A stacked ensemble classifier for the discovery of prokaryotic efflux proteins based on sequence information

Abstract

Background

Efflux proteins (EPs) are key transporters responsible for expelling antibiotics and other harmful substances from bacterial cells, thus contributing significantly to antimicrobial resistance. As such, the development of a reliable and efficient computational method for identifying EPs in bacteria and archaea is of utmost importance. However, current computational methods for EP identification are limited and suffer from various performance and utility constraints.

Results

This study presents EPDiscovery, a stacking-based ensemble classifier designed to predict prokaryotic EPs directly from their amino acid sequence. EPDiscovery employs the MLP algorithm to integrate the predictions of fifteen baseline classifiers. The test results demonstrated that EPDiscovery accurately identified EPs, achieving AUROC, accuracy, F-value, and Matthews correlation coefficient scores of 0.970, 0.926, 0.927, and 0.852, respectively. Additionally, EPDiscovery outperformed the existing representative method and demonstrated consistent and satisfactory performance, with AUROC scores exceeding 0.9 across three previously proposed datasets. Feature importance analysis using local interpretable model-agnostic explanations further confirmed EPDiscovery’s ability to discern the characteristics of EPs from protein sequence information.

Conclusions

EPDiscovery offers a reliable alternative for annotating efflux proteins in large-scale bacterial genome studies, thus accelerating the discovery of potential EPs and contributing to the advancement of antimicrobial resistance research.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12864-025-12039-1.

Introduction

Efflux proteins (EPs), also known as efflux pump proteins, are transporter molecules that actively expel a variety of substrates from cells to the external environment, and are present in both prokaryotic and eukaryotic cells [1]. EPs are integral to numerous physiological processes, including maintaining the appropriate potential and pH gradient across the cell membrane, intercellular signaling, microbial virulence, and the removal of unwanted metabolites and toxic substances from the cell [2]. Therefore, they play a vital role in maintaining cellular homeostasis and responding to environmental challenges [3–5]. One of the key mechanisms by which bacteria acquire resistance to antibiotics and bactericides is through efflux pump activity. This resistance arises when a substance structurally resembles the pump’s natural substrates or when the pump has broad substrate selectivity [2]. A deeper understanding of efflux pumps is essential for elucidating bacterial resistance mechanisms and for designing new antimicrobial strategies. Hence, accurately identifying EPs is of paramount importance.

Prokaryotic EPs are categorized into five major families based on their amino acid sequence and energy source for substrate export [6]. These families include ATP-binding cassette (ABC) transporter superfamily, multidrug and toxic compound extrusion (MATE) family, major facilitator superfamily (MFS), resistance-nodulation cell division (RND) family, and small multidrug resistance (SMR) family. While most of these families utilize the proton motive force for energy, the ABC family is unique in using ATP hydrolysis for substrate export [7]. Advances in genomic technologies have significantly expanded the identification of new members within these families [7]. However, traditional experimental approaches for EP identification are time-consuming, laborious, and expensive, making them unsuitable for large-scale studies. Computational methods, on the other hand, offer a more efficient and cost-effective solution for identifying EPs across large protein datasets, and can also provide insights into the functional study of efflux pumps. Consequently, there is a pressing need for computational approaches that leverage sequence information to identify EPs.

Machine learning (ML) techniques have gained prominence as powerful tools for identifying significant biological sequences based solely on sequence information [8]. There are currently ML-based methods to predict or analyze EPs, which provide a valuable complement to experimental studies. For instance, Taju et al. used convolutional neural networks combined with position-specific scoring matrices to identify EPs in membrane and transport proteins [9]. Pandey et al. established a support vector machine (SVM) model to classify bacterial EPs and determine their corresponding families [10]. More recently, Taju et al. proposed a method for identifying EPs using contextualized word embeddings from Bidirectional Encoder Representations from Transformers (BERT) coupled with an SVM classifier [11]. While these methods have demonstrated reasonable prediction accuracy, there is potential for improvement. These early methods were based on smaller datasets, limiting their predictive power. As the number of identified EP sequences continues to grow, there is an opportunity to develop a more comprehensive and accurate predictive classifier. Furthermore, these methods utilize only a few specific ML algorithms, leaving many widely used algorithms underexplored. For instance, ensemble learning [12]—an ML technique that integrates multiple learning algorithms to achieve superior predictive performance—remains unexploited. Thus, these methods may lack robustness in certain scenarios, which highlights the need for more powerful computational approaches.

