XGBoost-enhanced ensemble model using discriminative hybrid features for the prediction of sumoylation sites

Abstract

Posttranslational modifications (PTMs) are essential for regulating protein localization and stability, significantly affecting gene expression, biological functions, and genome replication. Among these, sumoylation a PTM that attaches a chemical group to protein sequences—plays a critical role in protein function. Identifying sumoylation sites is particularly important due to their links to Parkinson’s and Alzheimer’s. This study introduces XGBoost-Sumo, a robust model to predict sumoylation sites by integrating protein structure and sequence data. The model utilizes a transformer-based attention mechanism to encode peptides and extract evolutionary features through the PsePSSM-DWT approach. By fusing word embeddings with evolutionary descriptors, it applies the SHapley Additive exPlanations (SHAP) algorithm for optimal feature selection and uses eXtreme Gradient Boosting (XGBoost) for classification. XGBoost-Sumo achieved an impressive accuracy of 99.68% on benchmark datasets using 10-fold cross-validation and 96.08% on independent samples. This marks a significant improvement, outperforming existing models by 10.31% on training data and 2.74% on independent tests. The model’s reliability and high performance make it a valuable resource for researchers, with strong potential for applications in pharmaceutical development.

Introduction

Proteins are vital in all living cells of both prokaryotic and eukaryotic organisms because they perform numerous functions critical to Life, such as acting as enzymes, structural components, regulators of gene expression, and modulators of immune responses [1]. The PTM (Posttranslational Modification) process is essential for protein regulation. Sumoylation, another PTM process, includes SUMO proteins binding to target proteins’ lysine residues [2]. This modification regulates cellular activities, including nuclear-cytoplasmic transport, transcription, DNA repair, and protein interactions. Additionally, it is closely associated with the pathological mechanisms behind neurodegenerative diseases like Parkinson’s and Alzheimer’s, which are linked to issues such as protein misfolding and disrupted cellular homeostasis [3]. Focusing on sumoylation sites provides valuable insights into modern genetics. Researchers have identified around 450 protein modifications, including sumoylation [4], phosphorylation [5], acetylation [6], and ubiquitination [7]. These modifications affect various aspects of protein behavior, such as interactions, cellular location, enzymatic activity, and PTM processes. Given its ability to regulate a wide range of posttranslational events, sumoylation has become a focus of intense study, particularly its role in protein degradation, recycling, intracellular localization, and overall cellular function [8].

Recent studies [9–11] highlight the strong connection between sumoylation and diseases such as Alzheimer’s and Parkinson’s [12], which has led to the development of bioinformatics tools to map these sites. Computational biology has significantly contributed to drug discovery and modeling [13], with sumoylation recognized as a promising treatment target, particularly for cancer and related diseases. Several computational models have been created to predict and characterize sumoylation sites. Tools SUMOsp [14], GPS-SUMO [15], and SUMOsp2.0 [16] employ algorithms based on phosphorylation scoring to assess peptide similarities. Other approaches, such as SUMO_LDA [17], use linear discriminant analysis (LDA) but face challenges with complex data. More advanced models like SUMOPre [18], SUMOHydro [19], pSumo-CD [20], SumoPred-PLM [21], RSX_SUMO [22], and HseSUMO [23] incorporate machine learning methods, including decision trees and Support Vector Machines (SVM), to improve accuracy. Recently, Deep-Sumo [24], a deep learning model, was introduced by Khan et al. and achieved an impressive 96.47% accuracy, outperforming earlier methods. However, despite the success of these models, there is still room for refinement, particularly in addressing the limitations of traditional learning methods that depend on feature extraction and linear datasets.

In the present study, we have proposed the XGBoost-Sumo, a novel computational model that highly precisely identifies sumoylation sites using the classifier Extreme Gradient Boosting. The model uses Position-Specific Scoring Matrix with Discrete Wavelet Transform (PsePSSM-DWT) combined with Bidirectional Encoder Representations from Transformers (BERT) during feature extraction, while SHapley Additive exPlanations (SHAP) is used for feature selection to give better interpretability. Extensive validation using 10-fold cross-validation showed that XGBoost-Sumo performed very well, with accuracy as high as 99.68% on benchmark datasets and 96.08% on independent samples. This represents a significant improvement over existing models by 10.31% on the training data and 2.74% on the independent tests. To date, XGBoost-Sumo is known to be one of the best predictors in sumoylation site prediction. The following are this work’s contributions.

