NeXtMD: a new generation of machine learning and deep learning stacked hybrid framework for accurate identification of anti-inflammatory peptides

Abstract

Background

Accurate identification of anti-inflammatory peptides (AIPs) is crucial for drug development and inflammatory disease treatment. However, the short length and limited informational content of peptide sequences make precise computational recognition particularly challenging. While various machine learning and deep learning approaches have been explored, their limitations in feature representation and model integration hinder the effective discovery of novel AIPs.

Results

In this study, we present NeXtMD, a novel dual-module stacked framework that integrates both machine learning (ML) and deep learning (DL) components for accurate AIP identification. NeXtMD systematically extracts four functionally relevant sequence-derived descriptors—residue composition, inter-residue correlation, physicochemical properties, and sequence patterns—and utilizes a two-stage prediction strategy. The first stage generates preliminary predictions using four distinct encoding strategies and ML classifiers, while the second stage employs a multi-branch residual network (ResNeXt) to refine prediction outputs. Benchmark evaluations demonstrate that NeXtMD outperforms current state-of-the-art methods on multiple performance metrics. Moreover, NeXtMD maintains strong generalization capabilities when applied to unseen peptide sequences, showing its robustness and scalability.

Conclusions

NeXtMD offers a high-performance and interpretable computational framework for AIP identification, with significant potential to facilitate the discovery and design of peptide-based anti-inflammatory therapeutics. The architecture and methodological innovations of NeXtMD also provide a generalizable strategy that can be adapted to other bioactive peptide prediction tasks, supporting broader applications in therapeutic peptide development.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12915-025-02314-8.

Background

Inflammation is an essential biological response of the immune system to harmful stimuli, and its persistent activation is closely associated with various chronic diseases. For instance, obesity exacerbates kidney injury through mechanisms such as chronic inflammation, insulin resistance, and oxidative stress; clinical interventions for obesity-associated nephropathy typically emphasize lifestyle modifications and pharmacological treatments targeting metabolic pathways [1]. Currently, non-steroidal anti-inflammatory drugs, corticosteroids, and immunosuppressants are the primary clinical treatments; however, their prolonged use often results in drug resistance and adverse effects. Therefore, there is an urgent need to develop more precise, efficient, and safe anti-inflammatory strategies [2, 3]. Anti-inflammatory peptides (AIPs), typically composed of 5–50 amino acids, can effectively inhibit the production of pro-inflammatory factors and regulate key inflammatory signaling pathways, making them promising candidates for next-generation anti-inflammatory therapies [4]. Clinically, peptides such as LL-37 and PXL01 have demonstrated significant anti-inflammatory and tissue-repairing efficacy in trials for ulcerative colitis and postoperative tendon adhesion, respectively [5]. LL-37 has been evaluated in clinical trials for venous leg ulcers, showing dose-dependent activity and safety [6, 7]. PXL01 has completed Phase II clinical trials for preventing postoperative adhesions after tendon repair surgery and is progressing towards Phase III trials [6, 8].

With the rapid expansion in the number of identified peptide sequences, traditional experimental methods for evaluating their anti-inflammatory potential have become increasingly labor-intensive, costly, and inefficient. In contrast, computational approaches have emerged as critical tools for peptide function prediction due to their low cost and high efficiency [7, 9–14]. Currently, machine learning (ML)-based models employing algorithms such as Random Forest (RF) [15, 16], Support Vector Machine (SVM) [17–19], and eXtreme Gradient Boosting (XGBoost)—as exemplified by PreAIP, AIPstack, and TriStack—have achieved considerable success [8, 20–26]. In parallel, deep learning (DL) techniques, including Convolutional Neural Networks (CNN) [27], Long Short-Term Memory (LSTM), and attention mechanisms, have demonstrated robust capabilities in analyzing complex, large-scale peptide datasets [28–32]. However, existing computational methods face several limitations: (1) sequence feature extraction is often oversimplified, failing to capture comprehensive biochemical and physicochemical properties; (2) most prediction models rely on single classifiers, increasing susceptibility to bias and variance; and (3) few studies systematically investigate the integration of ML and DL models to enhance predictive accuracy [33].

To address these limitations, we propose NeXtMD, a dual-module stacked model that integrates both ML and DL approaches to accurately identify AIPs (Fig. 1). First, NeXtMD systematically incorporates four critical sequence-derived descriptors—Dipeptide Deviation Encoding (DDE), Composition of K-Spaced Amino Acid Pairs (CKSAAP), Physicochemical Properties grouped into 16 classes (PP16), and Auto-Correlation of Hydrophobicity (ACH)—to comprehensively capture both local and global peptide sequence characteristics [34]. Subsequently, we designed a two-stage prediction framework. In the first stage, predictions are generated through an ensemble of four ML classifiers—RF, XGBoost, Light Gradient Boosting Machine (LightGBM), and Gradient Boosting Decision Trees (GBDT)—employing a fivefold cross-validation strategy. In the second stage, a multi-branch residual refinement network further optimizes these preliminary predictions, thereby enhancing classification accuracy [35]. Experimental evaluations conducted on the AIP dataset demonstrate that NeXtMD significantly surpasses existing benchmark methods, underscoring its superior predictive performance and interpretability. Collectively, NeXtMD represents a robust computational tool with strong potential for advancing the discovery and development of clinically relevant AIPs. Furthermore, its robust performance across diverse external test sets highlights its transferability to the prediction of other categories of therapeutic peptides.

Fig. 1.

