Analysis of lipid–protein interactive oxidation and underlying pathways in aquatic products via multi-omics and machine learning: a case study of air-dried hairtail Authors: Yi Luo, Hui-Min Lin, Shucheng Liu, Soottawat Benjakul, Bin Zhang Published: 2026 ABSTRACT AbstractThe interactive oxidation mechanisms between lipids and proteins during aquatic product processing remain poorly understood, limiting quality control strategies. This investigation employed a comprehensive approach by integrating multi-omics data with seven machine learning models to elucidate the mechanisms of lipid and protein interactive oxidation in air-drying hairtail. The optimal XGBoost model achieved 0.98 AUC with 0.90 accuracy on test set. Machine learning combined with multi-omics has identified six oxidative biomarkers of lipid-protein interactions: receptor expression-enhancing protein 6-like, glutathione-S-transferase 1-like, peroxisome biogenesis factor 19, protein-glutamine gamma-glutamyltransferase 2-like isoform X2, dMePE (18:4/16:0), and PE (20:1/18:1). The analysis of the underlying pathways associated with these oxidative biomarkers revealed their involvement primarily in ferroptosis and the protein kinase A pathway. The lipid-protein interaction oxidation process in air-dried hairtail is facilitated by a TG2-crosslinked gel network. This study offers novel molecular targets and theoretical insights for developing targeted antioxidant interventions during processing.Keywords: Multi-omics, Machine learning, Oxidative biomarkers, Lipid oxidation, Protein oxidation Introduction The hairtail (Trichiurus lepturus), also known as cutlass fish, is widely distributed in the eastern Pacific Ocean and almost all coastal regions of China (Koddy et al., 2021). Hairtail, as one of the most valuable marine fish species, is rich in polyunsaturated fatty acids (PUFAs) and essential amino acids (Fang et al., 2022). Significantly, as a deep-sea species, hairtail is susceptible to spoilage due to its own high protein content, water content, and susceptibility to microbial contamination. Air-drying is recognized as the most cost-effective preservation method for aquatic products (Faisal et al., 2024). Thus, the conventional air-drying process is commonly used in coastal areas to extend the shelf life of hairtail, enhance nutritional value, and improve its flavor profile. As a representative processing method in coastal regions, traditional air-drying causes slight lipid and protein oxidation during the drying process, which benefits flavor improvement and texture modification (Zhao, He, et al., 2020; Zhao, Zhou, et al., 2020), imparting desirable flavor to dried aquatic products. However, oxidation is uncontrolled in food processing. The excessive oxidation not only generates unpleasant odors but also a large number of hazardous compounds that can harm consumer health and cause nutritional losses (Ma et al., 2025; Zhang, Fu, et al., 2025; Zhang, Hao, et al., 2025; Zhang, Zhong, et al., 2025). Zhao et al. (2022) investigated 39 significantly different lipid molecular species that were selected as the potential biomarkers of lipid and myofibrillar proteins (MPs) oxidation in hot-air-dried Penaeus vannamei. Cui et al. (2023) reported that drying significantly accelerates lipid oxidation in Scomberomorus niphonius. Ullah et al. (2023) reported temperature-dependent changes in the nutritive lipid profiles of Nile tilapia. Therefore, the air-drying process of aquatic products is primarily characterized by lipid and protein oxidation. Among these, unsaturated fatty acids are especially susceptible to auto-oxidation and thermal oxidation, which are triggered by exposure to light, temperature, metal ions, and oxygen (Li et al., 2023) Transition metal ions (particularly iron and copper) catalyze lipid oxidation via Fenton-type reactions; the bioaccumulation of such metal contaminants in marine organisms has been widely reported (Chan, Ali, et al., 2021), and their presence in muscle tissue can significantly accelerate oxidative processes during processing and storage. Lipid oxidation proceeds via free radical chain reactions with molecular oxygen, generating reactive oxygen species (ROS) that attack the amino acid side chains of MPs, particularly targeting methionine (Met), cysteine (Cys), and tryptophan (Trp) residues. This attack induces or accelerates MPs oxidation, leading to the production of alcohols, aldehydes, ketones, and volatile compounds (Zhang, Fu, et al., 2024; Zhang, Zhang, et al., 2024). Moreover, carbonyl compounds generated from protein oxidation react with lipids, promoting lipid oxidation and exacerbating product deterioration (Lan et al., 2025). Overall, lipid and protein oxidation forms a reciprocal cycle, ultimately degrading product quality through excessive oxidation. Beyond oxidation-induced quality deterioration, aquatic products and their processed forms are susceptible to other contamination risks. For instance, Ullah et al. (2022) highlighted the prevalence of banned Sudan dyes in food products available in South Asian markets, underscoring the importance of comprehensive quality monitoring. This emphasizes the need for robust analytical approaches, such as multi-omics integrated with machine learning, to detect and predict quality changes in processed seafood. In recent years, research on aquatic product oxidation has predominantly utilized single-omics techniques to analyze the impact of oxidation on the physicochemical properties of lipids and proteins. However, the mechanistic basis of the oxidative network underlying lipid–protein interactions remains poorly understood. Traditional chemical indices for oxidation evaluation are widely used, including sulfhydryl content, carbonyl value, peroxide value, and thiobarbituric acid value (TBARs) (Abeyrathne et al., 2021; Zhao et al., 2025). Although these methods can reflect the overall oxidation degree of food to some extent, they exhibit obvious limitations in revealing the dynamic changes of key biomarkers. However, the high computational power of machine learning (ML) enables the integration and analysis of diverse omics data, thereby facilitating the discovery of new biomarkers. ML models were developed to visualize changes in lipid and protein oxidation of mutton (Yi et al., 2025). Zhang, Fu, et al. (2024) and Zhang, Zhang, et al. (2024) revealed mutual promotion between protein conformational changes and PUFA peroxidation in salted large yellow croaker via lipidomics. Song et al. (2024) integrated ML with multi-omics approaches (lipidomics and metabolomics), revealing that phosphatidylcholine (PC), fatty acids, and phosphatidylethanolamine (PE) in air-fried salmon exhibit a strong correlation with lipid oxidation. Additionally, ML can elucidate the mechanisms underlying the formation of fishy volatile compounds derived from fatty acid oxidation. 