The benchmark dataset typically comprises positive samples (proteins capable of interacting with drugs) and negative samples (proteins incapable of interacting with drugs). We used Jamali et al.’s dataset [32]. To ensure a reliable comparison with current approaches. The 1611 druggable proteins have been obtained by using the Drug-Bank database, as previously reported. Similar sequences of these proteins, with respect to characteristics and composition, were eliminated using the Cluster Database at High Identity with Tolerance (CD-HIT) tool. Additionally, there are 1224 druggable proteins in the final set of positive samples. Likewise, the negative sample set was created by integrating the datasets presented by Bakheet and Doig [33]. First, the Swiss-Prot repository was used to obtain these sequences. Following the elimination of analogous sequences, 1300 non-druggable proteins persisted. The final benchmark dataset, which consists of 1218 druggable proteins, is employed in experimental data points amounting to 2518 druggable proteins, as shown in Table 1.

The normalization method is implemented on the data of each DDE vector feature profile, which comprises 400 elements. For the 400 vector characteristics extracted from the original protein profiles, the DDE model demonstrates the generation of score matrix functions. The initial step involves synchronizing all values with a common amino value. The frequency of amino acids is subsequently divided by the length of the sequence. In conclusion, formula X is used to scale all functional values. Means and standard deviations are derived using vector score feature profiles, which consist of 400 vector features. The DDE and 2D techniques are used to configure all vector profiles. The dipeptide composition, as described by Bhasin and Raghava (2004) [34], has been employed to forecast various protein sequence functions, as demonstrated by Dhanda et al. (2013) [35].

Hence, the Dipeptide Deviation from Expected Mean (DDE) method, developed by Saravanan and Gautham (2015) [36], Quantifies deviations in dipeptide frequencies within protein sequences to different enzymes from non-enzymes. The DDE function vector is derived using three parameters: (1) Observed dipeptide composition (D_c), (2) theoretical mean (TM), and (3) theoretical variance (Tv). (1)De(i)=niN

Here, De(i) represents the observed frequency of a dipeptide i in the sequence, where ni is the number of occurrences of the dipeptide i, and N=L−1 is the total number of dipeptides in a sequence of length L. (2)TM(i)=Ci1CN×Ci2CN

The theoretical mean (TM(i)) estimates the expected frequency of a dipeptide i, assuming independence between amino acids. Ci1 and Ci2 denote the counts of the first and second amino acid in dipeptide i across the entire dataset, while CN=61 represents the total codons (excluding three stop codons). (3)Tv(i)=TM(i)(1−TM(i))N

The theoretical variance (Tv(i)) quantifies the expected variability of a dipeptide’s i frequency under the independence assumption, where N is the total dipeptides in the sequence. (4)DDE(i)=Dc(i)−TM(i)TV(i)

The DDE score standardizes the deviation of observed dipeptide frequencies from their theoretical expectations, with higher absolute values indicating stronger deviations. A 400-dimensional feature vector is generated by computing DDE_(i) for all 400 possible dipeptides. (5)DDEp={DDE(i),...,…DDE(n)},where,i=1,2,…,400

The approach proposed according to Chen et al. [37] involves the extraction of sequential protein information, which is subsequently used to derive amino-acid frequency information. (6)g(m,n)=H(m,n)H(n),m∈{A,C,D,…,Y},n∈{W1,W2,....WL}

This method calculated the frequency of amino acid m within sliding windows of the sequence. Denotes the count of amino acid m in the n-th window, and H(n) is the window length. For example, a window size of 5 residues is slid across the sequence to capture local compositional biases.

The computation of EGAAC is derived from the categorization of amino acids. This study classifies amino acids into five distinct categories based on their physicochemical characteristics [38]. This method uses protein sequences to create numerical feature vectors, which are derived from their respective properties. Amino acid are grouped into five categories based on psychochemical properties: aliphatic (such as G, A, V, L, M, I), aromatic (F, Y, W), positively charged (K, R, H), negatively charged (D, E), and neutral (S, T, C, P, N, Q). While here (H(g,n)) is the count of amino acids in group g within the n-th window, and H(n) is the window length, for example, 5 residues. This feature extraction approach is highly effective in bioinformatics research fields, particularly in the prediction of druggable proteins. The calculation of EGAAC is determined by the following equation: (7)EGAAC(g,n)=H(g,n)H(n),g∈{g1,g2,g3,g4,g5},n∈{W1,W2,....WL}

Consequently, it is crucial to translate knowledge about protein sequences into numerical values that ensemble learning algorithms can comprehend and use [39–41]. This study utilized three methods for extracting characteristics from protein sequences. The three methods are.