Schematic illustration of the NeXtMD framework. A Overview of the NeXtMD architecture, which comprises four main components: feature extraction, initial prediction using machine learning classifiers, feature refinement via a multi-layer grouped residual network, and final classification. B Representation of the first module, including feature fusion and predictions generated by four distinct machine learning classifiers (RF, XGBoost, LightGBM, and GBDT). C Structure of Residual Networks neXt (ResNeXt) module, consisting of five sequential layers with multiple residual blocks, followed by the terminal classification layer

Results

Dataset characterization and sequence feature analysis

To comprehensively characterize the differences between AIPs and non-AIPs, we performed a systematic analysis of the dataset (Fig. 2). First, we examined the sequence length distribution and observed that AIPs predominantly range from 10 to 30 amino acids, with over 80% of sequences falling within this interval, whereas non-AIPs exhibit a broader and more uniform length distribution (Fig. 2A). This result aligns with the functional nature of AIPs, which are typically short bioactive peptides. Next, we compared the amino acid composition profiles of AIPs and non-AIPs. As shown in Fig. 2B, AIPs are notably enriched in basic and hydrophilic residues such as lysine (K), arginine (R), and threonine (T), whereas non-AIPs display higher frequencies of acidic residues such as aspartic acid (D) and glutamic acid (E). These differences may reflect the physicochemical preferences required for AIP biological activity, including membrane interaction and immune modulation.

Fig. 2.

Analysis of the AIPs dataset. A Sequence length distribution comparison between positive (AIPs) and negative (non-AIPs) samples within the dataset. B Average amino acid composition profiles of 20 canonical amino acids in AIPs and non-AIPs. C Heatmap visualization depicting the distribution of amino acid attribution scores derived from randomly selected sequences in the AIP dataset. D Heatmap visualization depicting the distribution of amino acid attribution scores derived from randomly selected sequences in the non-AIP dataset. E Positional preference of conserved residues identified within AIP sequences. F Positional preference of conserved residues identified within non-AIP sequences

To further explore sequence-specific features, we visualized amino acid attribution scores derived from model interpretation methods. The heatmap in Fig. 2C shows that AIP sequences tend to harbor high-contribution residues clustered at specific positions, suggesting the presence of functional hotspots. In contrast, attribution scores in non-AIP sequences are more diffusely distributed with lower overall contribution values, indicating a lack of conserved predictive patterns (Fig. 2, Fig. S1, 2). Consistent with these findings, sequence logo analysis revealed clear positional preferences in AIPs, where residues such as leucine (L), alanine (A), glutamic acid (E), glycine (G), and phenylalanine (F) are highly conserved across multiple positions (Fig. 2E). In comparison, non-AIPs lack such conserved motifs and display more scattered patterns (Fig. 2F). In addition, Fig. S3, 4 present the differences in the contribution values of AIPs and non-AIPs to the four categories of features selected for this study. Meanwhile, Fig. S5, 6 present histograms to visualize the difference in information. These results collectively indicate that AIPs exhibit distinct sequence characteristics and positional conservation, which can be effectively leveraged for computational prediction and classification.

Ensemble model NeXtMD can accurately predict AIPs

To evaluate whether the proposed NeXtMD method can accurately predict AIPs, we first systematically assessed the predictive performance of multiple individual ML models based on their area under the receiver operating characteristic curve (AUC) values. Four models with the highest AUC scores were selected to construct the first-layer ensemble architecture: RF (AUC = 0.8197), XGBoost (AUC = 0.8074), LightGBM (AUC = 0.8038), and GBDT (AUC = 0.8038). The ensemble of these four models achieved an AUC of 0.8149 on the test set, demonstrating the benefit of model integration (Fig. 3A).

Fig. 3.

Results of the NeXtMD model section with integrated comparison and fivefold cross-validation. A ROC curves obtained by a single ML algorithm. B ROC curves obtained from meta-feature inputs of ResNeXt using a single ML algorithm. C ROC curve of NeXtMD model under fivefold cross-validation. D PRC curve of NeXtMD model under fivefold cross-validation

Next, we evaluated and compared the classification performance of these four ML models, their ensemble, and two deep learning-based frameworks: the NeXtMD model and a ResNeXt-enhanced variant. As shown in Fig. 3B, all four individual ML models outperformed random classification (AUC = 0.5), each achieving AUC values above 0.8. The integrated ensemble model further improved performance, reaching an AUC of 0.8607 and an AUPRC of 0.8615, highlighting its strong generalization ability for proteomics-based feature learning.

In the comparison, we selected six model evaluation metrics: AUC, accuracy (ACC), precision, recall, F1-score, Matthews correlation coefficient (MCC). The specific comparison results can be found in Table 1. Finally, we investigated the effect of incorporating the ResNeXt deep learning module into the NeXtMD framework. Experimental results showed that the integration of ResNeXt consistently boosted the predictive accuracy of each individual ML classifier. More importantly, the NeXtMD ensemble—combining both ML and deep learning—outperformed ensembles built solely from traditional ML models. Notably, the performance of the NeXtMD ensemble using fivefold cross-validation surpassed that of the same model under tenfold cross-validation, suggesting that the model is robust and effective even with fewer validation folds (Fig. 3C, D).

Table 1.