7 volatile compounds were selected through random forest as significant contributors to the fishy off-flavor fingerprint in beef (Barragán-Hernández et al., 2024). Integrating lipidomics (lipid oxidation substrates), proteomics (protein oxidation substrates), and volatile organic compounds (oxidation products), ML enables systematic, multi-dimensional analysis of lipid and protein oxidation, facilitating accurate identification of key oxidative biomarkers and providing new perspectives on aquatic product oxidation mechanisms. This study takes air-dried hairtail as a research object and utilizes multi-omics data to develop ML predictive models for lipid and protein oxidation in hairtail and to identify key oxidative biomarkers that distinguish fresh from air-dried hairtail. The present study aims to comprehensively elucidate the intrinsic relationships between volatile compounds, protein and lipid oxidation in air-dried hairtail, to analyze the interactive oxidative pathways between lipids and proteins, and provide valuable molecular insights into oxidation mechanisms. We propose the following hypotheses: (1) Fresh hairtail undergoes lipid-protein interaction oxidation during the air-drying process. (2) Specific lipids and proteins showing strong correlations at the omics level can serve as predictive biomarkers of oxidative status, and (3) Transglutaminase-mediated cross-linking plays a central role in linking lipid and protein oxidation pathways during the air-drying process. Materials and methods Preparing air-dried hairtail Whole fresh hairtail (FH, average body weight of 300–400 g, commercial marine capture fishery products) individuals were purchased in December 2021 from an aquatic market (Zhoushan, Zhejiang province, China). All fish were already deceased upon procurement. The gills, viscera, tail, and the black peritoneal membrane of hairtail samples were manually excised. The treated samples were rinsed with distilled water, salted with 2% (w/w) solid NaCl, and subsequently air-dried outdoors for 2 d under natural conditions (temperature: 4–7 °C, relative humidity: 40–60%, wind speed: 2–6 m/s). Data collection and processing In the study, physicochemical indicators and multi-omics datasets (lipidomics, TMT proteomics, and volatile compounds) were derived from the Aquatic Product Processing Team of the College of Food and Pharmaceutical, Zhejiang Ocean University. Among these, lipidomics, TMT proteomics, and volatile flavor compounds were determined by LC-MS/MS, LC–MS/MS, and HS-GC-IMS, respectively. The physicochemical indicators and multi-omics data of fresh hairtail (FH) at 0 h and air-dried hairtail (FAH) at 48 h were collected, with 6 replicates set for each sample. Multi-omics data were generated as previously described (Liao et al., 2025; Liao, Ding, Du, et al., 2023; Liao, Ding, Wu, et al., 2023). Physicochemical parameters were collected based on two aspects: protein oxidation and lipid oxidation, including protein carbonyl content (PCC), total sulfhydryl content (TSC), total volatile basic nitrogen (TVB-N), lipid carbonyl content (LCC), peroxide value (POV), TBARs, pH, and color difference. Differential features in multi-omics datasets were screened using multi-dimensional analysis approaches, including variable importance in the projection (VIP), fold change (FC, FAH vs. FH), and P value. Notably, the Q2 intercept in the permutation test results was less than 0.05. The regression lines for R2 and Q2 exhibited an upward trend as permutational retention decreased, with both R2 and Q2 declining (Fig. S1). This indicates that the PLS-DA model is stable and reliable, avoiding overfitting and thereby preventing false positive results. The screening thresholds for differential lipids were set as P < 0.05, VIP ≥ 1, and FC (FAH vs. FH) = 1, whereas those for differential proteins were defined as P < 0.05 and FC (FAH vs. FH) > 1.2. Meanwhile, data on 75 identified volatile organic compounds (VOCs) were collected. To ensure data accuracy, the datasets were checked, with missing values and outliers removed or imputed using the K-nearest neighbor method. Additionally, the three omics datasets were normalized following their integration to eliminate the impact of scale differences among features on subsequent analyses. Oxidative feature selection and extraction An intermediate fusion strategy was employed to integrate the datasets, which included 193 differential proteins, 332 differential lipids, and 75 VOCs (Table S2). Then, we employed Pearson correlation coefficients to identify oxidation-related features. This involved analyzing the correlations between VOCs, lipid, and protein differential features of FAH, and physicochemical indicators. To identify strongly correlated oxidative biomarkers and reduce the problem of collinearity among oxidative features. Feature selection used a threshold of |r| ≥ 0.90 and a significance level of P < 0.05 for Pearson's correlation coefficient, considering these features highly correlated with lipid and protein oxidation. Additionally, Principal Component Analysis (PCA) was also performed to observe the clustering and dispersion patterns of samples. Wasserstein generative adversarial network with gradient penalty (WGAN-GP) WGAN-GP is an improved Generative Adversarial Network (GAN) model that not only generates realistic data samples to overcome the challenges posed by small sample limitations but also helps fix the instability and prevents mode collapse in traditional GANs. By introducing the Wasserstein distance, WGAN-GP provides more meaningful gradient information to enhance generator