The DDE, EGAAC, and EAAC. An approach to machine learning known as ensemble learning involves training many models. Commonly referred to as “weak learners,” they address the same problems and then combine them to achieve improved outcomes. By combining weak models, we can achieve a more precise model. The concept of Ensemble Learning encompasses three distinct models, including Bagging, Boosting, and Stacking. The fundamental framework comprises multiple machine learning models. The methodology known as CEL-IDPs Fig. 1 comprises five fundamental steps: (i) The researchers gathered datasets and benchmarking data, followed by preprocessing and eliminating redundant similarity, duplication deletion. (ii) Features were extracted to cover various characteristics of sequencing data. (iii) A feature representation learning methodology was utilized, and analysis was conducted using the t-distributed Stochastic Neighbor Embedding (t-SNE) algorithm. (iv) A predictor was constructed using a three-step distinct model: Bagging, Boosting, and Stacking. (v) The final classification model was constructed, as depicted in Fig. 1. The following sections have presented a comprehensive elucidation of each of these noteworthy phases.

Fig. 1 illustrates the proposed methodology’s comprehensive flow chart that depicts the organized process of our proposed methodology for identifying druggable proteins. In Section A, we compiled and improved a dataset of both druggable and non-druggable proteins, assuring data quality through redundancy. Section B emphasizes feature analysis, wherein we do SHAP analysis to elucidate feature significance, followed by t-SNE for data visualization. Furthermore, we obtained essential protein sequence descriptors, namely DDE, EAAC, and CGAAC. Section C delineates the predictive modeling phase in which various ensemble learning techniques-bagging, boosting, and stacking were employed to improve model generalizability and classification efficacy. The initial phase was the construction of a database containing both druggable and non-druggable proteins. Subsequently, the calculation of three families of protein composition descriptors was performed using Python. These descriptors included the 20 amino acid composition (AC), the Dipeptide deviation from expected mean (DDE), features length was 400, the Enhanced amino acid composition (EAAC) method, features length was 920, and the enhanced grouped amino acid composition (EGAAC) method, features length was 230. In the subsequent phase, Jupyter notebooks were utilized to implement Python/Scikit-learn-based machine learning classifiers [42]. These classifiers were constructed by integrating thirteen distinct types, which were derived from the combination of three descriptor families (DDE, EAAC, EGAAC). The classifiers utilized in this study encompassed a range of machine learning algorithms. First step, we conducted all analyses using the Bagging method first (Bagging Classifier) [43], second, the random forest (Random Forest Classifier) [44], and third, the decision tree (Decision Tree Classifier) [45]. In the second step, we used quantitative techniques to analyze the Boosting method, specifically, first AdaBoost [45], GB Boost [46], XGBM [47], LGBM, and CatBoost classifiers are used for the prediction of DPs sites. In the third step, we used the stacking method to combine quantitative techniques for analysis, which was based on first LR, SVC, NB [48], KNN [49], and DT [50] classifiers for the prediction of DPs sites. XGBoost (XGB) is an alternative ensemble method that utilizes sequential weak trees to rectify classification errors [47]. Using gradient boosting to classify is a well-established bootstrapping technique that relies on utilizing a succession of consecutive weak classifiers. A meta-estimator called the Ada-Boost classifier (AdaBoost) [45] starts the fitting process using a classifier that was created from the original dataset. Subsequently, it incorporates many iterations pertaining to the initial classifier, each with modified weights to account for cases that were improperly categorized. The Bagging classifier, also known as Bagging, shares similarities with AdaBoost. In Bagging, supplementary classifiers are produced from sub-sets of the initial dataset. The prediction model for machine learning was developed with distinct protein datasets. The positive set comprised 1219 druggable proteins, identified through their inclusion in the DrugBank database (www.drugbank.ca) and their correlation with FDA-approved pharmaceuticals. Conversely, the negative protein collection consisted of 1299 proteins that are not amenable to pharmacological targeting, as reported in a previous study.

The final machine learning prediction model was used to analyze three sets of cancer-associated protein lists. A total of 230 proteins were identified as essential in the BC dataset [51]. Additionally, 2353 proteins known to drive cancer were obtained from the Network of Cancer Genes [52], while 1365 proteins were classified as RNA-binding proteins (RBPs).