In this study, we introduce EPDiscovery, an ensemble learning-based classifier designed for the accurate prediction of prokaryotic EPs. EPDiscovery was developed using a curated EP dataset sourced from current, up-to-date databases. It utilized a diverse range of protein features, along with commonly used feature selection methods and ML algorithms, to construct baseline classifiers. Then the stacking strategy was employed to integrate the baseline classifiers. The performance of EPDiscovery was validated using independent and external test datasets and subsequently compared with other approaches, as well as an existing EP method. Additionally, the Local interpretable model-agnostic explanations (LIME) method was employed to provide interpretability for the predictions made by EPDiscovery. Our results demonstrate that EPDiscovery is an effective tool for identifying EPs from protein sequence data, offering significant potential for advancing research in this area.

Methods

Dataset Preparation

The details of the overall study are shown in Fig. 1. A carefully curated dataset is essential to developing an effective and reliable classifier. While some EP datasets have been built in prior studies [10, 11], the growing number of discovered EP sequences prompted us to reconstruct a dataset using the most recent database. EP sequences were derived from the UniProt database (https://www.uniprot.org/), released in May 2024. A protein was considered as an EP if it was annotated as belonging to one of the following families in the Transport Classification Database (TCDB, http://www.tcdb.org/): ABC transporter superfamily, MATE, MFS, RND, and SMR. The resulting sequences were filtered to retain only the reviewed and complete sequences from archaea and bacteria. While it is possible that some newly discovered EPs have not yet been incorporated into these databases, this approach prioritizes the reliability and accuracy of the sequence set used for model development. Furthermore, the combination of high-quality, manually curated protein sequences from UniProt with the domain-specific annotations provided by TCDB, a specialized resource for transport proteins, helps to mitigate potential biases in database annotations and supports the robustness of the classifier. Control sequences were selected by removing any EP sequences from the reviewed and complete sequences of archaea and bacteria in the UniProt database. To ensure the reliability of the datasets, the protein sequences containing ambiguous amino acids (e.g., ‘X’, ‘Z’, ‘B’, ‘J’, ‘O’, ‘U’) or stop codons (‘*’) were excluded. To minimize potential biases in subsequent analyses, the CD-HIT [13] program was then used to eliminate protein sequences with more than 40% identity. The threshold was chosen in accordance with previous studies [10, 14], while ensuring the retention of a sufficiently large set of representative protein sequences. Control sequences that clustered with EPs were also removed to prevent ambiguity, while representative sequences from each remaining cluster were retained. These procedures resulted in a curated dataset containing 2,148 EPs and 13,140 controls.

Fig. 1.

Schematic flowchart of the development of the EPDiscovery. It consists of dataset preparation, feature extraction, feature selection, baseline classifier construction, and the final ensemble classifier development

Given the extreme imbalance between EPs and controls in the curated dataset, classifiers built directly on it are prone to bias toward the prevalent labels. To mitigate bias, an equal number of EP and control sequences were randomly selected to compose datasets for training and evaluation. As illustrated in Figure S1, 1,718 (80%) of the EP sequences were randomly selected, along with an equal number of control sequences, to form the training dataset (n = 3,436). To maximize the use of control data, control sequences were sampled seven times without replacement to generate seven training datasets (denoted D1 to D7), each containing the same EP sequences. The remaining 430 EP sequences were divided into two equal parts, with an equal number of control sequences, forming the validation (n = 430) and test (n = 430) datasets (Table S1). This approach ensures that the number of EPs and controls in any dataset is balanced, with no overlap of sequences across the training, validation, and test datasets. Here, the training datasets were used to build baseline classifiers, the validation dataset was used to assess their performance and construct the ensemble classifier, and the test dataset was used to evaluate the ensemble classifier. To further verify the classifier, we obtained three previously proposed datasets as additional test datasets. Two of these are independent datasets from Taju2021 [11], comprising 113 EPs and 1,084 membrane proteins (referred to here as external dataset 1), and 113 EPs and 454 transport proteins (referred to here as external dataset 2). As the test dataset from Pandey2020 [10] was unavailable, we acquired the entire dataset from this work (referred to here as external dataset 3), which includes 210 EPs and 1,132 controls.

