sequence. 13 types of physicochemical properties have been previously used for computing these features. These include hydrophobicity, normalized Van der Waals Volume, polarity, polarizability, charge, secondary structures and solvent accessibility. These descriptors are calculated according to the following procedures: (i) The sequence of amino acids is transformed into a sequence of certain structural or physicochemical properties of residues; (ii) Twenty amino acids are divided into three groups for each of the seven different physicochemical attributes based on the main clusters of the amino acid indices of Tomii and Kanehisa. Pseudo-Amino Acid Composition (PAAC) This group of descriptors has been proposed [65, 66]. Let H1o(i), H2o(i),Moi for i = 1, 2, 3, … 20 be the original hydrophobicity values, the original hydrophilicity values and the original side chain masses of the 20 natural amino acids, respectively. They are converted to the following quantities by a standard conversion: where H2o(i) and Mo(i) are normalized as H2(i) and M(i) in the same manner. Next, a correlation function can be defined as: This correlation function is actually an averaged value for the three amino acid properties: hydrophobicity value, hydrophilicity value and side chain mass. Therefore, we can extend this definition of correlation function for one amino acid property or for a set of n amino acid properties. For one amino acid property, the correlation can be defined as: where H(Ri) is the amino acid property of amino acid Ri after standardization. For a set of n amino acid properties, it can be defined as: where Hk(Ri) is the k-th property in the amino acid property set for amino acid Ri. A set of descriptors called sequence order-correlated factors are defined as: where λ (λ < N) is an integer parameter to be chosen. Let fi be the normalized occurrence frequency of amino acid i in the protein sequence. Then, a set of 20 + λ descriptors called the pseudo-amino acid composition for a protein sequence can be defines as: where w is the weighting factor for the sequence-order effect and is set to w = 0.05 in antiaging-FL. PSSM profile (PSSM) This feature descriptor is extracted from the Position-Specific Scoring Matrix (PSSM) profile. The PSSM profile can be obtained by running PSI-BLAST against the uniref 50 database [67]. The PSSM descriptor is usually applied to encode the peptides with equal length. Each amino acid in the peptide is represented by a 20-dimensional vector. Enhanced Amino Acid Composition (EAAC) The Enhanced Amino Acid Composition (EAAC) feature type calculates the AAC based on the sequence window of fixed length (the default value is 5) that continuously slides from the N- to C-terminus of each peptide and can be usually applied to encode the peptides with an equal length. Grouped Amino Acid Composition (GAAC) In the GAAC encoding, the 20 amino acid types are further categorized into five classes according to their physicochemical properties, e.g. hydrophobicity, charge and molecular size [68]. The five classes include the aliphatic group (g1: GAVLMI), aromatic group (g2: FYW), positive charge group (g3: KRH), negative charged group (g4: DE) and uncharged group (g5: STCPNQ). GAAC descriptor is the frequency of each amino acid group, which is defined as: where N(g) is the number of amino acid in group g, N(t) is the number of amino acid type t, and N is the length of the protein/peptide sequence. Enhanced GAAC (EGAAC) The Enhanced GAAC (EGAAC) is also for the first time proposed in this work. It calculates GAAC in windows of fixed length (default is 5) continuously sliding from the N- to C-terminal of each peptide and is usually applied to peptides with an equal length. where N(g, win) is the number of amino acids in group g within the sliding window win and N(win) is the size of the sliding window win. Grouped Di-Peptide Composition (GDPC) The Grouped Di-Peptide Composition encoding is another variation of the DPC descriptor. It is composed of a total of 25 descriptors that are defined as: where Nrs is the number of tripeptides represented by amino acid type groups r and s, N is the length of a protein or peptide sequence. Grouped Tri-Peptide Composition (GTPC) The Grouped Tri-Peptide Composition encoding is also a variation of TPC descriptor, which generates 125 descriptors, defined as: where Nrst is the number of tripeptides represented by amino acid type groups r, s and t. N is the length of a protein or peptide sequence. CTDC Taking the hydrophobicity attribute as an example, all amino acids are divided into three groups: polar, neutral and hydrophobic. The Composition descriptor consists of three values: the global compositions (percentage) of polar, neutral and hydrophobic residues of the protein. The Composition descriptor can be calculated as follows: where N(r) is the number of amino acid type r in the encoded sequence and N is the length of the sequence. CTDT The Transition descriptor T also consists of three values [69, 70]: A transition from the polar group to the neutral group is the percentage frequency with which a polar residue is followed by a neutral residue or a neutral residue by a polar residue. Transitions between the neutral group and the hydrophobic group and those between the hydrophobic group and the polar group are defined in a similar way. The transition descriptor can then be calculated as: where N(r,s) and N(s,r) are the numbers of dipeptides encoded as “rs” and “sr” respectively in the sequence, while N is the length of the sequence. CTDD The Distribution descriptor consists of five values for each of the three groups (polar, neutral and hydrophobic) (Dubchak, et al., 1995; Dubchak, et al., 1999), namely the corresponding fraction of the entire sequence, where the first residue of a given group is located, and where 25, 50, 75 and 100% of occurrences are contained. For example, we start with the first residue up to and including the residue that marks 25/50/75/100% of occurrences for residues of any given group and then we simply divide the position of this residue by the length of the entire sequence. AAC + DPC This descriptor is a combination of AAC and the extension of dpc, measuring AAC and DPC information. In this descriptor, there are a total of 420 features. The first 20 features are the occurrence frequencies of 20 different amino acids extracted from AAC. The other 400 features are the dipeptide composition. AAC + TPC This descriptor is a combination of AAC and the extension of TPC, measuring AAC and TPC information. Structure feature extraction In this study, we applied seven structure-based feature extraction methods, including RSA, ASA, q3, q8, phi, psi, and disorder, to derive structural features from peptide sequences. Accessible Surface Area (ASA) refers to the actual solvent-accessible surface area of a residue, which directly quantifies the degree of solvent exposure for each amino acid. Relative Solvent Accessibility (RSA) is defined as the ratio of a residue's ASA to its maximum possible ASA value, enabling the identification of surface-exposed residues. The q3 feature simplifies secondary structure classification into three categories: α-helix, β-strand, and coil, facilitating rapid assessment of global folding patterns in peptides. In contrast, the q8 feature provides a more granular classification of secondary structures. The backbone dihedral angles, phi (φ) and psi (ψ), represent the torsion angles around the N–Cα and Cα–C bonds, respectively. Specifically, the φ angle describes the rotation between the C atom of the preceding residue and the Cα atom, while the ψ angle describes the rotation between the Cα atom and the N atom of the following residue. The disorder feature represents the intrinsic disorder tendency of residues, allowing the identification of disordered regions within peptide structures. Together, these structural descriptors provide a comprehensive assessment of peptide conformations, secondary structure organization, and residue-level interactions. ESM model for feature extraction ESM (Evolutionary Scale Modeling) is a Transformer-based protein language model developed by Meta AI, capable of extracting evolutionary information, structural features, and contextual semantics directly from amino acid sequences. In this study, we employed the ESM-1b model to extract high-dimensional features (1024-dimensional vectors) encoded by the Transformer from peptide sequences. These extracted features were subsequently used for the construction of predictive models. Feature selection Feature selection is a crucial step in constructing a robust machine learning model. In this study, we employed a greedy search algorithm called Recursive Feature Elimination (RFE). By iteratively eliminating irrelevant or redundant features and retaining the most pertinent ones, RFE enhances the model’s generalization capability and performance. The essence of the RFE algorithm lies in repeatedly training the model and filtering features to identify the optimal subset. Data augmentation Data augmentation is a technique that generates diverse new samples through supervised transformations of the original training data, thereby enhancing model generalizability. In this study, we employed a data augmentation strategy based on a Generative Adversarial Network (GAN), a generative deep