• We developed a robust computational model for accurately predicting sumoylation sites built on XGBoost through a standardized learning procedure.

• Employ PsePSSM-DWT and BERT methods to convert peptide sequences into feature vectors effectively.

• The SHAP algorithm implements an effective feature extraction strategy that removes unnecessary and noisy information, improving the quality of the retrieved features.

• A thorough set of performance assessment measures shows how reliable and successful the model predicts sumoylation locations.

Framework model

The suggested model’s design is presented in this part, with Fig. 1 illustrating. A detailed explanation of every element that makes up the model is given to provide a thorough comprehension.

Fig. 1.

Proposed computational model framework

Benchmark dataset

A suitable training dataset in machine learning and bioinformatics is essential to creating a working prediction model. The selection of a benchmark dataset plays a significant role in determining the model’s overall performance. For this paper, to further ensure credibility and reliability during the training and validation process, we specifically opted for a benchmark dataset from [25]. The following visualization showcases the selected dataset, reflecting our assurance of transparency in the model development process.

1

Here, denotes sumoylation site sequences and non-sumoylation site sequences. The Compendium of Protein Lysine Modification (CPLM) [25] provides detailed information on 12 types of lysine PTMs. Using CD-HIT, we filtered out sequences with over 40% similarity, resulting in a benchmark dataset with less than 40% similarity, comprising 780 sumoylation and 21,353 non-sumoylation site sequences. Addressing the dataset’s imbalance is a common challenge in classification tasks. Several methods, such as undersampling and oversampling, are extensively used to manage this issue [26].

To address the imbalance in our dataset, we employed the Near Miss algorithm [27] using Python, applying undersampling techniques to reduce the number of non-sumoylation sites for a balanced dataset. The Near Miss algorithm is a sampling technique used to address class imbalance by selecting instances of the majority class that are closest to cases of the minority class. Unlike oversampling, which can lead to overfitting by duplicating minority class instances, undersampling ensures a more balanced distribution without repetition. As a result, we achieved 780 positive and 780 negative samples, creating a balanced dataset for our subsequent analyses and model training. This balance ensures the reliability of our computational method.

Furthermore, to evaluate the model’s generalizability [28] and to reflect better real-world conditions, we perform the two independent validation, i.e., balance-independent and imbalance-independent validation. For balance-independent validation, one-fifth of the benchmark dataset was set aside. We used a balanced set of 312 samples (i.e., 156 positive and 156 negative). For imbalance-independent validation, we used an unbalanced dataset from [25] to reflect better real-world conditions, i.e., 500 samples with 200 positive and 300 negative examples, as shown in Table 1. Importantly, no repeated samples from the training data in the test set, ensuring an unbiased evaluation of the model’s performance.

Table 1.

Benchmark dataset samples of sumoylation site sequences

Dataset	Positive Sequences	Negative Sequences
Training	780	780
Independent validation (balance Dataset)	156	156
Independent validation (imbalance Dataset)	200	300

Feature encoding schemes

In this section, we implement feature encoding schemes, as most predictive machine learning models require numerical data, which complicates peptide sequence representation. Feature extraction aids in this process, but choosing the right features is essential for model accuracy. The encoding must capture the sequence’s structure and key characteristics.