In Table 2, the results of our study indicate a requirement for improved outcomes in order to demonstrate that the proposed methodologies of bagging, boosting, and stacking improve the performance of these three feature extraction methods. DDE, EAAC, and EGAAC approaches. Results were considered significant. DDE features were used with bagging (RF obtained 66.87%), boosting (XGBM obtained 66.66%), and stacking (Stacking obtained a score of 68.33%). After that, EAAC features used bagging (Bagging obtained 66.66%), boosting (CatBoost obtained 56.66%), and stacking (Stacking obtained a score of 66.66%). Third EGAAC features were considered significant and outstanding, with bagging (RF obtained 71.66%), boosting (XGBM obtained 71.66%), and stacking (Stacking obtained a score of 68.33%). The highest level of performance is achieved by employing the EGAAC feature extraction method with bagging, boosting, and stacking methods.

Results were considered significant. DDE features were used with bagging (RF obtained 66.87%), boosting (XGBM obtained 66.66%), and stacking (Stacking obtained a score of 68.66%), as shown in Table 3. The initial data set, referred to as the independent dataset, was created using the DDE model and the bagging, boosting, and stacking strategies. Table 3 illustrates this. The stacking method exhibits the greatest accuracy among ensemble learning techniques, attaining an accuracy score of 68.33%, as illustrated in Fig. 5. Additionally, the stacking process produces precise outcomes.

After that, EAAC features used bagging (Bagging obtained 66.66%), boosting (CatBoost obtained 56.66%), and stacking (Stacking obtained a score of 66.66%), as shown in Table 4. This is feasible due to the significantly larger amount of data in the dataset compared with dataset 1. It makes it possible for the model to enhance its learning efficiency, as illustrated in Fig. 6. Furthermore, it is possible that the data employed has a distinct pattern. According to the experimental results on the dataset, the boosting methods AdaBoost, XGB, and Light Gradient Boosting, as well as the random forest Bagging model, attained an average accuracy of 56.66%. Boosting has greater accuracy than bagging and stacking, as evidenced by the prior dataset, with the stacking method achieving the highest accuracy at 66.66%.

Third EGAAC features were considered significant and outstanding results with bagging (RF obtained 71.66%), boosting (XGBM obtained 71.66%), and stacking (Stacking obtained a score of 68.33%), as shown in Table 5. The highest level of performance is achieved by employing the EGAAC feature extraction method using bagging, boosting, and stacking techniques, as illustrated in Fig. 7. Table 5 displays the results of ensemble learning experiments performed on a separate dataset utilizing the EGAAC model. The accuracy acquired from the experimental findings of three ensemble learning approaches, specifically Bagging, Boosting, and Stacking, varied across three distinct datasets.

Prioritizing previously identified genes based on the Pharmacogenomics Knowledgebase (PharmGKB), the interpreter for cancer genomes, and the Consensus Strategy of the Pan-Cancer Atlas project, we conducted comprehensive analyses of genetic changes, signaling pathways, protein interactions, networks, protein expression, dependency maps, and enrichment maps in a prior study. These investigations enabled us to identify critical proteins linked to the etiology of breast cancer. Predicting the draggability of breast cancer (BC) proteins using machine learning (ML) techniques has the potential to provide important information about potential biomarkers, potential treatment targets, and upcoming clinical trials. This strategy could improve precision medicine and global cancer pharmacogenomics while reducing ethnic bias [56,57].

To enhance the interpretability of our predictive model, we utilized SHAP (SHapley Additive exPlanations) to analyze the impact of specific characteristics in distinguishing druggable and non-druggable proteins. Unlike traditional feature selection methods, through the quantification of each feature’s influence on the model’s decision-making mechanism, SHAP offers a simple way to gauge the value of features [58].

Fig. 11 illustrates the SHAP-based feature analysis applied to three different feature encoding methods: Dipeptide Deviation Encoding (DDE), Enhanced Amino Acid Composition (EAAC), and Extended Grouped Amino Acid Composition (EGAAC). The SHAP plots reveal the most influential features contributing to the classification process. Features with high SHAP values strongly influence the model’s decision in favor of druggable proteins (positive impact) or non-druggable proteins (negative impact). These findings offer deeper insights into the biochemical properties that play a crucial role in druggability prediction, thereby strengthening the biological relevance of our model.