learning framework comprising two components: a generator and a discriminator. The generator, composed of five fully connected layers, is responsible for producing synthetic peptide samples, while the discriminator, consisting of four fully connected layers, is trained to distinguish between real and synthetic inputs. During the adversarial training process, both the generator and discriminator iteratively improve their capabilities. Eventually, the generator becomes proficient enough to produce realistic samples that the discriminator can no longer reliably distinguish from real ones. Notably, during training, the discriminator is updated five times for every single update of the generator to maintain training balance. Both networks were optimized using the Adam optimizer with a learning rate of 1e-5. In addition, we adopted a data augmentation strategy based on conservative amino acid substitutions. In this method, amino acids are grouped according to physicochemical properties such as polarity and charge, allowing for biologically equivalent substitutions within peptide sequences. Unlike other augmentation techniques, this approach does not involve amino acid insertions or deletions and preserves the original physicochemical nature of the sequence. Machine learning algorithms The K-nearest neighbor (KNN) algorithm is a classic classification algorithm based on distance measurement. The support vector machine (SVM) algorithm is a linear classification method based on maximum margin classification. Logistic regression (LR) algorithm models the relationship between input and output variables to predict the value of output variable. Random forest (RF) is an ensemble of multiple decision trees, each trained on randomly selected features and data subsets. XGBoost is an efficient and scalable Gradient Boosting Decision Tree (GBDT) algorithm, which is a type of boosting method that progressively enhances the overall performance of the model by continuously constructing new weak classifiers. In this study, the implementation of the KNN, RF, SVM and LR were conducted by the scikit-learn library, while XGBoost is implemented by the xgboost library. We conducted the grid search strategy to optimize the parameter based on fivefold cross-validation. Attention mechanism and convolutional neural network (CNN) The attention mechanism is a fundamental technique in deep learning that enables models to selectively focus on different parts of an input sequence, assigning varying weights to each component to highlight information that is more relevant to the task at hand. Multi-head attention extends this concept by computing multiple self-attention heads in parallel, allowing the model to simultaneously capture dependencies at different positions and thereby enhancing its representational capacity. Convolutional Neural Networks (CNNs) are widely used deep learning architectures known for their ability to automatically learn spatial or temporal patterns through local receptive fields, weight sharing, and hierarchical feature extraction. In this study, we implemented a predictive model that integrates CNN with a multi-head attention mechanism. The model architecture comprises one embedding layer, two convolutional layers, and one multi-head attention layer. During the training phase, the model was trained for 5000 epochs with a learning rate of 1e-5 and a batch size of 32. Additionally, we employed a binary cross-entropy loss function combined with a sigmoid activation function to optimize the model. Performance measurements In this work, we employed six metrics to evaluate the performance of predictive models, including Sensitivity (SE), Specificity (SP), Accuracy (ACC), Precision, Matthews Correlation Coefficient (MCC), and F1-score. These metrics are calculated using the following formulas: where TP, TN, FP, and FN denote the counts of true positives, true negatives, false positives, and false negatives, respectively. SE measures the ability to predict positives, SP measures the ability to predict negatives, and Precision calculates the proportion of true positives among all positive predictions. The F1-score is the harmonic mean of Precision and Sensitivity. ACC and MCC are utilized to assess the overall performance of the predictive model. Receiver operating characteristic curve The receiver operating characteristic (ROC) curve is used to evaluate the overall performance of predictive model [71]. It is plotted by calculating sensitivity against 1-specificity under different thresholds. The area under the ROC curve, termed as AUC, is often used as a metric to access the overall performance. The AUC score ranges from 0.5 to 1. The