Protein feature extraction

The amino acid sequence of a protein encodes essential information that governs its properties, such as composition, sequence patterns, and physicochemical characteristics [15]. Since various protein properties can contribute to EP identification, this study integrated multiple protein features to select the optimal combination for EP identification. Using the iFeature program [16], a comprehensive profile was extracted, consisting of 17 protein descriptors and encompassing 2,726 sequence features. The descriptors are amino acid composition [17] (AAC), amphiphilic pseudo-amino acid composition [18, 19] (APAAC), composition of k-spaced amino acid group pairs [20] (CKSAAGP), composition [21] (CTDC), distribution [21, 22] (CTDD), transition [21, 22] (CTDT), conjoint triad [23] (CTriad), dipeptide deviation from expected mean [24] (DDE), dipeptide composition [17] (DPC), grouped amino acid composition [25] (GAAC), geary [26], grouped tripeptide composition [16, 17] (GTPC), moran [27], normalized Moreau-Broto [28] (NMBroto), pseudo-amino acid composition [18, 19] (PAAC), quasi-sequence-order descriptors [29] (QSOrder), and sequence-order-coupling number [30] (SOCNumber). Their relevance to EP classification was assessed using ROC analysis across the training, validation, and test datasets.

Classifier construction

A prediction pipeline was developed to determine whether a protein sequence is an EP. This pipeline employed a ten-fold cross-validation approach, where, in each iteration, 10% of the data was withheld for testing, and the remaining 90% was used for hyperparameter tuning, feature selection, and classifier training. In the pipeline, the first step was data standardization, employing the Z-score method to remove the mean and scale the data to unit variance. The second step, feature filtering, utilized the F-test to select features that distinguish between EP and control sequences. The third step, feature selection, involved nine methods, grouped into three categories: least absolute shrinkage and selection operator (LASSO), selecting features from model (SFM), and recursive feature elimination (RFE). Among these, SFM uses an ML algorithm to select features based on their importance or contribution to the model, while RFE identifies the most important features by recursively removing less important features and building a model on the remaining ones. Since both the SFM and RFE rely on an estimator, extra-trees (ET), logistic regression (LR), random forest (RF), and XGBoost (XGB) algorithms are employed as estimators respectively. To address potential redundancy, we implemented feature filtering and selection procedures prior to model construction, thereby reducing dimensionality. The classification step exploited 11 classification algorithms, including AdaBoost (AB), CatBoost (CB), gradient boosting (GB), k-nearest neighbors (KNN), light gradient boosting machine (LGB), multi-layer perceptron (MLP), support vector machine (SVM), ET, LR, RF, and XGB. For each possible combination of nine feature selection methods, and 11 classification algorithms, we trained the models sequentially across seven training datasets, yielding a total of 693 distinct models. As a classification task, the model trained in this study is referred to as a classifier, named according to the order of “training dataset - feature selection method - classification algorithm”.

The trained classifiers that achieved AUROC scores of 0.95 or higher were evaluated on the validation dataset. For the classifiers with AUROC scores ≥ 0.95, the Max-Relevance and Min-Redundancy (mRMR) [31] method was employed to select the baseline classifiers. mRMR is a feature selection technique designed to find features that have the highest relevance to the target variable while minimizing redundancy among features [32]. This study treated the classifier predictions as features to implement classifier selection. Finally, the stacking strategy was applied to integrate the baseline classifiers and create the ensemble classifier. In this process, both the number of baseline classifiers and the classification algorithm were optimized to attain the best performance.

Classifier evaluation

Several commonly used evaluation metrics were employed to assess the performance of the selected features and classifiers for EP prediction. These metrics include accuracy (ACC), F-value, Matthews’ correlation coefficient (MCC), sensitivity (SN), specificity (SP), and area under the receiver operating characteristic curve (AUROC). Among these, AUROC is particularly valuable as it accounts for the trade-off between sensitivity and specificity, offering a more comprehensive assessment of classifier performance across different threshold settings. Thus, AUROC was chosen as the primary metric for performance comparison in this study. The remaining metrics were calculated using the following formulas:

where TP, TN, FP, and FN represent the numbers of true positives, true negatives, false positives, and false negatives, respectively.