optimization (Yin et al., 2025). The WGAN-GP was constructed with slight modifications based on the method described by Zhang, Fu, et al. (2025), Zhang, Hao, et al. (2025) and Zhang, Zhong, et al. (2025). The generator and discriminator were fully connected deep neural networks that had undergone parameter optimization after 4000 training iterations with a gradient penalty weight of 50. Specifically, the generator network architecture consisted of an input layer, multiple hidden layers, and an output layer. The input layer consists of a noise vector configured as a 128 dimensional random vector sampled from a Gaussian distribution N (0,1) and a class label noise vector. Residual connections were introduced in hidden layers, with Rectified Linear Unit (ReLU) as the activation function, while the output layer utilized a tanh activation function. The discriminator network comprised an input layer, five hidden layers, and an output layer, with Leaky ReLU activation functions applied in the hidden layers. Both networks employed the Adam algorithm with an adaptive learning rate, which was set to 0.0002, configured with first and second moment estimates (beta1 = 0.9, beta2 = 0.999), and set the batch size to 256. Previous research indicates that a 10-fold expansion yields the best balance between dataset enlargement and synthetic data quality enhancement, supplying sufficient diversity to construct stable machine learning models. Machine learning models In this study, to comprehensively compare the performance of different algorithms on this dataset and to ensure objectivity in model selection, 7 machine learning models were selected for the classification of FH and FAH, including Random Forest (RF), Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), K-Nearest Neighbor (KNN), Decision Tree (DT), Multilayer Perceptron (MLP), and Naive Bayes (NB). Finally, the dataset was split into training and test sets at a 7:3 ratio using the train_test_split function in the sklearn library. Stratified sampling (stratify = labels) was applied to maintain a consistent class distribution. Hyperparameters were optimized using GridSearchCV, and model performance was evaluated via 5-fold cross-validation to avoid overfitting. Hyperparameter tuning for machine learning models was adapted from the methods described by Gu et al. (2024), Li et al. (2024), and Lu et al. (2022), with specific values detailed in Table 1. Based on the classification task requirements, evaluation metrics including Accuracy, Precision, Recall, F1 score, and Area Under Curve (AUC) were selected to assess model performance. Table 1Machine learning model hyperparameters.Machine learning modelHyperparametersRFcriterion = gini, max_depth = 3, min_samples_split = 5, n_estimators = 200SVMC = 0.05, kernel = linearXGBoostlearning_rate = 0.05, max_depth = 2, n_estimators = 100, subsample = 0.7KNNn_neighbors = 20, p = Euclidean distanceDTCriterion = gini, max_depth = 2, min_samples_leaf = 10, min_samples_split = 2MLPActivation = relu, hidden_layer_sizes = (256,128), learning_rate = invscaling, learning_rate_init = 0.001, max_iter = 500, solver = sgdNBvar_smoothing = 1 × e−8 Machine learning model hyperparameters. Oxidation biomarker identification Based on the model performance evaluation results, the optimal machine learning model was selected for feature selection in the FH and FAH datasets. In this study, the intersection of recursive feature elimination (RFE), Lasso, and RF feature selection methods was employed to identify the optimal feature subset. These features exhibited a strong correlation with lipid and protein oxidation, and obtained lipid and protein oxidation biomarkers in FAH. Meanwhile, the SHAP (SHapley Additive exPlanations) algorithm was applied for global model interpretation. Using the best-performing machine learning model, we computed the Shapley values for each feature and ranked their importance, thereby enhancing the credibility and transparency of the model. SHAP values were calculated using the SHAP library in Scikit-learn for Python 3.8. Data statistics and analysis The PCA, Lasso, and Machine learning classification models were constructed using the Scikit-learn library in Python 3.9 (Python Software Foundation, Beaverton, USA). following packages: scikit-learn (version 1.5.2), TensorFlow (version 2.16.2), and SHAP (version 0.47.2). Pearson correlation analysis of oxidative biomarkers was performed using the Pearson function from the scipy library module in Python. Data visualization was conducted using Origin 2021, BioRender (https://www.biorender.com/), and chiplot (https://www.chiplot.online/). Results and analysis Multi-omics data characterization and feature selection To compare the discriminative trends of VOCs, differential lipids, and proteins between FH and FAH, PCA analysis revealed distinct differences between FH and FAH samples, with the first two principal components explaining 85.3% (PC1 78.60% and PC2 6.70%) of the total variance. As shown in Fig. 1-A, the extracted principal components effectively captured the majority of significant variations and differences in the original data, reflecting the primary relational structure among variables. Given that highly correlated features enhance model interpretability and predictive performance, Pearson correlation analysis was conducted between physicochemical indices and VOCs, differential lipids, and proteins. This analysis aimed to identify key oxidation-associated features in FAH. As shown in Fig. 1-B and D, 122 features exhibited a strong correlation with oxidation in FAH (|r| ≥ 0.90, P < 0.05). The first two principal components accounted for 88.4% (PC1 83.10% and PC2 5.30%) of the total variance. This indicates that the oxidation-related features enhanced model interpretability, while the retained principal components effectively captured the dominant variation patterns and reduced noise from the original features. This suggests that features screened by Pearson coefficients are more effectively integrated to explain the raw multi-omics data and show strong associations with lipid and protein oxidation in FAH. Thus, based on VOCs, different lipid and protein integrated datasets can be used for the identification of oxidative