The NeXtMD model setup cross-validation K = 5 vs. K = 10

NeXtMD	AUC	ACC	Precision	Recall	F1-score	MCC
K = 5	0.8607	0.7995	0.8077	0.7130	0.7500	0.5883
K = 10	0.8504	0.7852	0.7557	0.7389	0.7472	0.5606

NeXtMD outperforms state-of-the-art methods in AIP prediction

To comprehensively assess the predictive power of NeXtMD, we benchmarked it against several state-of-the-art AIP prediction models, including TriStack, AIPStack, TriNet, and three feature-specific variants of PPTPP (cls, prb, and fus) [36]. All competing models were retrained locally using the same AIP dataset and their official implementations to ensure a fair comparison.

As summarized in Fig. 4 and Table 2, NeXtMD consistently outperformed all other models across five out of six evaluation metrics, including AUC (0.8607), precision (0.8062), recall (0.7134), F1-score (0.7513), and MCC (0.5884). Although TriStack slightly surpassed NeXtMD in accuracy (ACC = 0.8219 vs. 0.7996), NeXtMD still achieved a balanced and robust performance profile, particularly excelling in AUC, the most critical indicator for classification tasks involving imbalanced data.

Fig. 4.

Comparison of NeXtMD with other methods on AIP test set

Table 2.

The various metrics comparison of NeXtMD and other models

Model	AUC	ACC	Precision	Recall	F1-score	MCC
NeXtMD	0.8607	0.7996	0.8062	0.7134	0.7513	0.5884
TriStack	0.7900	0.8219	0.7900	0.7080	0.7496	0.5830
AIPStack	0.7650	0.8120	0.7680	0.6630	0.7020	0.5130
PPTPP_cls	0.7000	0.8120	0.7130	0.5240	0.6080	0.4000
PPTPP_prb	0.7500	0.8060	0.7710	0.5940	0.6820	0.4930
PPTPP_fus	0.6500	0.7890	0.7880	0.2920	0.4210	0.3210
TriNet	0.6950	0.7940	0.6870	0.5540	0.6030	0.3580

Among the comparison models, TriStack and AIPStack performed relatively well, yet their AUC (0.7900 and 0.7650, respectively) and F1-scores (0.7496 and 0.7020) still lagged behind NeXtMD. The PPTPP variants showed varying performance depending on the feature type, with PPTPP_prb performing best among them (AUC = 0.7500), but still falling short of NeXtMD in most metrics. TriNet, in contrast, exhibited the weakest performance across all metrics. It is worth noting that TriStack performs best on the ACC metrics and presents a more balanced advantage over the other models in the comparison.

Taken together, these results demonstrate the superior predictive capability and generalization ability of NeXtMD, especially in balancing precision and recall. Its consistent top-ranking performance across multiple metrics highlights its strong potential as a next-generation computational tool for AIP prediction.

NeXtMD learns discriminative features for accurate AIP classification

To visually and quantitatively assess the discriminative capacity of NeXtMD, we conducted unsupervised clustering analyses on both the original input features and the deep feature embeddings learned by the model. Specifically, we applied two widely used dimensionality reduction techniques: uniform manifold approximation and projection (UMAP) and t-distributed stochastic neighbor embedding (t-SNE).

As illustrated in Fig. 5A and C, the raw feature space revealed a high degree of overlap between AIP (red) and non-AIP (blue) samples, indicating limited separability of the input features. In contrast, the embeddings generated by NeXtMD show clear and consistent inter-class separation (Fig. 5B and D). Notably, NeXtMD transformed the previously entangled distributions into distinct, non-overlapping clusters with well-defined boundaries, especially along the second dimension in the UMAP and t-SNE plots.

Fig. 5.

Visualization of the test set's original features and the learned representations obtained by NeXtMD. Red points represent positive (AIP) samples and blue points represent negative (non-AIP) samples. A shows the t-SNE visualization of the original features obtained from the test set. B shows the t-SNE visualization of the learned features after NeXtMD was trained on the AIP samples. C shows the UMAP visualization of the original features obtained from the test set. D shows the UMAP visualization of the learned features after NeXtMD was learning the AIP database samples

To quantitatively support these observations, we computed average silhouette scores, which were significantly higher for NeXtMD-derived representations than for the raw features. Moreover, a Wilcoxon signed-rank test confirmed that this improvement in cluster separation was statistically significant (p < 0.001).

Together, these findings highlight NeXtMD’s ability to learn task-relevant, discriminative representations that facilitate better AIP classification. The consistency of results across three independent non-linear projection methods further demonstrates the robustness and stability of the learned feature space. This approach exemplifies the growing trend in translational bioinformatics, where model-driven representation learning, combined with traditional exploratory tools, contributes to a deeper understanding of biological data and supports more accurate downstream predictive tasks.

NeXtMD combines classifiers and features for predictive advantage

We conducted ablation experiments to investigate the mechanisms behind NeXtMD’s performance, targeting both ML components and sequence-derived feature descriptors. NeXtMD combines four ML classifiers and a deep residual network with four sequence-based features. We assessed the impact of removing individual classifiers or descriptors on performance.

In model ablation experiments, we removed each classifier—RF, XGBoost, LightGBM, and GBDT—and evaluated the performance changes. As shown in Fig. 6A and C, NeXtMD consistently outperformed ablated variants across metrics such as ACC, precision, recall, F1-score, and MCC. Small AUC increases when RF or XGBoost was removed likely resulted from decision boundary shifts rather than meaningful improvements. For instance, recall increased without RF and ACC improved without XGBoost (Table 3). However, these improvements were offset by declines in other metrics, showing reduced global performance. This confirms the essential role of each meta-classifier.

Fig. 6.