Position specific scoring Matrix

The PSSM is the most geriatric and successful sequence alignment used to contrive the distant protein relatives by aligning multiple sequences. The first method, introduced by Gribskov et al. [29], captures the residue-level sequence-based similarities and structural characteristics using profile-based PSSM. In the last years, PSSM has been proven to be an efficient predictor for a series of protein types like hormone-binding proteins [30], DNA and RNA-binding proteins, antifreeze proteins [31], antiviral peptides [32, 33], antioxidant proteins [34, 35], interaction proteins or Snare proteins, as well as for various membrane proteins [36]. PSSM for a protein sequence is computed using PSI-BLAST, which encodes the structural and evolutionary information in root symbols for residues of biological sequences. These categorical transitions are represented by the amino acid substitution scores offered by the matrix, compounding residue transformation probabilities over different residues (A to V, A to W). It reveals residues under evolutionary constraint, as positive scores (i.e., those with a higher score) tend to occur frequently and be evolutionarily conserved substitutions, and negative scores correspond to rare or unlikely substitutions. The PSSM matrix denoted by H is mathematically expressed as in Eq. 2.

2

Where h_{i, j} denoted the residue frequency of a peptide sample’s i^th and j^th amino acid. L denotes the length of a certain peptide sequence. Twenty amino acids are altered, and their corresponding residues are shown in rows of matrix ‘H,’ representing the amino acid residues in a peptide sample. The PSSM data are normalized using the sigmoid function below. (X) 0→1

3

In this paper, we applied protein pseudo and DWT methods to the PSSM matrix to achieve high discriminative evolutionary descriptors.

Pseudo-PSSM

Position-specific scoring Matrix (PSSM) offers valuable evolutionary information, but machine learning algorithms such as SVM, RF, and KNN face challenges when handling protein sequences of varying lengths. Furthermore, PSSM doesn’t consider the order of sequences or correlations between residues [37]. To address these concerns, we present PsePSSM, a method for computing residue frequencies that considers sequence-order features. PsePSSM has applications in various fields, including DNA-binding protein systems, bioinformatics, proteomics, and predicting non-protein structures [38].

In this study, PsePSSM creates an L*20 vector from a protein structure with the length L by utilizing the PSI-BLAST tool in conjunction with the Swiss-Prot database [39, 40]. This approach guarantees that vector length is consistent across peptide samples by comparing the correlation between residues separated by “d” amino acids and determining the average score for each amino acid in the PSSM matrix. The following is one way to depict the PsePSSM vector for a peptide sample.

4

Where,,denotes the mean score of all amino acid residues. Which are mutated to q amino acid in peptide sample ‘S’

Whereis the sequence ordering details of the peptide sample, q represents the amino acid, and is the contiguous distance.

Discrete Wavelet transform (DWT)

Nanni et al. [41] proposed using the Discrete Wavelet Transform (DWT) to analyze biological samples’ frequency and residue data. DWT is highly effective for signal compression and noise reduction tasks, breaking down amino acid sequences into multiple levels to reveal hidden patterns. This method transforms the PSSM matrix of a peptide sequence into an image using specific parameters. The resulting PSSM image matrix is divided into various resolution levels to extract essential features that may not be captured through sequential encoding. Each resolution level is separated into detailed coefficients (high-frequency data) and approximation coefficients (low-frequency data) [42, 43]. Research suggests that the low-frequency components typically carry more valuable information than high-frequency ones.

Since extracting features from these low-frequency components can be challenging, we decompose the level-1 approximation into further components, HF1 and LF1, to uncover additional hidden features [44]. This subsequent decomposition enables the extraction of highly discriminative characteristics. Ultimately, the characteristics from high and low frequencies at each level are combined into a transformed feature set labeled HF1 and LF1.

6

In this case, wavelet function, i.e., w(j-q/p), input signal, i.e., y(j), and p and q represent the scaling and translating variables. S(p, q) indicates the transformation coefficients. Since more decomposition creates redundant and noisy characteristics, we limit our decomposition to two layers. Thus, a unique encoding technique, PsePSSM-DWT, is devised using 2-level DWT transformations into the PsePSSM matrix.