The figure demonstrates SHAP (SHapley Additive exPlanations)-based feature analysis applied to different datasets (DDE, EAAC, and EGAAC) using three different machine learning techniques: Bagging, Boosting, and Stacking. While SHAP values explain the importance of features for the model’s prediction, the figure does not make the relevant biology behind these traits as they relate to drug targets for identification, as you observed. Guided visualization and interpretation of SHAP-Inspired results: To improve the interpretability and utility of the SHAP-based research image shown in this study, it is important to annotate these top-ranked features (by the SHAP because these features are considered top-ranked features) with known biology and biological relationships. Conduct checks on the literature to ensure that these top-ranking features (e.g., “DDE115,” “EAAC616,” “EGAAC146,” etc.) actually correspond to biological functions or diseases that are relevant to the new drug target prediction. Explore whether these characteristics are linked to known biomarkers, genes, proteins, or pathways related to diseases such as cancer, diabetes, or neurodegenerative diseases. A functional relationship can be established if the features in the datasets are related to gene expression or protein activities; otherwise, use Gene Ontology (GO) analysis or pathway databases (e.g., KEGG, Reactome, or STRING). For example, one might question whether the genes linked to the affected traits are involved in specific biological processes like cell growth, death, or immunological response. Use bioinformatics techniques to determine if these features are similar to known functional domains, molecular functions, or biological processes. Often, highly maintainable features in biological datasets represent core regulatory proteins, enzymes, and receptors that are involved in drug-target interactions. Cross-reference resulting core traits with biological databases (e.g., UniProt, Ensembl, and/or GeneCards) to identify their implication in established pharmacological targets. This might provide strong support that these core characteristics may serve as potential targets for therapy. Once features are associated with biological pathways or molecular mechanisms, experimental validation such as CRISPR-based knockdown experiments or small molecules screen can validate the potentiality of the features as therapeutic targets. This approach, by linking SHAP-based features to biological literature, can contribute to a better understanding of their impact on drug discovery and possibly help identify new drug targets. It would also improve the interpretability of the SHAP study and bridge computational models with biological insights.

To further investigate the separability of druggable and non-druggable proteins in the learned feature space, the nonlinear dimensionality reduction method t-Distributed Stochastic Neighbor Embedding (t-SNE) was utilized [59]. Fig. 12 presents the t-SNE projections for the DDE, EAAC, and EGAAC feature encodings, providing a 2D visualization of the distribution of protein sequences. Each dot in the scatter plots represents a protein, with druggable proteins marked in blue and non-druggable proteins in red. The clustering patterns observed in the t-SNE plots suggest that different feature representations have varying degrees of discriminatory power. While some encodings exhibit better separation between druggable and non-druggable proteins, others display overlapping distributions, indicating that feature selection and fusion strategies may further improve classification performance.

To further analyze the underlying structure of the collected features and enhance the t-SNE and SHAP-based interpretations, Principal Component Analysis (PCA) was performed on all three feature types: DDE, EAAC, and EGAAC. PCA is a well-established linear dimensionality reduction method that converts high-dimensional data into a lower-dimensional space, preserving the directions of highest variance using orthogonal principle components [60].

Fig. 13 illustrates that the initial two principal components (PC1 and PC2) were employed to depict the class distribution across each feature set. Indicating that this descriptor captures discriminative patterns in the feature space, the PCA projection for DDE features showed a comparatively clearer separation between druggable and non-druggable proteins. Despite some reported class overlap in EAAC and EGAAC descriptors, significant grouping trends remained apparent. This indicates that even with a linear projection, the derived features retain class-relevant signals.

Although PCA may not properly represent intricate nonlinear manifolds like t-SNE, its linear projections provide an interpretable and complementary view of the spatial distribution of samples in the feature space. The congruence of PCA visualizations with t-SNE mappings and SHAP-derived feature importance substantiates the validity and consistency of our engineered feature sets across various interpretive dimensions, emphasizing the robustness of the proposed feature design strategy and its appropriateness for druggable protein prediction, even when assessed through different dimensionality reduction methodologies.