Comparison ablation test results to validate the contribution of ML algorithms and feature selection in NeXtMD models. A Histogram plots showing the performance of NeXtMD when each of the four ML models is ablated one at a time (retaining the other three). B Histogram plots showing performance when each of the four feature descriptors is ablated one at a time. C ROC curves of NeXtMD with one ML models removed at a time. D ROC curves of NeXtMD with one feature descriptor removed at a time

Table 3.

The various metrics comparison of NeXtMD and model ablation

Model	AUC	ACC	Precision	Recall	F1-score	MCC
NeXtMD	0.8607	0.7996	0.8062	0.7134	0.7513	0.5884
w/o RF	0.8152	0.7923	0.7761	0.7026	0.7054	0.5147
w/o XGBoost	0.8338	0.8022	0.7726	0.6712	0.7022	0.5234
w/o LightGBM	0.8367	0.7832	0.7813	0.6964	0.7037	0.5438
w/o GBDT	0.8233	0.7864	0.7714	0.6529	0.7014	0.5246

We also found that each sequence descriptor provides complementary, non-redundant information (Fig. 6B, D, and Table 4). The performance drop upon feature removal underscores the importance of combining multiple descriptors. Together with the model ablation results, this confirms that both the ensemble classifier and multi-descriptor features are crucial for maximizing NeXtMD’s predictive performance in AIP classification.

Table 4.

The various metrics comparison of NeXtMD and model ablation

Model	AUC	ACC	Precision	Recall	F1-score	MCC
NeXtMD	0.8607	0.7996	0.8062	0.7134	0.7513	0.5884
w/o DDE	0.8438	0.7763	0.7452	0.7124	0.7046	0.5396
w/o CKSAAP	0.8165	0.8037	0.7657	0.6928	0.7166	0.5581
w/o PP16	0.8152	0.7864	0.7791	0.7117	0.7145	0.5276
w/o ACH	0.7703	0.8107	0.7816	0.7063	0.7170	0.5734

In addition, to verify the necessity of ResNeXt stacking, we tried logistic regression and MLP as the second layer of stacking for comparison (Fig. S7–9). The NeXtMD model built with ResNeXt demonstrates a clear advantage in this task.

NeXtMD exhibits excellent generalizability across diverse AIP datasets

To further evaluate the predictive efficacy and generalization capability of the NeXtMD model in AIP classification, we conducted external test set validations by replacing the test data with previously unused samples. Specifically, three independent external test sets were constructed by selecting the complementary subsets (i.e., non-overlapping samples) of the DeepAIP, BertAIP, and AIP-DeepEnC datasets relative to our original dataset, ensuring that no overlap with the training data occurred. This strategy enabled rigorous out-of-distribution testing of the model [37]. We observed that despite substantial differences in sequence length distributions across the three processed external validation datasets, sequence lengths predominantly clustered within the range of 15–20 amino acids. Furthermore, the length distribution trends of positive and negative samples were generally consistent within each external validation set (Fig. S10–12).

As shown in Fig. 7 and Table 5, despite slight performance fluctuations across external datasets, NeXtMD consistently maintained strong generalization and stability. On the DeepAIP-complement test set, NeXtMD achieved an AUC of 0.9374 and ACC of 0.9238, notably higher than on the internal test set (AUC = 0.8607, ACC = 0.7996). The model also showed strong recall (0.7969) on DeepAIP and competitive performance on AIPs-DeepEnC (AUC = 0.7938, F1 = 0.6882). Although precision was lower on BertAIP (0.4881), recall remained high (0.8656), suggesting some trade-off. Overall, these results highlight NeXtMD’s robust external validity and its ability to generalize beyond the training distribution.

Fig. 7.

Performance analysis of NeXtMD on external test sets. A ROC curves of NeXtMD evaluated on the DeepAIP-derived external test set containing unfamiliar AIPs. B ROC curves of NeXtMD evaluated on the BertAIP-derived external test set containing unfamiliar AIPs. C ROC curves of NeXtMD evaluated on the AIPs-DeepEnC-derived external test set containing unfamiliar AIPs. D Radar plot omparing performance on the internal test set versus external test sets across evaluation metrics

Table 5.

The various metrics comparison of the initial AIP test set and other external validation test sets

Dataset	AUC	ACC	Precision	Recall	F1-score	MCC
NeXtMD	0.8607	0.7996	0.8062	0.7134	0.7513	0.5884
DeepAIP	0.9374	0.9238	0.6097	0.7969	0.5766	0.4026
BertAIP	0.7128	0.7412	0.4881	0.8656	0.6242	0.4206
AIPs-DeepENC	0.7938	0.7150	0.6022	0.7857	0.6882	0.4245

Given the high complementarity observed between the DeepAIP-derived AIPs and the original AIP test set, we integrated the two to construct an augmented AIP dataset [38]. This expanded dataset was subsequently re-partitioned into new training and testing subsets for retraining and evaluation. As shown in Fig. 8 and Table 6, the re-trained NeXtMD achieved substantially improved performance across all evaluation metrics. Specifically, the AUC increased from 0.8607 to 0.9783, ACC from 0.7996 to 0.9317, and F1-score from 0.7513 to 0.9200. Notably, recall rose sharply from 0.7134 to 0.9502, and MCC improved from 0.5884 to 0.8618, indicating a significant gain in both sensitivity and overall predictive reliability.

Fig. 8.