Bidirectional encoder representations from transformers (BERT)

Recently, transformer-based models like BERT have made significant progress in NLP by capturing contextual and semantic relationships [45]. This study uses the BERT architecture to extract features from peptide sequences, treating amino acids as words. Specifically, we use ProtBERT-BFD [46], which combines protein-based BERT embeddings with the Big Fantastic Database (BFD) for enhanced feature representation. Each amino acid is tokenized from the peptide sequences, with a unique “CLS” token added to represent aggregated features for prediction [47]. Each sequence is padded to a fixed length of 200 using PAD, and the SEP token is utilized to separate sequences. The tokenized sequences are transformed into 1024-dimensional feature vectors using global average pooling. The prediction process in deep learning models begins with feeding this extracted information into their input layer. Figure 2 shows the ProtBERT-BFD model’s general mechanism.

Fig. 2.

ProtBERT-BFD model of word embedding

Hybrid features

To extract semantic associations and contextual information from the peptide sequences, we used ProtBERT-BFD. Conversely, the PsePSSM-DWT is utilized to gather changed features of evolution. Additionally, to improve the training features’ discriminative ability with high predictive outcomes, we fused the extracted vectors—220D of PsePSSM-DWT and 1024D of BERT—to create a hybrid vector, 1244D, by making the necessary adjustments to offset the separate vectors’ weaknesses.

7

SHAP feature selection

Understanding the biological significance of selected descriptors can be challenging, as machine learning models are often seen as “black boxes” due to their complex Nature. In machine learning, discussions about data shape revolve around the size of data groups and structure, which is critical for data owners when determining how to process and manage datasets. The performance and effectiveness of machine learning algorithms can vary depending on the dataset used for training. Therefore, understanding data shape is crucial for vital preprocessing tasks like splitting, feature selection, and normalization [48]. Data science relies on well-organized data to aid decision-making by providing essential insights. To give credit where credit is due in machine learning models, SHAP [49] employs cooperative game theory. It assigns a numerical value to each feature, reflecting its impact on improving prediction accuracy. SHAP measures the difference in contributions from selected features inside and outside the model, highlighting each feature’s influence on the overall outcome [50]. Equation 8 provides a formal mathematical representation of this process, capturing the incremental effect of adding feature I to different subsets of features.

Where, , represents the SHAP value for the feature i, N, is the set of all features, S, is a subset of features excluding I, is the model’s prediction given features in S, and is the model’s prediction given features in S plus feature

Figure 3 presents the top 10 features; each row represents a specific feature, with dots showing the feature values in a given instance. Red dots indicate positive feature values, while blue dots represent negative values. The horizontal axis displays the SHAP values, indicating how each feature influences the model’s predictions [51]. Positive SHAP values mean the feature increases the likelihood of Sumo (positive outcome), while negative values increase the likelihood of non-Sumo (negative result). We experimented with feature groups of various sizes and found that using the highest 163 features (i.e., 50D of PsePSSM-DWT and 113D of BERT) significantly enhanced the model’s performance.

Fig. 3.

SHAP-selected features impact the model output

Figure 4 shows the results of a LIME study [52] conducted on a randomly chosen instance after the SHAP analysis for group classification, which allowed us to delve further into the function of individual features. Class 1 denotes a positive class, while class 0 denotes a negative class.

Fig. 4.

Using LIME on instances chosen at random following SHAP feature selection

Samples visualization via tSNE

To evaluate the effectiveness of the extracted features, t-distributed Stochastic Neighbor Embedding (t-SNE) [53] was used to visualize the hybrid feature vector before and after feature selection, as shown in Fig. 5. The t-SNE maps revealed distinct clusters, helping to differentiate between Pos⁺ and Neg⁻ sequences. Although false negatives and positives appeared between the true positive and accurate negative clusters, their occurrences were rare. The t-SNE method effectively represents both local structural details and global relationships.