larger score a predictor achieves, the better performance it has. Supplementary Information Additional file 1. Table S1-S10. Table S1: comprehensive details of each anti-aging peptide. Table S2: The length of anti-aging and non-anti-aging peptides. Table S3: The model performance of 40 feature descriptors on the AAP400 dataset. Table S4: Comparison of model performance based on structural features of peptides and ESM-1b model. Table S5: Sources of images in Fig. 5. Table S6: The performance of predictive models constructed using GAN-augmented datasets. Table S7: Sources of images in Fig. 7. Table S8: Sources of images in Fig. 7. Table S8: Performance of antiaging-FL predictor on AAP400 and test datasets. Table S9: Sources of images in Fig. 10. Table S10: Sources of images in Fig. S1. Additional file 2. Figures S1-S3. Fig. S1: The performance of five different machine learning algorithms using the ESM-derived features. Fig. S2: The latent distribution of the data during deep learning modeling from different magnitudes of data augmentation. Fig. S3: Construction of deep learning models based on data augmentation and alternative embedding strategy by integrating features derived from DDE, AAC_TPC, CKSAAP (k = 2), and CTriad (k = 4). Acknowledgements We greatly thank Dr. Jianwen Chen (Ohio State University, USA) and Dr. Wei-hua Chen (Huazhong University of S&T) for critical review of the manuscript. We also greatly thank Dr. Qihao Zheng (Shanghai Artificial Intelligence Laboratory) for advice of predictive model. Abbreviations Anti-aging peptides Generative adversarial network Convolutional neural network Receiver operating characteristic Area Under Curve Amino Acid Composition Grouped Amino Acid Composition Enhanced Amino Acid Composition Di-Peptide Composition Tri-Peptide Composition Composition/Transition/Distribution Conjoint Triad Composition of k-spaced Amino Acid Pairs Position-Specific Scoring Matrix Dipeptide Deviation from Expected Mean Pseudo-Amino Acid Composition Support Vector Machine Matthews correlation coefficient Accuracy Sensitivity Specificity Relative Solvent Accessibility Accessible Surface Area Random Forest Logistic Regression Multi-Layer Perceptron K-Nearest Neighbors Evolutionary Scale Modeling Recursive Feature Elimination Authors’ contributions Z.Z. conceived and designed the study. Z.Z. and Y.C. implemented the pipeline, constructed the predictor, developed the codes, and performed all experiments. Z.Z., Y.C., S.W., G.C., M.W., Y.P, and E.L. performed data analyses. Z.Z., Y.C., and G.C. discussed and interpreted all results. Z.Z., Y.C., and E.L. wrote and critically revised the manuscript. All authors read and approved the final manuscript. Funding The work was supported by grant from the Wannan Medical College Key Project Research Fund (WK2023ZZD08), and Shandong Province Science Foundation for Youths (ZR2024QC312). Data availability All data generated or analyzed during this study are included in this published article and its supplementary information file. All of the codes and materials are freely available at GitHub (https://github.com/zhangzhiyuan257/anti_aging) and Zenodo (10.5281/zenodo.17470123). Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Additional file 1. Table S1-S10. Table S1: comprehensive details of each anti-aging peptide. Table S2: The length of anti-aging and non-anti-aging peptides. Table S3: The model performance of 40 feature descriptors on the AAP400 dataset. Table S4: Comparison of model performance based on structural features of peptides and ESM-1b model. Table S5: Sources of images in Fig. 5. Table S6: The performance of predictive models constructed using GAN-augmented datasets. Table S7: Sources of images in Fig. 7. Table S8: Sources of images in Fig. 7. Table S8: Performance of antiaging-FL predictor on AAP400 and test datasets. Table S9: Sources of images in Fig. 10. Table S10: Sources of images in Fig. S1. Additional file 2. Figures S1-S3. Fig. S1: The performance of five different machine learning algorithms using the ESM-derived features. Fig. S2: The latent distribution of the data during deep learning modeling from different magnitudes of data augmentation. Fig. S3: Construction of deep learning models based on data augmentation and alternative embedding strategy by integrating features derived from DDE, AAC_TPC, CKSAAP (k = 2), and CTriad (k = 4). Data Availability Statement All data generated or analyzed during this study are included in this published article and its supplementary information file. All of the codes and materials are freely available at GitHub (https://github.com/zhangzhiyuan257/anti_aging) and Zenodo (10.5281/zenodo.17470123).