Classifier interpretation

LIME is a widely used technique that provides interpretable and faithful explanations of classifier predictions by learning an interpretable model locally around the prediction [33]. In this study, classifier interpretability was achieved using the LIME framework. Local regression coefficients quantified each feature’s contribution to EP prediction, with positive coefficients favoring EP classification and negative coefficients opposing it. A global explanation was derived via the submodular pick method, and the top 20 features with the highest absolute coefficients were identified as major contributors to predictive accuracy.

Results

Performance of different protein features in predicting EPs

A total of 2,726 features covering 17 protein descriptors were evaluated using seven training datasets to explore their predictive utility in EP classification (Fig. 2A). ROC analysis revealed that 1,287 features had an AUROC score greater than 0.70, with 67 features exceeding an AUROC score of 0.80. In addition, nine features, belonging to CTDC, CTDT, GAAC, QSOrder, and SOCNumber, achieved an average AUROC score greater than 0.80 across seven training datasets (Fig. 2B). These features also demonstrated strong performance on the validation and test datasets (Fig. 2C). These results highlight the potential of these features in EP identification and underscore the value of sequence information for EP prediction.

Fig. 2.

Performance of the 2726 protein features in the classification of EPs. A Statistics of the features in 17 protein descriptors included in this study. B Top ten protein features with the best predictive performance for EPs in seven training datasets. C Validation of the top ten features for prediction of EPs on the validation (left) and test (right) datasets

Creating a classifier pipeline with diverse feature engineering and ML techniques

Given that certain features may contribute minimally to the characteristics of EPs, their inclusion in the model could negatively impact classification performance. Thus, feature selection is crucial for identifying the optimal feature set for EP prediction prior to classifier construction. To evaluate the classification power of various protein features, we developed a pipeline that systematically tested 693 potential models. These models combined nine feature selection methods and 11 classification algorithms across seven training datasets (see “Methods”). The platform began with data standardization using the Z-score method. Next, the F-test was applied to select features that differentiate between EP and non-EP sequences. Third, to prevent overfitting, we employed a suite of nine feature selection methods to limit the number of features. Finally, 11 different classifiers were trained using each selected feature set. For each combination, the process was repeated 10 times, with 10% of the sequences held out for evaluation and the remaining 90% used for hyperparameter tuning, feature selection, and classifier training. The average performance of each of the 693 classifiers was evaluated using the AUROC score across the 10 held-out folds.

Of the 693 trained classifiers, 357 (51.5%) achieved a classification AUROC score > 0.95, with a maximum AUROC score of 0.969 (Fig. 3A, Table S2). Moreover, a subset of 46 combinations of feature selection and classification algorithms (highlighted with a green band in Fig. 3A) consistently performed well across all seven training datasets, achieving a mean AUROC score of 0.960. The performance of the 357 high-performing classifiers was evaluated on the validation and test datasets. Of the classifiers with an AUROC score > 0.95 on the training dataset, 336 (94.1%) also reached an AUROC score > 0.95 on the validation dataset (Fig. 3B, Table S2). These 336 classifiers involved nine different feature selection methods and nine classification algorithms, and also exhibited an AUROC score > 0.928 for the test dataset.

Fig. 3.

Construction of the baseline classifiers for predicting EPs using the training dataset. A AUROC scores were computed for each of the 693 baseline classifier results from the training datasets. The AUROC scores were based on the average performance of each classifier across 10 iterations, with 10% of samples being held out each time. B Performance of the top ten classifiers for prediction of EPs on the validation (left) and test (right) datasets