This figure shows how various models (Random Forest, Gradient Boosting, Support Vector Machine, Decision Tree, Logistic Regression, Voting (Soft), Stacking, Bagging) performed on the Precision-Recall (PR) curve. Each of these models has an AUC-PR score. Here is an overview of the AUC-PR comparison on all models. The best AUC-PR is achieved by RF, Voting (Soft), and Stacking (0.71 for all these models) in this competition. Bagging scores the worst with an AUC-PR of 0.61. Area Under the Precision-Recall Curve (AUC-PR) The AUC-PR is a performance metric used to evaluate how well a model can distinguish between positive and negative classes, particularly for imbalanced datasets, such as that employed in Druggable Protein Identification (refer to Fig. 14). The precision-recall curve (PR curve) allows us to understand to what degree the models are identifying the “druggable” proteins (positive class) vs. controlling the trade-off between minimizing the false positives (which are the proteins that are incorrectly labeled as druggable) and maximizing the precision (which is the correct recognition of druggable proteins). AUC-PR Curve Summary: Random Forest (AUC-PR = 0.71): It does a great job at retrieving the druggable proteins; the performance is high at all recall values. This indicates that RF does a good job of identifying druggable proteins while avoiding increasing the number of false positives. Gradient Boosting (AUC-PR = 0.69): The performance of this model is a bit lower than the Random Forest model, but it is still able to distinguish between druggable and non-druggable proteins quite well. Its performance isn’t as good as Random Forest at first glance, but it’s still good. SVM (AUC-PR = 0.63): Doesn’t crush Random Forest or Gradient Boosting, for sure. The plot visually demonstrates that it struggles in reducing the false positive rate, which translates into lower precision across different levels of recall. Decision Tree (A-UC-PR = 0.68): Better than SVM but not as good as Random Forest or Gradient Boosting. The curve indicates that the Decision Tree has difficulty in finding druggable proteins, especially when a large recall is achieved. Logistic Regression (AUC-PR = 0.65): It doesn’t work as well as the tree-based models (Random Forest, Gradient Boosting, Decision Tree). It has a smaller AUC-PR, indicating a less effective balance of precision and recall. Voting (Soft) (AUC-PR = 0.71): The Voting model (Soft) performs similarly to Random Forest. By applying a soft voting ensemble to average the predictions of multiple models, we can expect good performance while recognizing druggable proteins. Stacking (AUC-PR = 0.71): Stacking, similar to Voting (Soft), takes the best of a few models and has a good AUC-PR.

Under this setting, the best performer in these figures, Random Forest and Voting (Soft), demonstrates similar results. Bagging (AUC-PR = 0.61): This model shows the lowest AUC-PR value, indicating lower AUC-PR value, which indicates less competitiveness and more false positives. This indicates that Drn can identify druggable proteins with more difficulty than the other models in the study. The models, such as Random Forest, Voting (Soft), and Stacking, perform the best in terms of predicting druggable proteins, with the AUC-PR score of 0.71. They strike the best balance between precision and recall, crucial for finding proteins that drugs can target. Bagging, in contrast, does horribly, with AUC-PR well lower, which implies it may not be as good at discovering druggable proteins.

Description of Table 7: Return on Capital, AUC-PR, and MCC (Matthews Correlation Coefficient) metric scores. This table displays the evaluation measures AUC-PR (Area Under the Precision-Recall Curve) and MCC (Matthews Correlation Coefficient) for various model performances. These metrics are frequently employed in machine learning to evaluate the efficacy of categorization algorithms. This is an analysis of the columns and values in the table:

The initial column, denoted by numbers 0 to 8, serves as an index that presumably signifies several models or experiments for which the metrics are calculated. This statistic assesses the balance between precision and recall at various thresholds, however the area under the curve (AUC) provides a comprehensive evaluation of model performance across all potential classification levels. Elevated AUC-PR values signify superior performance, with values approaching 1 denoting an ideal model and values nearing 0 reflecting subpar performance. MCC quantifies the quality of binary classifications. It considers both true positives and true negatives, and is seen as more relevant than other accuracy metrics when addressing imbalanced datasets. MCC values span from −1 to +1, with +1 denoting a flawless forecast, 0 signifying a random guess, and −1 representing a wholly erroneous prediction. MCC: 0.320107 A moderate MCC score, signifying acceptable model performance, if not exemplary. Analogous to the AUC-PR score, the MCC value is marginally lower, indicating a relatively diminished model performance. The MCC of 0.316226 surpasses that of Index 1, signifying a marginally superior classification accuracy. This table provides a summary of model performance across various metrics (AUC-PR and MCC). It emphasizes areas where specific models exhibit superior or inferior performance, and also illustrates the effect of absent data (NaN values) on various models.