biomarkers. The heatmap clustering results in Fig. 1-C are aligned with the PCA findings, demonstrating significant differentiation of oxidation-related features between FH and FAH samples. This consistency further validated the appropriateness of the selected oxidation features for subsequent machine learning classification model analysis. As depicted in the scree plot of Fig. 2-A, the first seven principal components accounted for 99% of the cumulative variance, retaining most of the original dataset's information while reducing dimensionality. The feature vectors generated via PCA dimensionality reduction were more effectively used for subsequent DA analysis, avoiding the curse of dimensionality. Data preprocessing. (A) Principal component analysis of differential features in multiple-omics; (B) Principal component analysis of oxidized features; (C) Clustered heatmap of oxidized features; (D) Pearson analysis of differential features with physicochemical indicators. Data augmentation and data quality assessment. (A) Oxidized feature scree plot; (B)Loss function trend plot; (C) Kernel density estimation plot; (D) Generated and original data boxplot. Data augmentation and quality assessment Previous studies have shown that a large number of labeled sample data are typically required to obtain efficient and accurate classification models (Sabanci et al., 2022). It is observed that the sample size of omics datasets is often limited, and using a small number of samples for training machine learning model may lead to overfitting. Therefore, appropriate data augmentation was performed based on the previously selected multi-omics oxidation feature dataset to simulate the feature distribution of the original data and generate similar data. The original data of 122 oxidation features were amplified 10-fold by applying the WGAN-GP framework, increasing the sample size from 12 to 120. Although the small initial sample size may have impacted the model's generalization capability, this feasibility validation of machine learning combined with multi-omics represents a valuable exploration. As shown in Fig. 2-B, C, and D, the quality and rationality of the data synthesized by the WGAN-GP model were evaluated through multi-dimensional methods to determine whether WGAN-GP could learn the distribution pattern of original data and generate data similar to the features. The distribution consistency between synthetic and original data is a key indicator for evaluating synthetic data quality. The results showed that the Fréchet distance (FID) value stabilized at approximately 2.56, where a lower FID value indicates higher distribution similarity (Shen et al., 2024). Meanwhile, the kernel density estimation plot showed a high structural similarity in the distribution between synthetic and original data, and the Kolmogorov-Smirnov test confirmed no significant differences in dimensions or distribution between the synthetic and original datasets (P > 0.05) (Gbashi et al., 2023). The boxplots showed that the median, upper quartile, and lower quartile were highly similar, with small median differences and limited outliers, indicating consistency in the data range and distribution between synthetic and original data. Therefore, the WGAN-GP model is effective for data augmentation of multi-omics data, and the generated data can be further used to establish machine learning classification models. However, we also proposed that future studies could incorporate a larger initial sample to optimize the model's robustness and generalizability further. Machine learning model development and performance To accurately identify oxidation biomarkers in FAH, RF, SVM, XGBoost, DT, KNN, MLP, and NB machine learning models were developed using the dataset after data augmentation (DA) on 122 oxidation features screened by Pearson coefficients. Meanwhile, GridSearchCV was applied to tune the hyperparameters of the seven machine learning classification models for performance optimization. Model classification performance was evaluated using Accuracy, Precision, Recall, F1 score, and AUC (Table 2). On the training set, the AUC values were as follows: RF, SVM, and XGBoost achieved an AUC of 1, while DT and MLP achieved an AUC of 0.99. NB and KNN exhibited AUC values of 0.98 and 0.95, respectively. Among these, XGBoost demonstrated the best performance, with its accuracy, precision, and F1-score, surpassing those of other models. Notably, its precision rate reached 1.00, demonstrating the capability of the XGBoost model to accurately predict positive examples in the training data with overall outstanding performance. The SVM and DT showed comparable performance, both achieving an accuracy and F1-score of 0.957. In contrast, KNN, NB, and MLP exhibited relatively weaker performance on the training set, with accuracy ranging from 0.913 to 0.924. Table 2Model performance evaluation.ModelDatasetAccuracyPrecisionRecallF1 scoreRFTrain0.9780.9780.9780.978Test0.9000.8331.0000.909SVMTrain0.9570.9380.9780.957Test0.9000.8331.0000.909XGBoostTrain0.9891.0000.9780.989Test0.9000.8331.0000.909DTTrain0.9570.9380.9780.957Test0.9500.9091.0000.952K-NNTrain0.9240.8980.9570.926Test0.9000.8331.0000.909MLPTrain0.9130.8960.9350.915Test0.9000.8331.0000.909NBTrain0.9240.8980.9570.926Test0.9000.8331.0000.909 Model performance evaluation. On the test set, model performances were more comparable. This can be attributed to the relatively low complexity of the classification task, the substantial oxidative variations observed in the samples, and the adoption of a unified feature set across all constructed models. DT reached the highest accuracy and F1 score, reflecting its superior comprehensive performance on test data. As shown in Fig. 3-A and B, XGBoost and DT had AUC values of 0.98 and 0.97, respectively, indicating that XGBoost slightly outperformed DT in overall performance. Based on the overall analysis of all evaluations of performance indicators, XGBoost was selected as the optimal model for classifying FH and FAH, serving as the basis for subsequent identification of oxidation biomarkers in hairtail. Model classification performance analysis. (A) Training set subject operating characteristic curve (ROC-Curve); (B) Test set subject