Fig. 5.

t-SNE visualization of (A) without feature selection Hybrid Training features (B) SHAP-based Selected features

Further visualization of the extracted features using t-SNE reduced the high-dimensional vectors into a 2D space (see Fig. 5). In Fig. 5(A), the hybrid features show some overlap between positive and negative samples, leading to partial separation but insufficient for precise classification. In contrast, Fig. 5(B) clearly distinguishes between the two classes, highlighting the superior performance of SHAP-selected optimal features in differentiating between Sumo and non-Sumo, compared to the hybrid features in Fig. 5(A).

eXtreme Gradient Boosting (XGBoost)

XGBoost is a gradient-boosting library developed to provide a more flexible, scalable, and faster machine learning process [54]. It utilizes an ensemble learning algorithm called gradient boosting, which creates a predictive model by combining the outputs of multiple weak learners, typically decision trees. XGBoost also includes a specialized matrix class that enhances data storage and access efficiency during model training and evaluation, ensuring high performance [55, 56]. XGBoost minimizes the mean squared error among actual and predicted values in regression tasks. The library offers several loss functions to steer the optimization process, including Huber loss, absolute error, and squared error. The objective function evaluates the overall model performance, while the loss function compares predicted values to actual ones, as demonstrated in Fig. 6.

Fig. 6.

The Workflow of the XGBoost algorithm

XGBoost does the training in stages, adding new trees one after another to correct mistakes made by the previous trees; this would be guided by the negative gradient of the loss function [57]. Cross-validation is one of the essential techniques for assessing performance in XGBoost, whereby dividing the data into subsets provides multiple training and validation rounds. Users can choose between binary and multiclass classification-oriented objectives, such as logistic regression or Softmax, when working with an XGBoost classifier because this library is designed to play nicely with scikit-learn and can confidently switch back and forth between XGBoost’s native API and the scikit-learn API without compatibility issues, allowing for easy use of other machine learning libraries—the optimized parameters of the proposed model as shown in Table 2.

Table 2.

XGBoost classifier optimized parameters

Classifier	Parameter	Value
XGBoost	n_estimators	200
	Random state	42
	lambda (reg_alpha)	0.1, 1
	Objective Function	Binary logistics
	alpha (reg_lambda)	0.1, 1
	Booster	Gbtree
	gamma	0.5
	Col sample by level	0.5, 0.8, 1.0
	Min child weight	10
	Max depth	20
	Learning rate	0.01

Performance evaluation

The performance evaluation matrix provides a handy tool to find the weaknesses and strengths of the different ML models [58]. It has better-defined metrics to validate the model’s performance in many aspects. “Accuracy” assesses the overall validity of the predicted; “Precision” measures how many of its optimistic predictions are genuinely positive. Recall or Sensitivity is the model’s capability to accurately classify all relevant instances, while Specificity measures how well the model avoids misclassifying negatives [59]. The MCC is very efficient in the case of imbalanced datasets because it considers TP, FN, TN, and FP for its calculation [60], hence providing an overall measure. Other necessary measures are classifier discrimination-ROC Curve and AUC, which estimate the model’s ability to discriminate between classes. The confusion matrix illustrates how many items have been correctly and wrongly predicted concerning each class [61]. These measures enable full model performance review, which may be very important when trying to meet the growing requirements of machine learning technology [62].

9

10

11

12

Experimental results and analysis

Performance analysis of cross-validation Scheme

We evaluated the suggested model’s performance using various feature extraction methods, including PsePSSM-DWT and BERT, and different hybrid features using multiple cross-validation tests. The results of the XGBoost model, tested using 5-fold cross-validation across different sequence formulation techniques, are presented in Table 3. XGBoost achieved superlative performance with hybrid features compared to PsePSSM-DWT and BERT features. For example, before applying feature selection, the XGBoost model’s median success rate was 98.91%. To further enhance its performance, we implemented a feature selection method to reduce the dimensionality of the hybrid feature space. This led to a notable improvement, with the accuracy increasing to 99.08%, as highlighted in Table 3.

Table 3.

5-fold cross-validation performance evaluation using XGBoost mode

Method	ACC (%)	SP (%)	SN (%)	F1 Score (%)	MCC
PsePSSM-DWT	98.72	98.68	98.75	98.74	0.974
BERT	98.78	98.85	98.71	98.70	0.976
Hybrid feature (Before feature selection)	98.91	98.89	98.92	98.93	0.978
Hybrid feature (After feature selection)	99.08	99.10	99.05	99.04	0.982

Moreover, Table 4 shows the accuracy improvements exerted on the individual and combined features for the XGBoost classifier based on a 10-fold cross-validation test. As illustrated in Table 4, the XGBoost classifier performance was higher when using hybrid features than other individual feature methods. For instance, the model achieved the highest average accuracy, 99.02%, before applying feature selection. When feature selection was used, its accuracy further increased to 99.68%. This means the XGBoost classifier has optimum predictive effectiveness on hybrid features in a 10-fold cross-validation setting.