Effectiveness of AIP dataset enhancement based on DeepAIP exogenous AIPs and comparison with initial AIP dataset. A ROC curves of NeXtMD trained and evaluated on the enhanced AIP dataset. B Radar plot comparing the initial AIP dataset and the enhanced dataset across evaluation metrics. C Histogram comparing the distribution of evaluation metrics between the initial and enhanced datasets

Table 6.

The various metrics comparison of AIP dataset and enhanced dataset

Dataset	AUC	ACC	Precision	Recall	F1-score	MCC
Origin	0.8607	0.7996	0.8062	0.7134	0.7513	0.5884
Enhanced	0.9783	0.9317	0.8916	0.9502	0.9200	0.8618

These results demonstrate that incorporating high-quality external data not only enhances the model’s discriminative capacity but also significantly improves its generalization. The improved metrics underscore the potential of NeXtMD to support more accurate and robust AIP prediction in broader peptide-based bioinformatics applications.

NeXtMD demonstrates strong transferability across peptide tasks

To explore the transferability and broader applicability of NeXtMD, we evaluated its performance on the prediction of antimicrobial peptides (AMPs), which are functionally and mechanistically related to AIPs. In clinical contexts, inflammatory responses often co-occur with bacterial infections, and recent studies suggest that AIPs and AMPs can exert synergistic therapeutic effects, enhancing treatment efficacy [39]. Leveraging this biological relevance, we tested whether a model trained solely on AIP data could be effectively applied to AMP prediction.

The AMP dataset was preprocessed and partitioned using the same pipeline as the AIP dataset before being fed into NeXtMD for evaluation [40, 41]. Comparative performance analyses against two representative models, TriStack and TriNet, revealed that while NeXtMD did not outperform all competitors across every individual metric, it achieved the highest overall predictive performance. NeXtMD reached an AUC of 0.9706, substantially surpassing the baseline models (Fig. 9 and Table 7).

Fig. 9.

Validation of NeXtMD’s model transferability to AMP prediction. A ROC curves of NeXtMD evaluated on the AMP dataset. B Radar chart comparing NeXtMD, TriStack, and TriNet on the AMP dataset. C Histogram comparing the performance of NeXtMD, TriStack, and TriNet on the AMP dataset

Table 7.

The various metrics comparison of NeXtMD, TriStack, and TriNet on AMPs

Model	AUC	ACC	Precision	Recall	F1-score	MCC
NeXtMD	0.9706	0.9294	0.9158	0.9055	0.9160	0.8189
TriStack	0.9247	0.9220	0.9170	0.9340	0.9250	0.8440
TriNet	0.8687	0.8891	0.8876	0.9182	0.9070	0.7864

These findings confirm that NeXtMD possesses strong cross-domain transferability and can generalize effectively to related peptide classification tasks without requiring model retraining. This demonstrates its practical value in real-world biomedical scenarios and highlights its potential for clinical and translational applications, especially in the joint modeling of functionally associated peptides such as AIPs and AMPs [42]. The ability of NeXtMD to adapt to different peptide types, such as AIPs and AMPs, showcases its versatility and makes it a promising multi-functional peptide prediction tool. This capability enables NeXtMD to efficiently model diverse peptide functions, which is crucial for developing tools that can predict and characterize a wide range of peptides with various biological roles.

Discussion

In this study, we developed NeXtMD, a novel computational framework that integrates traditional machine learning techniques with a ResNeXt-based deep learning module to enhance the prediction accuracy of AIPs. By employing diverse sequence-derived descriptors—including DDE, CKSAAP, PP16, and ACH—NeXtMD effectively captures both local and global information from peptide sequences. This comprehensive feature extraction is essential for delineating the subtle biochemical variations that govern peptide functionality.

The ResNeXt architecture is particularly well-suited for modeling peptide sequences due to its ability to efficiently handle complex, hierarchical data structures. With its grouped residual blocks and skip connections, ResNeXt can learn deep, multi-level representations that capture intricate patterns and dependencies in peptide sequences, including long-range interactions between amino acids. This architecture’s capacity to model complex non-linear relationships is crucial for identifying subtle but critical features that govern peptide activity, contributing to the observed performance gains.

The ensemble approach in NeXtMD leverages four advanced classifiers (XGBoost, Random Forest, LightGBM, and GBDT) through a stacking strategy validated by fivefold cross-validation [43]. Each classifier contributes probabilistic predictions that form a meta-feature vector subsequently refined by the deep learning module. The ResNeXt component, designed with grouped residual blocks and skip connections, further enhances feature representation by uncovering complex non-linear relationships in the data. This fusion strategy significantly improves performance across various metrics such as ROC AUC, accuracy, precision, recall, F1 score, and MCC.

Biologically, the integrated features provide critical insights into the molecular underpinnings of AIP activity. For instance, the DDE and CKSAAP descriptors quantify local sequence patterns, reflecting the underlying dipeptide distributions, while the physical—chemical features—such as hydrophobicity, charge, polarity, and secondary structure propensities—offer clues regarding peptide solubility, membrane interaction, and overall stability. Furthermore, ACH features capture the global dependency of these properties along the peptide chain. Such detailed characterization not only improves prediction accuracy but also aids in the biological interpretation of peptide efficacy, thereby facilitating hypothesis generation for experimental validation [44, 45]. Supporting this perspective, clinical evidence from chronic kidney disease demonstrates that statin therapy significantly reduces inflammation markers such as C-reactive protein (CRP) and high-sensitivity CRP (hs-CRP), highlighting the practical implications of pharmacological interventions targeting inflammatory pathways as well as the potential benefits of analyzing the molecular characterization of AIP activity in the treatment of inflammation [46].