operating characteristic curve (ROC-Curve); (C) Lasso error plot; (D) Lasso path plot. Identification of oxidative biomarkers In this study, the oxidation features in FAH were identified through correlation analysis, and seven machine learning classification models were developed. Following a comprehensive evaluation of model performance, XGBoost was selected to implement the XGBoost-REF feature selection technique. As shown in Fig. 3-C and D, Lasso, XGBoost-REF, and RF feature selection techniques were performed on 122 oxidation features, identifying 12, 35, and 61 oxidative biomarkers, respectively. To minimize false positives, we selected the intersection of biomarkers from the three methods as the oxidation biomarkers. As shown in Fig. 4-A and Table S4, the Venn plot revealed six common oxidation features across all methods, including four differential proteins and two differential lipids: receptor expression-enhancing protein 6-like (REEP6; TRINITY_DN3201_c0_g1_i2_orf1), glutathione S-transferase omega-1-like (GSTO1; TRINITY_DN15333_c0_g1_i1_orf1), peroxisomal biogenesis factor 19 (PEX19; TRINITY_DN3869_c0_g1_i1_orf), protein-glutamine gamma-glutamyltransferase 2-like isoform X2 (TG2; TRINITY_DN6280_c0_g1_i1_orf1), dMePE (18:4/16:0), and PE (20:1/18:1). It is noteworthy that oxidative biomarkers do not encompass VOCs. However, this may be due to the limited number of VOC characteristics detectable by GC-IMS, as well as the inherently complex and dynamic nature of VOCs. The three methods identified the following compounds: 2,3-dimethylpyrazine, 2-methylpropanal, 3-methylbutanal, and benzaldehyde. These are VOCs generated by the oxidative interaction between lipids and proteins (Table S3). These VOCs are produced via Strecker degradation, whereby lipid oxidation products react with amino acids. Oxidative biomarker identification and characterization importance analysis. (A) Oxidative biomarker Venn plot; (B) SHAP summary plot; (C) Characterization importance bar chart; (D) SHAP waterfall plot. The SHAP algorithm was employed to quantify feature importance and visualize machine learning model results, thereby enhancing the interpretability of the classification model (Lundberg et al., 2020). XGBoost-SHAP analysis quantified the average magnitude of each impact of the oxidation biomarker on predictions. As displayed in Fig. 4-C, the bar chart of average absolute SHAP values revealed that REEP6 and PE (20:1/18:1) were the most critical oxidation biomarkers for predicting oxidation in FAH. PEX19 and TG2 also played key roles in the model decision-making process. Additionally, the SHAP algorithm further enabled exploration of how input variables influenced specific classification outcomes. The SHAP values of the six selected biomarkers, ranked by feature importance, are shown in the summary plot of Fig. 4-B. This summary plot illustrates varying contributions of different oxidation biomarkers to hairtail oxidation: REEP6, GSTO1, and dMePE (18:4/16:0) exhibited positive effects on model output as their values increased, whereas higher values of PEX19 and PE (20:1/18:1) had negative effects. The negative SHAP values of PEX19 and PE (20:1/18:1) can be attributed to their susceptibility to oxidation, with the potential for oxidative degradation or modification under severe oxidative conditions. The SHAP values of dMePE (18:4/16:0) were concentrated around 0, indicating a relative contribution to the model output. The waterfall plot (Fig. 4-D) provided a local interpretation: PEX19 exerted the largest positive push (SHAP +1.19), whereas REEP6 provided the strongest opposing force (−1.25). This local pattern clearly illustrates how oxidation biomarkers jointly determine the model's decision. These results confirm that the six oxidation biomarkers are critical indicators of lipid and protein oxidation in FAH. Overall, the SHAP algorithm analysis improved the interpretability and transparency of the model, providing clear insights into oxidative biomarker prediction. Biomarker expression patterns and correlations To explore the dynamic changes of oxidation reactions and potential relationships between lipid and protein oxidation in FAH, hierarchical clustering heatmaps and correlation analyses were performed for the screened oxidative biomarkers. As shown in Fig. 5-A and B, the hierarchical clustering heatmaps visually exhibited expression differences of oxidative biomarkers, with FH and FAH distinctly separated. The clustering results of oxidative biomarkers revealed two main categories: PEX19, TG2, and PE (20:1/18:1) were clustered together, showing significantly higher contents in FH than in FAH. The other cluster consisted of REEP6, GSTO1, and dMePE (18:4/16:0), whose contents significantly increased during hairtail air-drying compared to FH. Studies have reported that PE, composed of glycerol, fatty acids, phosphoric acid, and ethanolamine, exhibits potent antioxidant properties. It undergoes the Maillard reaction to generate conjugates with also enhanced antioxidant activity (Wang, Fan, et al., 2024; Wang, Jia, et al., 2024; Wang, Zang, et al., 2024). Dimethylphosphatidylethanolamine (dMePE), a derivative of PE, serves as an intermediate product of phospholipid degradation and is associated with oxidation-induced lipid degradation pathways. Saleh et al. (2019) reported that transglutaminase plays a key role in the γ-glutamyl cycle, maintains normal redox status, and exerts antioxidant and anti-inflammatory effects against tissue damage. PEX19 is essential for peroxisome maturation. Peroxisomes contain catalase and various oxidases that decompose H2O2 and reduce free radical production (Hettema & Motley, 2009). However, oxidases in peroxisomes significantly promote fatty acid β-oxidation. Collectively, oxidative biomarkers in hairtail muscle cells primarily exhibit antioxidant effects after hairtail death. During air-drying, factors such as temperature, air humidity, and NaCl disrupt the redox balance (Zhao, He, et al., 2020; Zhao, Zhou, et al., 2020). Studies have shown that dMePE contents in salt-cooked Parapenaeopsis hardwickii and Penaeus vannamei significantly increase with drying time (P < 0.05) (Wang, Fan, et al., 2024; Wang, Jia, et al., 2024; Wang, Zang, et