Table 4.

10-fold cross-validation performance evaluation using XGBoost Mode

Method	ACC (%)	SP (%)	SN (%)	F1 Score (%)	MCC
PsePSSM-DWT	98.76	98.78	98.73	98.79	0.975
BERT	98.82	98.78	98.85	98.82	0.976
Hybrid features (Before feature selection)	99.02	99.02	99.01	99.04	0.980
Hybrid features (After feature selection)	99.68	99.62	99.74	99.74	0.994

The performance of XGBoost-Sumo was further examined using the AUC metric, a key indicator of binary classifier accuracy. A higher AUC score directly translates to better model performance. As illustrated in Fig. 7, XGBoost-Sumo delivered outstanding results, achieving an AUC of 0.992 before the feature section and 0.998 after feature selection using 10-fold cross-validation. These findings highlight the model’s exceptional predictive capabilities, particularly when using selected features on the 10-fold cross-validation method. Additionally, Fig. 8 presents a confusion matrix that delves deeper into the performance of the XGBoost classifier, showcasing its effectiveness in the prediction with selected features vector on the 10-fold.

Fig. 7.

10-Fold cross validation performance comparison of AUC

Fig. 8.

Confusion matrix of the proposed model

Others learning classifiers comparison

This section compares the proposed model’s performance against other commonly used ML methods using hybrid features. K-Nearest Neighbor (KNN), Linear Regression (LR), Random Forest (RF), SVM, and Naive Bayes (NB) are considered for classifier comparison. LR is a standard statistics method employed to classify things into two categories by calculating the probability of each outcome. A maximum likelihood estimation method determines the relationship between a dependent variable with two possible outcomes and one or more independent factors. LR is valued for being simple, easy to understand, and effective in real-world applications like diagnosing medical conditions and assessing credit risk.

In image processing, KNN is a widely employed instance-based and non-parametric method. The KNN algorithm classifies data by calculating Euclidean distances between instances. RF is an ensemble learning method generally used in regression and classification dilemmas. It generates many decision trees created using the bootstrap technique on randomly selected training samples. Support Vector Machine is a robust classification method usually used in bioinformatics applications. It can solve linear and nonlinear problems by determining an optimal hyperplane to separate classes. Based on Bayes’s theorem and assuming feature independence, NB is a basic yet effective probabilistic classifier. It works well on large datasets, such as text classification tasks like fake recognition. Even with its uncomplicatedness, NB often provides high accuracy in various applications. Table 5 shows the performance of these classifiers.

Table 5.

Performance comparison with other Learning algorithms

Method	ACC (%)	SP (%)	SN (%)	F1 (%)	MCC
RF	98.20	98.04	98.36	98.36	0.964
LR	98.48	98.64	98.32	98.32	0.969
KNN	98.49	98.59	98.38	98.38	0.970
NB	98.57	98.67	98.46	98.46	0.971
SVM	98.61	98.59	98.63	98.63	0.972
XGBoost-Sumo	99.68	99.62	99.74	99.74	0.994

From Table 5, XGBoost achieved the maximum accuracy of 99.68% among all classifiers. SVM followed with an accuracy of 98.61%. For the MCC value, which represents model stability, the best value of 0.994 was obtained by XGBoost, while the SVM value was the next highest at 0.972. As illustrated in Fig. 9, XGBoost-Sumo delivered outstanding results, achieving an AUC of 0.998 using 10-fold cross-validation on an optimized feature set.

Fig. 9.

AUC performance comparison with commonly used classifiers

Performance comparison with existing models

This section compares the proposed model with the existing benchmark methods referenced in [20, 23], ²⁴and²⁸. These recently developed approaches have built prediction models using ML methods. We conducted performance evaluations using 10-fold cross-validation on established benchmark datasets to assess the effectiveness of our model and the benchmarks. For comparison, Table 6 summarizes the comparison of the results with the existing methods.