Despite these advances, there are notable limitations. The current dataset, composed exclusively of peptides shorter than 40 amino acids, may not represent the full diversity of AIPs. It remains uncertain whether longer peptides might also exhibit significant anti-inflammatory effects, underscoring the need for future validation on larger and more diverse datasets. Additionally, although the ensemble of ML classifiers and the integration with deep learning provide robust performance, the overall computational complexity remains high. The rich but high-dimensional feature space may introduce redundancy, potentially impeding efficiency. To address this, future work could incorporate feature selection methods such as minimal redundancy maximal relevance (mRMR) or analysis of variance (ANOVA) to streamline input dimensions.

Looking ahead, further improvements in model performance may be achieved by exploring advanced deep learning architectures, including Transformer-based models and graph neural networks, which offer enhanced capabilities for capturing long-range dependencies and complex structural relationships [46, 47]. Moreover, the substitution of conventional normalization layers with dynamic tanh activation functions could improve both model fitting and computational efficiency [35, 48]. Optimizing ensemble strategies and refining gradient-based criteria also present promising avenues for further enhancements [49, 50].

From a translational perspective, the accurate computational identification of AIPs is crucial for accelerating drug discovery processes and enabling the development of novel therapeutics for inflammatory diseases. For example, targeting Bromodomain and Extra-Terminal (BET) proteins, which regulate critical inflammatory gene expression pathways, has shown considerable therapeutic potential in both preclinical and clinical studies. Therefore, integrating insights from BET inhibitor research may provide valuable directions for the future development of personalized anti-inflammatory therapeutics [51]. Additionally, the transcription factor Nrf2, known for its regulatory role in ferroptosis—a form of iron-dependent cell death—has been implicated in mitigating oxidative stress and inflammation in neurological diseases. Exploring Nrf2-mediated pathways could thus offer complementary strategies to advance anti-inflammatory interventions in related therapeutic areas [52, 53]. The biological interpretability of NeXtMD, facilitated by its integrated feature extraction framework, links computational predictions directly with underlying biochemical mechanisms—a critical aspect in clinical decision-making. The demonstrated transferability of the model to related peptide types, such as AMPs, further underscores its potential utility across diverse biomedical applications. The precise prediction of antimicrobial peptides by NeXtMD could offer novel therapeutic approaches for combating invasive fungal infections, addressing critical issues like antifungal drug resistance and enhancing patient outcomes [54].

In summary, while NeXtMD exhibits significant improvements in predicting AIPs by integrating multifaceted feature descriptors with a robust ensemble and deep learning strategy, future research must address dataset limitations and computational challenges. Continued refinement of both feature selection and model architectures will not only enhance predictive performance but also broaden the practical applicability of this approach in drug discovery and clinical therapies for inflammation-related conditions.

Conclusions

In this study, we introduced a novel computational framework, NeXtMD, which effectively integrates traditional ML algorithms with the ResNeXt deep learning module to significantly enhance the accuracy of AIP identification. NeXtMD innovatively utilizes four complementary sequence-derived descriptors—DDE, CKSAAP, PP16, and ACH—to systematically capture a comprehensive range of local and global features from short peptides, thereby addressing the limitations of existing methods that rely on simplistic feature extraction and single classifier architectures. Benchmark experiments demonstrate that NeXtMD outperforms current state-of-the-art predictors, exhibiting exceptional predictive performance and robustness in AIP classification. Ablation studies, incorporating both model and feature-level assessments, validate the crucial roles of the four selected features, the five ML classifiers, and the ResNeXt module within the framework. Furthermore, evaluations on an independent, unseen test set confirm the generalizability of NeXtMD, while additional results indicate that expanding the training dataset can further enhance its performance. Lastly, tests conducted on an AMP dataset effectively demonstrate the transferability of NeXtMD and its potential applicability to the prediction of other peptide types.

Methods

Dataset processing

In this study, we leveraged a dataset composed exclusively of AIPs. The data were sourced from the Immune Epitope Database (IEDB), a publicly accessible repository offering epitope data and predictive tools [51]. The dataset was constructed by extracting experimentally validated linear peptides and epitopes, classified into positive and negative samples. Positive samples consist of those peptides that have been empirically demonstrated, in human or murine T cell assays, to induce anti-inflammatory cytokines such as IL-10, IL-4, IL-13, IL-22, TGF-β, and IFN-α/β. In contrast, negative samples were defined as those peptides that were verified not to stimulate these cytokines. Ultimately, a curated set of 4,194 non-redundant AIP samples was assembled for model training and evaluation (Table 8). In addition, the dataset was divided into training and testing subsets in a 5:1 ratio. Subsequent analyses confirmed that the sequence length distribution and amino acid abundance of both positive and negative samples were well balanced in the processed AIP dataset. Moreover, a fivefold cross-validation approach was employed on the training data to mitigate potential overfitting risks due to excessive model training.

Table 8.

The AIP datasets used in this study

Data	AIP	non-AIP	Total
Training set	1,398	2,097	3,495
Test set	280	419	699

Feature selection

NeXtMD is a stacked computational framework integrating machine learning and deep learning strategies, consisting of two modules, specifically designed for the identification of AIPs. The overall architecture of the NeXtMD model can be divided into four primary steps: (1) feature extraction, (2) feature learning using four ML classifiers, (3) feature re-encoding by ResNeXt, and (4) functional classification of AIPs on the test set.