al., 2024; Zheng et al., 2025). Additionally, severe damage to the muscle cell membrane in FAH exacerbates oxidation, leading to rapid accumulation of oxidation intermediates and products (Wang et al., 2022). Correlation analysis of oxidized biomarkers. (A) Oxidized biomarker content; (B) Hierarchical clustering heat map analysis of oxidized biomarkers; (C) Correlation analysis of oxidized biomarkers. Note: * P < 0.05; ** P < 0.01; *** P < 0.001. To comprehensively assess potential correlations among oxidative biomarkers in FAH, Pearson correlation analysis results are presented in Fig. 5-C. GSTO1 was significantly positively correlated with dMePE (18:4/16:0) (r = 0.88, P < 0.001) and REEP6 (r = 0.72, P < 0.01). PE (20:1/18:1) was significantly positively correlated with both PEX19 (r = 0.63, P < 0.05) and TG2 (r = 0.64, P < 0.05). REEP6 showed a strong positive correlation with dMePE (18:4/16:0) (r = 0.91, P < 0.001), while PEX19 was significantly positively correlated with TG2 (r = 0.94, P < 0.001). Conversely, REEP6 was significantly negatively correlated with PEX19 (r = −0.96, P < 0.001), TG2 (r = −0.97, P < 0.001), and PE (20:1/18:1) (r = −0.71, P < 0.01), Similarly, dMePE (18:4/16:0) negatively correlated with PEX19 (r = −0.88, P < 0.001), TG2 (r = −0.84, P < 0.001), and PE (20:1/18:1) (r = −0.71, P < 0.01). These results suggest that coordinated changes in PEX19, TG2 and PE (20:1/18:1) influence oxidative interactions with lipids and proteins. Potential oxidation mechanisms of interaction lipid and protein oxidation biomarkers To reveal the molecular mechanisms of lipid-protein interactive oxidation during hairtail air-drying, we explored the potential pathways of six oxidative biomarkers (Fig. 6). Under normal physiological conditions, the production and elimination of ROS in cells maintain a dynamic balance. Disruption of this balance triggers oxidative stress, characterized by ROS accumulation, reduced antioxidant activity, and increased membrane permeability (Fan et al., 2025). Thereby, after slaughter and exsanguination, blood circulation ceases, rapidly depleting nutrients to muscle tissues and disrupting oxygen delivery, forcing skeletal muscle cells into a state of enforced death under adverse physiological conditions (Li et al., 2022). (A) Potential lipid and protein oxidation pathways for oxidative biomarkers; (B) Mechanisms predicting lipid and protein interaction oxidation. During this period, glycolysis produces low-level ATP, while mitochondria maintain basic functions and activate endogenous enzymes and signaling factors (Fan et al., 2024). REEP6, primarily located near ER-mitochondria contact sites to maintain mitochondrial homeostasis (Fan et al., 2022). Thus, the cell is subjected to oxidative stimulation, and the ER is similarly subjected to oxidative stress, which activates G protein-coupled receptors (GPCRs) on cell membranes via REEP6, triggering downstream adenylate cyclase to generate cyclic adenosine monophosphate (cAMP) (Björk et al., 2013; Hotamisligil, 2010; Son et al., 2022). As a second messenger, cAMP activates the protein kinase A (PKA) pathway, enhancing phosphorylation of hormone-sensitive lipase (HSL). PKA, an upstream signaling molecule of adenosine monophosphate-activated protein kinase (AMPK), activates AMPK by phosphorylating Thr 172 (Yan et al., 2014). Activated AMPK inhibits lipogenesis, promotes adipose triglyceride lipase (ATGL) activity, and induces fatty acid oxidation (Kim et al., 2016; Kohjima et al., 2008). Phosphorylation of HSL and ATGL promotes lipid hydrolysis, generating free fatty acids and fragmenting intramuscular lipid droplets into smaller units to facilitate lipid oxidation (Zhao, He, et al., 2020; Zhao, Zhou, et al., 2020). Meanwhile, Short-chain fatty acids produced by lipolysis undergo oxidative phosphorylation in mitochondria, generating abundant ROS. But PUFAs (such as EPA and DHA) and some of the branched-chain fatty acids in aquatic products require peroxisomal oxidation to form shorter fatty acids, producing H2O2. Peroxisomes also participate in polyamine oxidation and amino acid catabolism. In a hypoxic environment, abnormal expression of divalent metal ion transporter 1, ferroportin 1, and transferrin receptor leads to intracellular Fe2+ overload. Free Fe2+ in unstable iron pools is oxidized to Fe3+ via the Fenton reaction, generating •OH (Wang, Fan, et al., 2024; Wang, Jia, et al., 2024; Wang, Zang, et al., 2024). Eventually, the intracellular redox system fails to maintain relative homeostasis, leading to massive accumulation of lipid peroxides. Intracellular TG2 is inactive under normal conditions but activated by H2O2 or UV-induced calcium influx during oxidative stress (Shin et al., 2004). At this point, TG2 plays a pivotal role in inducing lipid-lipid, lipid-protein, and protein-protein interactive oxidation. It has been reported that during cell death, TG2 is thought to be responsible for cross-linking proteins to the membrane, preventing cellular contents leakage, forming a gel network of protein-protein and protein-lipid conjugates (Jeong et al., 2009). This structural change can facilitate the immobilization of ROS, PUFAs, and lipid peroxides within the three-dimensional gel network (Chen et al., 2021; Zeng et al., 2025). ROS produced by mitochondria, H₂O₂ produced by peroxisomes, •OH reactive oxygen species, lipid peroxides, and electrophiles (carbonyl groups, aldehydes, ketones) are trapped within this network, where they attack protein side chains and PUFAs, thereby forming alkyl radicals (R•). This cyclic reaction of R• with O2 continues until radical quenching terminates the chain reaction. However, unstable ROOH decomposes into small-molecule oxidation products, such as aldehydes, ketones, alcohols, volatile organic acids, and epoxides. Aldehydes and ketones from lipid oxidation bind to lipids and proteins, promoting oxidation. Thus, TG2-induced gel networks trigger lipid-protein and protein-protein interactive oxidation (Ali Khan et al., 2024). Discussion The oxidative interaction of lipids and proteins in air-dried hairtail is a complex process. The six oxidative biomarkers identified in this study exert different effects on the oxidative interaction of lipids and