Development of an ensemble classifier with the stacking strategy

To build a composite ensemble classifier, a stacking strategy was employed to combine the classifiers that performed well (AUROC score > 0.95) on both the training and validation datasets. Prior to this, baseline classifiers were selected using the mRMR algorithm, based on the principle of minimum redundancy and maximum relevance. The number of baseline classifiers was optimized, with the ensemble classifier performing best when 15 baseline classifiers were used (Figure S2A). These 15 baseline classifiers were trained on six different training datasets, incorporating seven feature selection methods and five classification algorithms. These classifiers include two AB algorithm-based (D2-RFERF-AB and D4-RFEET-AB), four GB algorithm-based (D3-LASSO-GB, D4-SFMET-GB, D5-SFMXGB-GB and D7-SFMRF-GB), two LGB algorithm-based (D4-SFMXGB-LGB and D3-RFERF-LGB), six MLP algorithm-based (D7-SFMET-MLP, D6-RFEET-MLP, D5-RFELR-MLP, D3-SFMXGB-MLP, D2-SFMXGB-MLP and D2-SFMRF-MLP), and one SVM algorithm-based (D7-RFERF-SVM) classifiers. We used the t-SNE algorithm to further evaluate the predictive power of combinations of these classifiers. As shown in Fig. 4A, EP and control sequences were generally clustered separately in both the validation and test datasets. Unsupervised cluster analysis revealed distinct classification patterns for EP and control sequences in both datasets (Fig. 4B). These findings suggest that integrating these classifiers provides valuable insights for EP classification. To obtain the optimal ensemble classifier, the classification algorithm was further optimized, with the MLP algorithm demonstrating the greatest advantage (Figure S2B). Consequently, a stacked ensemble classifier, named EPDiscovery, was constructed based on 15 baseline classifiers with the MLP algorithm.

Fig. 4.

Predictive potential of the combined baseline classifiers in the classification of EPs. A t-SNE visualization of the protein sequences in validation (left) and test (right) datasets according to the prediction probabilities produced by 15 baseline classifiers. The red and blue dots represent EPs and controls, respectively. B The clustering heatmap of sequences in validation (left) and test (right) datasets based on the 15 classifiers prediction probability similarity (distance = “Euclidean” and method = “ward.d2”)

Validation and evaluation of EPDiscovery

The predictive performance of EPDiscovery was validated using the test dataset. ROC curve analysis revealed that EPDiscovery achieved an AUROC score of 0.970 (Fig. 5A). The performance of EPDiscovery achieved values of 0.926, 0.927, 0.852, 0.944, and 0.907 for ACC, F-value, MCC, SN, and SP, respectively (Table 1). Additionally, EPDiscovery was evaluated on three external datasets. The AUROC scores of EPDiscovery for external datasets 1, 2, and 3 were 0.968, 0.909, and 0.928, respectively (Table 1; Fig. 5B, C, and D). The overall accuracy across these three datasets was 0.939, 0.840, and 0.891, respectively. Validation results from multiple datasets confirmed the efficacy of EPDiscovery in distinguishing between EPs and controls.

Fig. 5.

Validation of EPDiscovery using the test dataset (A), and three external datasets from Taju et al. (B, C) and Pandey et al. (D). Shown are ROC curves and confusion matrices. On the confusion matrix plot, the green cells correspond to observations that are correctly classified. The red cells correspond to incorrectly classified observations. Both the number of observations and the percentage of the total number of observations are shown in each cell. The column on the far right of the plot shows the percentages of all the samples predicted to belong to each class that are correctly (precision) and incorrectly (false discovery rate) classified. The row at the bottom of the plot shows the percentages of all the samples belonging to each class that are correctly (recall) and incorrectly (false negative rate) classified. The cell at the bottom right of the plot shows the overall accuracy

Table 1.

Performance metrics for evaluating EPDiscovery on multiple test datasets

Dataset	AUROC	ACC	F-value	MCC	SN	SP
Test dataset	0.970	0.926	0.927	0.852	0.944	0.907
External dataset 1	0.968	0.939	0.736	0.719	0.903	0.943
External dataset 2	0.909	0.840	0.685	0.610	0.876	0.830
External dataset 3	0.928	0.891	0.873	0.780	0.840	0.933

To assess the effectiveness of the ensemble strategy employed in the development of EPDiscovery, we compared its performance with that of its constituent baseline classifiers. As shown in Fig. 6 A, EPDiscovery exhibited a higher AUROC score than the 15 constituent classifiers. Further, we compared the stacking strategy with two alternative ensemble learning strategies: majority voting and averaging. The majority voting strategy aggregates the most frequent prediction outcomes from all constituent classifiers, while the averaging strategy computes the geometric or ranking average of all constituent classifiers’ prediction probabilities. The stacking strategy yielded a higher AUROC score than the other ensemble learning strategies (Fig. 6B), suggesting that stacking effectively combines the strengths of the individual classifiers to achieve superior and more stable performance.