After completing the preparation of the AIP dataset, to implement the feature learning of the four ML classifiers of the NeXtMD model, we carried out the feature extraction based on the above AIP dataset. Specifically, NeXtMD initially extracts multiple functional sequence descriptors, including DDE, CKSAAP, PP16, and ACH. Among these, the DDE descriptor is represented as a 2400-dimensional vector capturing local dipeptide patterns, while CKSAAP is encoded into a 400-dimensional vector to characterize local sequence information. The PP16 descriptor, represented by a 28-dimensional vector, encodes comprehensive physicochemical properties based on statistical classification of amino acids. Simultaneously, the ACH descriptor, represented by a 5-dimensional vector, generates autocorrelation-based hydrophobicity features computed from amino acid hydrophobicity scales. Subsequently, these four descriptor vectors are integrated and fed into the dual-module stacked modeling framework for downstream analysis.

The DDE descriptor quantifies deviations between the observed and theoretically expected occurrences of each dipeptide (two consecutive amino acids) in protein sequences based on predefined amino acid usage frequencies. By evaluating the differences between observed frequencies and their theoretical expectations, DDE effectively captures characteristics of local sequence structures. Specifically, dipeptide frequencies are counted to form a 400-dimensional feature vector suitable for subsequent classification and predictive analysis.

The DDE formulas are as follows:

$$E_{\mathit{ij}}=p_i{p}_j(L-1)$$ 1

$${\sigma }_{\mathit{ij}}=\sqrt{\left(L,-,1\right)p_i{p}_j\left(1,-,p_i,p_j\right)}$$ 2

$${\mathit{DDE}}_{\mathit{ij}}=\frac{N_{\mathit{ij}}^{\mathit{obs}}-E_{\mathit{ij}}}{{\sigma }_{\mathit{ij}}}$$ 3

The CKSAAP descriptor systematically extracts amino acid pair combinations separated by a fixed interval k (where k ranges from 0 to 5) by traversing all possible amino acid pairs within the protein sequence [55]. It calculates the occurrence frequency for each amino acid pair at these intervals and normalizes these frequencies by the total number of possible combinations at the corresponding interval. This strategy captures correlations spanning both local and long-range positions, thereby generating a comprehensive 2400-dimensional feature vector (6 intervals × 20 × 20 amino acid pairs) for subsequent input into the NeXtMD model.

The CKSAAP formulas are as follows:

$$N_{\mathit{ij}}^k=\text{Counts of amino acid pairs}(i,j)\text{with interval}k$$ 4

$$f_{\mathit{ij}}^k=\frac{N_{\mathit{ij}}^k}{L-k-1}$$ 5

$$\text{The feature vector is the splice of all}{f}_{\mathit{ij}}^k\text{under}k=\text{0,1},...,5$$ 6

The PP16 descriptor comprehensively characterizes the global physicochemical properties of protein sequences by encoding each amino acid with multiple physicochemical indices [56], including hydrophobicity, charge, polarity, α-helix propensity, β-sheet propensity, and volume. Specifically, it calculates both the sum and mean values of these indices across the entire sequence. Hydrophobicity is described using 16 dimensions, whereas charge, polarity, α-helix propensity, β-sheet propensity, and volume each contribute two dimensions, resulting in a 28-dimensional feature vector. This descriptor thus provides essential global sequence information effectively utilized by the NeXtMD model.

The PP16 formulas are as follows:

$$S_x^{(k)}=\sum _{i=1}^L{x}_i^{(k)}$$ 7

$${\mu }_x^{(k)}=\frac{S_x^{(k)}}{L}$$ 8

where $x_i^{\left(k\right)}$ denotes the value of the $i\text{th}$ amino acid in the sequence t the $k\text{th}$ physicochemical index, and $L$ is the length of the sequence.

The ACH descriptor quantifies hydrophobicity characteristics within protein sequences using a predefined hydrophobicity scale. Specifically, this approach computes autocorrelation values at multiple lag intervals (lag ranging from 1 to 5) by first centering the hydrophobicity values of amino acids relative to the sequence mean, and then calculating the average product of these centered hydrophobicity scores between each amino acid and subsequent residues at defined lags. Consequently, this descriptor captures local dependency patterns in sequences, resulting in a 5-dimensional feature vector that provides intrinsic physicochemical sequence information for integration into the NeXtMD model.

The ACH formulas are as follows:

$$\mu =\frac{1}{L}\sum _{i=1}^L{x}_i$$ 9

$$\text{AC}\left(k\right)=\frac{1}{L-k}\sum _{i=1}^{L-k}(x_i-\mu )(x_{i+k}-\mu )(k=1,2,\cdots ,5)$$ 10

Model construction

After completing the feature extraction described above, we carried out the construction and implementation of NeXtMD. This module integrates four ML meta-classifiers for initial classification of peptide sequences. Specifically, the RF classifier aggregates predictions from an ensemble of decision trees, whereas XGBoost, LightGBM, and GBDT classifiers are based on gradient boosting frameworks, iteratively refining predictions through sequential addition of base learners to minimize predictive error. Following training and prediction via fivefold (K = 5) cross-validation, each peptide sequence obtains four distinct prediction scores, which subsequently serve as meta-features for further integration into the downstream modeling process.