proteins in air-dried hairtail. PE (20:1/18:1) provides precursors for lipid oxidation, and the primary amino groups of biomolecules and proteins act as potential targets for covalent modification by lipid peroxidation products (Breitling-Utzmann et al., 2001). Studies have shown that PE reacts with 13-hydroperoxyoctadecadienoic acid or other polyunsaturated fatty acid peroxides to increase lipid peroxidation levels, generating N-(hexanoyl) ethanolamine (Tsuji et al., 2003). For instance, oxidation of PE and dMePE occurs at unsaturated bonds, forming phospholipid hydroperoxides that decompose into long or short-chain products (Zhang, Dai, et al., 2023; Zhang, Xie, et al., 2023). Long-chain fatty aldehydes (hexadecanal and octadecanal) were derived from plasmalogen hydrolysis (Zhang, Fu, et al., 2025; Zhang, Hao, et al., 2025; Zhang, Zhong, et al., 2025). Oxidative degradation of PE also produces 1-octen-3-ol and (E)-2-decenal (Li, Bai, et al., 2025; Li, Liu, et al., 2025). Meanwhile, the PE methylation pathway involves three consecutive transmethylation steps, all catalyzed by PE methyltransferase (PEMT). The enzyme converts PE first into methyl-PE, then into dMePE, and finally into phosphatidylcholine (PC) (Xu et al., 2022). Compared to PE, the choline group enhances fatty acyl oxidative stability, possibly explaining the higher dMePE content in FAH (Chen et al., 2024). PEX19 and TG2 disrupt the original redox system, tending toward an oxidative direction and generating abundant free radicals that contribute to lipid and protein oxidation during hairtail air-drying, ultimately leading to the accumulation of substances such as REEP6 and dMePE (18:4/16:0). REEP6 excessive expression of significantly reduces muscle Fe2+ content, which correlates negatively with Fe2+ levels (Xia et al., 2024). This may occur because REEP6 facilitates the Fenton reaction between Fe2+ and H2O2, promoting Fe3+ and free radical production. Moreover, REEP6 is highly enriched in adipocytes, regulating protein oxidative phosphorylation levels and promoting adipocyte β-oxidation (Ye et al., 2013). Meanwhile, salt plays a key role in promoting lipid oxidation in FAH (Chan et al., 2025; Chan, Hasan, et al., 2021). Salt concentrations of 1–5% enhance lipase activity, and increasing salt content fragments lipid droplets into smaller units, accelerating triglyceride hydrolysis (Zhao, He, et al., 2020; Zhao, Zhou, et al., 2020). Meanwhile, in the presence of NaCl and REEP5 (a homolog of REEP6 with high sequence similarity), lipid droplets fragment into smaller ones, increasing their specific surface area. Beyond oxidative stress, the air-dried hairtail is influenced by microbial activity and endogenous protective enzymes. Naveed et al. (2025) demonstrated that lysozyme extracted from jellyfish exhibits potent antibacterial and antibiofilm properties, suggesting that natural antimicrobial proteins in aquatic organisms may protect against spoilage. Similarly, our identification of GSTO1 as an oxidative biomarker highlights the role of endogenous antioxidant systems in maintaining product quality during air-drying (Hussain Chan et al., 2021). The PE in the phospholipid bilayer acts as an antioxidant, consuming ROS to delay oxidative stress and forming lipid peroxides. Similar to glutathione peroxidase 4 (GPX4), GSTO1 protects cells from lipid peroxides by shielding macromolecules from electrophilic attacks (Thapliyal et al., 2023). However, Zhang, Dai, et al. (2023) and Zhang, Xie, et al. (2023) demonstrated that ferroptosis-induced downregulation of GSTP1 results from HECT E3 ubiquitin ligase-mediated polyubiquitination and proteasomal degradation of GSTP1. Hossain et al. (2023) reported that GSTO1 activity is downregulated under hypoxic conditions. Following slaughter, Hairtail muscle primarily undergoes anaerobic oxidation via the glycolytic pathway, creating an intracellular hypoxic environment that reduces GSTO1 antioxidant defense. Ultimately, a large amount of lipid peroxides accumulates in the muscle tissue of the hairtail, damaging cell membranes, disrupting cellular function, and inducing ferroptosis. Recent studies highlight ferroptosis as a key regulator of postmortem meat quality under hypoxia, with hypoxia-inducible factor 1α mediating protein oxidation and cell death (Zhu et al., 2025). Chen et al. (2023) also revealed that ferroptosis is the primary mode of cell death in grass carp erythrocytes under saline conditions (mimicking a nutrient-deficient environment) and in serum. PE decides cellular sensitivity to ferroptosis, while TG2 enhances this sensitivity. Peroxisomes serve as carriers for lipid peroxides that propagate ferroptosis, whereas REEP6 acts as a ferroptosis inhibitor. GSTO1 directly scavenges the secondary oxidation product 4-hydroxynonenal (4-HNE) to prevent ferroptosis (Alves et al., 2025). Gagliardi et al. (2023) identified the existence of an unknown compensatory mechanism between TG2 and GPX4 with a potential role in regulating cellular ferroptosis. In conclusion, oxidative biomarkers are intimately linked to cellular ferroptosis, a process driven by the overproduction of reactive oxygen species ROS, including free radicals and lipid peroxides. Transglutaminases are a family of thiol- and Ca2+-dependent acyl transferases that catalyze the formation of a covalent bond between the γ-carboxamide groups of peptide-bound glutamine residues and various primary amines (ε-amino groups of lysine), forming crosslinked or aminated proteins (Sarang et al., 2005). A series of findings indicate that TG2 is inextricably linked to lipid and protein oxidation in food. It was found that peroxides and TBARs were significantly increased in emulsions with microbial transglutaminase (MTGase) added to Oncorhynchus mykiss, Merluccius merluccius, and Merluccius capensis recombinant meat compared to restructured meat with no added MTGase, independent of fat content (Moreno, Carballo, & Borderías, 2010; Moreno, Javier Borderías, & Baron, 2010). Similar findings show enhanced protein oxidation (elevated carbonyl content) in Scomberomorus niphonius with MTGase addition (Pan et al., 2019). This may be due to the fact that cross-linking of TG2-acting proteins induced changes in the protein