Fig. 6.

Performance comparison of EPDiscovery with other modeling strategies and existing methods using the test datasets. A ROC curves for EPDiscovery and its constituent baseline classifiers. B ROC curves for ensemble classifiers trained based on different ensemble strategies. C ROC curves for BacEffluxPred and EPDiscovery. D Classification metrics for evaluating the performance of EffluxBERT and EPDiscovery on the external dataset 1 (upper) and dataset 2 (lower)

To further evaluate the performance of EPDiscovery, it was compared with the existing methods, BacEffluxPred and EffluxBERT. On the test dataset, BacEffluxPred achieved an AUROC score of 0.805, while EPDiscovery surpassed this with a higher score (Fig. 6C). Furthermore, when comparing the performance on two external datasets, EPDiscovery also demonstrated comparable performance to EffluxBERT, especially with higher MCC and SN scores than EffluxBERT (Fig. 6D). This demonstrates that EPDiscovery has the potential to serve as an improved alternative for predicting EPs.

Explaining the predictions of EPDiscovery

To gain insight into the learning process, we applied the LIME method to analyze the feature importance of EPDiscovery. First, we examined the contributions of the 15 constituent baseline classifiers to EPDiscovery’s predictions using the test dataset. We observed that the top four classifiers had the most significant effects, including D3-RFERF-LGB, D5-RFELR-MLP, D2-SFMXGB-MLP, and D3-SFMXGB-MLP (Fig. 7A). Next, we identified the most important protein features for each of these constituent classifiers, with the top 20 features displayed in Fig. 7B and Figure S3. Further analysis revealed that the key features were primarily from protein descriptors such as CKSAAGP, DDE, and Moran (Fig. 7F). Notably, CKSAAGP and DDE maintained consistent importance across diverse EP families (Figure S4), emphasizing their critical role in EPDiscovery’s classification process.

Fig. 7.

Interpretation of EPDiscovery. A LIME analysis for 15 constituent baseline classifiers of EPDiscovery on the test dataset. Top 20 important features of D3-RFERF-LGB (B), D5-RFELR-MLP (C), D2-SFMXGB-MLP (D) and D3-SFMXGB-MLP (E). On the explanation plot, the coefficient represents the impact on the prediction output, while the positive and negative values represent the positive and negative contributions to the prediction output. (F) The distribution of protein descriptors for the top 20 important features in the 15 baseline classifiers. HC: hydrophobicity

Deployment of the EPDiscovery program

EPDiscovery was developed as a standalone command-line program to efficiently analyze large quantities of protein sequences derived from genomes or metagenomes. This program allows the user to input an arbitrary number of amino acid sequences, automatically extracts features, invokes EP classifiers to make predictions, and outputs their predicted probability and binary classification of EPs. The EPDiscovery program is currently available at https://github.com/ying-jc/EPDiscovery.

Discussion

EPs are critical in bacterial resistance to a broad range of antibiotics due to their ability to expel structurally diverse compounds. Beyond antibiotic resistance, EPs also play essential roles in bacterial stress responses, virulence, biofilm formation, and modulation of host interactions [34]. Accurate prediction of EPs can offer valuable insights into efflux pump function and guide subsequent experimental research. To this end, this study introduces EPDiscovery, a novel stacked ensemble classifier that leverages sequence information instead of structural information to accurately predict EPs. The method integrated multiple feature-encoding schemes capturing compositional, evolutionary, and physicochemical properties, alongside a series of popular ML algorithms for baseline classifier construction. Multiple strategies were employed to mitigate overfitting: (1) performing feature filtering and selection before classifier training; (2) conducting hyperparameter tuning strictly within a nested cross-validation framework to avoid information leakage; and (3) validating model performance on independent test and external datasets. To enhance its predictive performance, the feature selection method, the selection of the baseline classifier, and the ensemble algorithm were individually optimized. Experimental results showed that EPDiscovery outperformed its constituent classifiers and alternative ensemble strategies, confirming the effectiveness of the stacked ensemble approach adopted in this study.