The ML formulas are as follows:

$$\text{RF}:f_{\text{RF}}\left(x\right)=\frac{1}{N}\sum _{i=1}^N{h}_i(x)$$ 11

$$\text{Gradient boosting method}:f\left(x\right)=\sum _{t=1}^T{f}_t(x)$$ 12

This module initially constructs a four-dimensional meta-feature vector from the prediction scores generated by the four base classifiers, which is subsequently expanded into a 64-dimensional representation through a fully connected layer. The expanded vector is then input into multiple stacked ResNeXt blocks. Within each ResNeXt block, the input is equally partitioned into four groups, each independently undergoing a sequence of operations including a Dense layer (with 64 units and ReLU activation), BatchNormalization layer, Dropout regularization (dropout_rate = 0.3), and a second Dense layer. The outputs from these groups are summed with a linearly transformed skip-connection input, followed by ReLU activation, facilitating effective residual learning. Each ResNeXt block uses a cardinality of 4, meaning that each block has 4 branches, with 16 units per branch. After passing through five consecutive ResNeXt blocks, deep hierarchical features are extracted. Finally, a fully connected layer with sigmoid activation performs binary classification, enabling precise functional annotation of peptide sequences.

The DL formulas are as follows:

$$z=\text{ReLU}(W_{0}x+b_0)$$ 13

$$x_{\text{block}}=\text{ReLU}\left(\sum _{i=1}^C,F_i,\left(x_i\right),+,W_{\text{skip}},x\right)(C=4)$$ 14

$$\widehat{y}=\sigma (W_0{x}_{\text{final}}+b_0)$$ 15

The model evaluation indicator

The evaluation metrics used in this study include ACC, precision, recall, F1-score, and MCC, and AUC [54, 57–60]. These metrics were computed as follows: true positives (TP) represent correctly classified positive samples, true negatives (TN) represent correctly classified negative samples, false positives (FP) represent negative samples incorrectly classified as positive, and false negatives (FN) represent positive samples incorrectly classified as negative [61, 62]. The F1-score, calculated as the harmonic mean of precision and recall, provides a more comprehensive performance evaluation, while MCC further integrates all possible combinations of TP, FP, TN, and FN. Therefore, we propose accuracy, F1-score, and MCC as more robust evaluation metrics compared to using precision or recall alone. Moreover, ROC AUC values were further assessed, with the ROC curve defined by plotting false positive rate (FPR) on the x-axis against true positive rate (TPR) on the y-axis.

The model evaluation indicator formulas are as follows:

$$\text{Precision}=\frac{\text{TP}}{\text{TP}+\text{FP}}$$ 16

$$\text{Recall}=\frac{\text{TP}}{\text{TP}+\text{FN}}$$ 17

$$\text{ACC}=\frac{\text{TP}+\text{TN}}{\text{TP}+\text{TN}+\text{FP}+\text{FN}}$$ 18

$$F1-\text{score}=2\times \frac{\text{Precision}\times \text{Recall}}{\text{Precision}+\text{Recall}}$$ 19

$$\text{MCC}=\frac{\text{TP}\times \text{TN}-\text{FP}\times \text{FN}}{\sqrt{(\text{TP}+\text{FP})\times (\text{TP}+\text{FN})\times (\text{TN}+\text{FP})\times (\text{TN}+\text{FN})}}$$ 20

$$\text{AUC}=\underset{-\infty }{\stackrel{\infty }{\int }}\text{TPR}(t)\text{dFPR}(t)$$ 21

Abbreviations

AIPs

Anti-Inflammatory Peptides

ML

Machine Learning

DL

Deep Learning

ResNeXt

Residual Networks neXt

SVM

Support Vector Machine

XGBoost

eXtreme Gradient Boosting

CNN

Convolutional Neural Networks

LSTM

Long Short-Term Memory

DDE

Dipeptide Deviation Encoding

CKSAAP

Composition of K-Spaced Amino Acid Pairs

PP16

Physicochemical Properties grouped into 16 classes

ACH

Auto-Correlation of Hydrophobicity

LightGBM

Light Gradient Boosting Machine

GBDT

Gradient Boosting Decision Trees

UMAP

Uniform Manifold Approximation and Projection

t-SNE

t-distributed Stochastic Neighbor Embedding

PCA

Principal Component Analysis

AMPs

AntiMicrobial Peptides

CRP

C-Reactive Protein

hs-CRP

high-sensitivity CRP

mRMR

minimal Redundancy Maximal Relevance

ANOVA

ANalysis Of VAriance

BET

Bromodomain and Extra-Terminal

IEDB

Immune Epitope DataBase

TP

True Positives

TN

True Negatives

FP

False Positives

FN

False Negatives

FPR

False Positive Rate

TPR

True Positive Rate

Authors’ contributions

Conceptualization: H.L.2, F.D., and J.F.; Data curation: C.X. and H.Y.; Methodology: C.X., H.L.1 and Y.W.; Writing—original draft: C.X. and X.L.; Writing—review and editing: H.L.2, and F.D. All authors read and approved the final manuscript.

Funding

This work was supported by the National Natural Scientific Foundation of China (62402089, U24A20789), Sichuan Science and Technology Program (2025ZNSFSC1465), Undergraduate Elite Experience Program of University of Electronic Science and Technology of China (2024GFTY018), and China Postdoctoral Science Foundation (2023TQ0047, GZC20230380).

Data availability

All data generated or analysed during this study are included in this published article, its supplementary information files and publicly available repositories. The sources of the analysed datasets contain BertAIP, DeepAIP, AIP-DeepEnC, TriStack, and the relevant information is included in the public repository. In addition, the data and code used for training and testing the NeXtMD model in this study are available in the Zenodo repository (10.5281/zenodo.15667176) and Figshare repository (10.6084/m9.figshare.29322215.v1).

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.