structure that altered the iron-binding site on the protein, bringing iron and peroxides closer and triggering oxidation (Kellerby et al., 2006). The cross-linking effect of TG2 resulted in a more compact spatial structure of the muscle, increasing collision frequency between ROS, fatty acids, and proteins in the gel network to promote interactive oxidation, acting as an oxidation amplifier. Additionally, proteins also release ammonia in response to TG2, which enhances lipid oxidation, promotes free radical production, and inhibits antioxidant enzyme activity. Concurrently, the progressive hydrolysis of proteins and lipids creates favorable conditions for the cross-oxidation between lipids and proteins. Consequently, the three-dimensional gel network induced by TG2 promotes the oxidation interaction of lipids and proteins in air-dried hairtail. Deamidation of TG2 increased the content of negatively charged amino acids (Glu and Asp), resulting in a decrease in surface hydrophobicity and promoting the aggregation of hydrophobic amino acids inside the protein (Trp, Met, and Tyr) (Pei et al., 2023). In the gel network, oxidative modification of MPs by ROS. Cys is particularly susceptible to sulfhydryl oxidation under oxidative conditions. This reaction leads to the formation of intramolecular disulfide bonds with adjacent amino acid sulfhydryl groups (Li, Bai, et al., 2025; Li, Liu, et al., 2025). This may explain why TG2 reduces free sulfhydryl content, promotes disulfide bond formation, and enhances surimi gel properties (Liu et al., 2025). Hydrophobic amino acids undergo a dual oxidation process in which ROS attack Tyr and Met, leading to the formation of active methionine sulfoxide and methionine sulfone intermediates. Then, peroxides derived from fatty acids or amino acids can bind to other amino acids, participating in the formation of disulfide and dityrosine bonds, thereby inducing protein cross-linking and aggregation (Chen et al., 2025; Dorta et al., 2019). Notably, aldehydes can covalently bind to the ε-amino group of lysine, and this interaction is irreversible. The carbonyl group reacts with a basic residue (N-terminal amino group) to produce a Schiff base (Poojary & Lund, 2022). MTGase can compete with free MDA for the modification of MPs, affecting MDA-induced MPs oxidation (Lv et al., 2021). Therefore, it can be concluded that lipid-lipid, lipid-protein, and protein-protein interactions influence the oxidation of FAH. Conclusions In this study, we integrated VOCs, lipidomics, and proteomics data to construct seven machine learning algorithms, uncovering the oxidative biomarkers and molecular mechanisms of lipid and protein interaction oxidation during FAH processing. The XGBoost model achieved accurate classification between fresh hairtail and air-dried hairtail (AUC = 0.98). A total of six oxidative biomarkers of air-dried hairtail were identified: REEP6, GSTO1, PEX19, TG2, dMePE (18:4/16:0), and PE (20:1/18:1). The mechanism suggests that REEP6 promotes lipolysis and polyunsaturated fatty acid oxidation through activation of the PKA/AMPK signaling pathway. The TG2-mediated protein-lipid cross-linking network synergistically amplifies oxidative interaction of lipid and protein effects, forming a lipolysis-ferroptosis-cross-linking oxidative cascade. This study establishes a multi-omics and machine learning-based potential pathway for aquatic product oxidation, providing molecular targets for deciphering the oxidation mechanism of FAH. Additionally, future research focusing on oxidative biomarkers should further investigate their dynamic evolution patterns and the underlying mechanisms governing flavor formation in aquatic products. The six lipid and protein interaction-derived oxidative biomarkers identified in this study in air-dried hairtail still have some shortcomings and limitations. In future research, targeted analyses could be conducted, and the variations in these oxidative biomarkers in other seafood products could be investigated. Furthermore, the impact of these oxidative biomarkers on the quality of air-dried hairtail is of great importance in actual production and processing. CRediT authorship contribution statement Yi Luo: Writing – review & editing, Writing – original draft, Formal analysis, Data curation. Hui-Min Lin: Writing – review & editing, Supervision, Formal analysis, Data curation, Conceptualization. Shucheng Liu: Writing – review & editing, Supervision, Investigation. Soottawat Benjakul: Writing – review & editing, Supervision, Investigation. Bin Zhang: Writing – review & editing, Methodology, Investigation, Funding acquisition, Conceptualization. Ethical statement No studies involving humans or animals were conducted in the present research. As no live animal experimental procedures were carried out in this study, compliance documentation and ethical approval under the Animals (Scientific Procedures) Act 1986, EU Directive 2010/63/EU or equivalent laboratory animal guidelines are not required. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Acknowledgments This study was funded by the Zhejiang Province Commonweal Projects (LTGN24C200008), and National Natural Science Foundation of China (Grant No. U23A20263), and the Zhejiang Natural Science Foundation Project (Grant No. LRG25C200001), and the National Key Research and Development Program of China (Grant No. 2021YFD2100504). Footnotes Supplementary data to this article can be found online at https://doi.org/10.1016/j.fochx.2026.104355. Contributor Information Hui-Min Lin, Email: linhuixiaomin@126.com. Bin Zhang, Email: zhangbin@zjou.edu.cn. Supplementary data PLS-DA model analysis results. Supplementary Table S1 Quality control information for omics data. Supplementary Table S2 VOCs identified using three feature selection methods. Supplementary Table S3 Oxidation biomarker data information. Data availability Data will be made available on request. References Associated Data Supplementary Materials Supplementary Table S1 Quality control information for omics data. Supplementary Table S2 VOCs identified using three feature selection methods. Supplementary Table S3 Oxidation biomarker data information. Data Availability Statement Data will be made available on request.