Table 6.

Existing models performance comparison

Method	ACC (%)	MCC	SN (%)	SP (%)
pSumo-CD 20	72.80	0.494	53.60	92.10
HseSUMO 23	89.50	0.790	89.50	89.50
Deep-Sumo 24	96.47	0.929	96.71	96.25
PSSM-Sumo 28	98.71	0.974	97.72	99.68
XGBoost-Sumo	99.68	0.994	99.74	99.62

As shown in Table 6, the proposed XGBoost-Sumo model outperforms the existing models by a considerable margin. For instance, the highest accuracy by XGBoost-Sumo was 99.68%, while the second best among compared predictors, PSSM-Sumo, reached an accuracy of 98.71%. Besides, XGBoost-Sumo obtained an MCC of 0.994, which is a high value, considerably outperforming Pssm-Sumo’s score of 0.974. Fairly noteworthy from this result is that the model XGBoost-Sumo has the best performance, improving the average success rate of the benchmark methods to 10.31%.

Performance comparison on an independent dataset

The accurate measure of a prediction model’s generalization lies in its performance on unseen data. To assess the robustness of our developed model, we tested it on an independent dataset, splitting the data into 80% for training and 20% for testing. The results from this evaluation are presented in Table 7. Among existing methods, PSSM-Sumo demonstrated solid performance, achieving an accuracy of 94.45% and an MCC of 0.912. However, the proposed model outperformed existing methods, delivering a higher accuracy of 96.08%, Specificity of 95.42%, Sensitivity of 96.77%, and an MCC of 0.922. XGBoost-Sumo has the best performance, improving the average success rate of the benchmark methods as high as 2.74%.

Table 7.

Independent dataset performance comparison with existing models

Method	ACC (%)	MCC	SP (%)	SN (%)
Deep-Sumo 24	92.23	0.892	93.53	90.93
PSSM-Sumo 28	94.45	0.912	92.03	96.87
XGBoost-Sumo	96.08	0.922	95.42	96.77

Furthermore, the XGBoost-Sumo model was evaluated on an independent imbalanced dataset to accurately reflect the natural imbalance in real-world data. Its performance was compared with other ML models, as detailed in Table 8. From Table 8, XGBoost-Sumo achieved the highest accuracy at 95.58% and demonstrated superior Specificity, Sensitivity, and MCC compared to other methods. Among conventional classifiers, SVM achieved the highest accuracy at 93.54%, outperforming both RF and KNN. Hence, the enhanced performance results using both the independent datasets (balanced and imbalanced) ensure the effectiveness and validity of our model.

Table 8.

Performance comparison on independent imbalance dataset

Method	ACC (%)	MCC	SP (%)	SN (%)
XGBoost-Sumo	95.58	0.913	93.42	97.76
SVM	93.54	0.872	91.67	95.41
KNN	92.04	0.843	88.56	95.61
RF	91.80	0.837	89.34	94.29

Conclusion

When accurately identifying sumoylated sites, the assessed XGBoost-Sumo demonstrated excellent dependability. XGBoost-Sumo, which uses an optimized learning algorithm as a sophisticated feature extraction method, beats previous models in this field. Thorough 10-fold cross-validation was used to validate the model’s correctness, and the results showed good generalization performance in practical situations. Furthermore, tests of XGBoost-Sumo versus popular machine learning methods demonstrated unique advantages that resulted in more precise sumoylation site predictions.

For future directions, we will examine the integration of transfer learning to enhance adaptation across various datasets, and refining model architecture via hyperparameter optimization and ensemble techniques may augment performance and resilience. Furthermore, we proposed parallel programming methodologies to improve the model’s efficiency, speed, and scalability in future endeavors.

Author contributions

All author contributed equally.

Funding

This research is not funded.

Data availability

The datasets used and/or analyzed during the current study are available on Github link https://github.com/salman-khan-mrd/Sumo.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.