In recent years, multiple tools for predicting EPs have been developed using ML techniques, including BacEffluxPred [10] and EffluxBERT [11]. Although employing different feature extraction methods, both tools utilized the SVM as their classification algorithm. This study also confirmed that the SVM algorithm performed effectively on the validation and test datasets, and was selected as one of the baseline classifiers to construct EPDiscovery. In contrast to these works, this study created a curated dataset containing a greater number of EP sequences, derived from the most recent database (Table S1). Demonstrating the effectiveness of EPDiscovery in EP prediction necessitates a comparison of its performance with these previously released tools. Nevertheless, due to differences in the test dataset, the results of this study are not directly comparable to prior works. A direct comparison of the performance of these tools would be more reliable if conducted on the same dataset. Currently, BacEffluxPred is the only tool available as a standalone program capable of large-scale EP identification. Comparative analysis revealed that EPDiscovery outperformed BacEffluxPred, underscoring its substantial role in EP prediction. In addition, validation using previously proposed datasets aids in evaluating the classifier’s performance. Across three external datasets from Taju2021 [11] and Pandey2020 [10], EPDiscovery demonstrated satisfactory performance, especially comparable to that of EffluxBERT, further affirming its robustness in EP identification.

Despite their widespread adoption, ML models largely remain black boxes. Understanding the predictions of ML models can enhance their trustworthiness and promote broader application [33]. This study employed the LIME method to interpret the predictions of EPDiscovery by examining the contributions of its constituent classifiers and protein features. Four baseline classifiers played a significant role in EPDiscovery’s predictions, three of which were MLP-based models, highlighting the advantages of the MLP algorithm in EP prediction. Notably, several protein descriptors garnered considerable attention, particularly CKSAAGP, DDE, and Moran, which were utilized by nearly all constituent classifiers. This underscores EPDiscovery’s capacity to discern the characteristics of EPs from protein sequence information alone, rather than merely memorizing data labels.

This study has several limitations that warrant acknowledgment. Firstly, a limited number of EP sequences were included in this study, despite efforts to retrieve as many sequences as possible from the latest protein databases. Incorporating more sequences for model training is widely recognized as beneficial for improving model performance. Nonetheless, the proposed classifier demonstrated a robust ability to predict EPs, surpassing existing tools and achieving the objectives of this study. Secondly, this study did not deploy EPDiscovery as an online web server but instead released it as a standalone program. This decision stems from the challenges associated with analyzing large quantities of protein sequences and integrating them into other genome analysis platforms. Web servers may also face accessibility issues due to network failures or maintenance. Finally, due to the goal of this study, EPDiscovery is currently unable to categorize EPs, which remains an important area for future work. Future releases of EPDiscovery will include functionality to predict EP categories, enabling a more comprehensive analysis of EP predictions.

Conclusion

This study presents EPDiscovery, a stacking-based ensemble classifier that integrates a series of ML-based classifiers to accurately and efficiently predict prokaryotic EPs based solely on sequence information. The development of EPDiscovery is expected to not only expedite the discovery of novel EPs but also enhance biologists’ understanding of EP functions. EPDiscovery is poised to become a reliable resource for EP prediction, particularly in large-scale bacterial genome studies.

Authors’ contributions

Q. W and Q. Y. performed all the experiments and prepared the initial draft of the manuscript. Z. T. assisted the data curation and visualization. T. X. and J. Y. conceived and designed the research, interpreted the results, and revised the manuscript. All authors read and approved the final manuscript.

Funding

This study was supported by grants from the Wenzhou Municipal Science and Technology Bureau (Y20210099), Science and Technology Project of Inner Mongolia Autonomous Region, China (2025YFSH0144), and Fundamental Research Funds for Wenzhou Medical University, China (Grant Number: KYYW201919).

Data availability

Project name: EPDiscovery; Project home page: https://github.com/ying-jc/EPDiscovery; Operating system(s): Linux Programming language: Python; Other requirements: Python 3.9.18 or higher; License: Apache Licence 2.0; Any restrictions to use by non-academics: licence needed.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.