Sequence-based prediction of drug–target binding using machine learning, deep learning and ensemble models without 3D structural information Authors: Nazife Çevik, Taner Çevik, Ahmet Gürhanlı Published: 2026 ABSTRACT Accurate prediction of drug–target interactions (DTIs) is a fundamental challenge in early-stage drug discovery, particularly in the absence of reliable three-dimensional structural information. In this study, we propose a fully sequence-based DTI prediction framework that eliminates dependence on structural data while achieving docking-comparable predictive performance. The proposed framework introduces a unified representation that systematically integrates physicochemical protein descriptors, protein 3-gram sequence motifs, and sequence-like drug encodings into a single feature space, enabling effective learning across heterogeneous models. A diverse set of machine learning, deep learning, and ensemble classifiers is evaluated under stratified five-fold cross-validation with class imbalance correction using Synthetic Minority Over-sampling Technique (SMOTE). Beyond individual models, the framework incorporates advanced ensemble strategies, including a stacking classifier that combines Random Forest, Support Vector Machine, and Logistic Regression, resulting in robust performance with ROC-AUC values exceeding 0.90 and a maximum AUC of 0.914. Importantly, the framework explicitly addresses model interpretability through feature importance analysis, revealing biologically meaningful protein sequence motifs associated with binding interactions. To further substantiate the reliability of the proposed approach, molecular docking experiments are conducted on a subset of predicted drug–target pairs, and the observed agreement between docking scores and predicted binding probabilities provides independent validation. Collectively, this study demonstrates that carefully engineered sequence-derived representations, coupled with optimized ensemble learning, constitute a scalable, interpretable, and computationally efficient alternative to structure-dependent DTI prediction methods. Introduction Drug–target interactions (DTIs) are fundamental to therapeutic development, as the pharmacological efficacy of a small molecule relies on its ability to selectively bind to and modulate a specific protein target. Conventional strategies for characterizing DTIs depend on experimental assays or structure-based molecular docking, where protein–ligand binding is estimated through conformational sampling and scoring functions1. Despite its widespread adoption, docking accuracy is often constrained by the availability and quality of three-dimensional (3D) protein structures, assumptions regarding receptor flexibility, and sensitivity to scoring-function parametrization2,3. These limitations have stimulated an increasing transition toward learning-based approaches that infer binding directly from molecular data, particularly when structural information is incomplete, uncertain, or unavailable4,5. Early computational efforts in DTI prediction leveraged similarity-based assumptions, exploiting chemical and sequence-level resemblances to infer potential interactions. Building on these foundations, learning-based methods have progressively evolved to incorporate increasingly expressive molecular representations and predictive models. In recent years, data-driven approaches operating directly on molecular sequences have gained prominence due to their scalability and applicability to targets lacking reliable structural annotations. Despite substantial progress, several challenges remain unresolved. Many high-performing computational frameworks implicitly rely on structural proxies or predicted spatial features, which may not generalize across diverse protein families. Purely sequence-based models, while more broadly applicable, often employ opaque representations that limit interpretability and hinder biological insight. In addition, methodological issues related to dataset construction, class imbalance, information leakage, and evaluation protocols have been shown to inflate reported performance if not carefully addressed6,7. These limitations highlight the need for DTI prediction frameworks that are not only accurate, but also transparent, robust, and biologically interpretable. Interpretability has therefore emerged as an essential consideration in computational drug discovery. Beyond predictive performance, there is increasing demand for models that can elucidate which molecular and sequence-level features drive binding predictions and whether these features align with established biophysical principles. Explainable artificial intelligence techniques offer promising tools in this regard; however, systematic interpretability analyses remain limited in fully sequence-based DTI pipelines. Motivated by these challenges, this study develops a comprehensive sequence-based drug–target interaction prediction framework that functions independently of three-dimensional structural data while upholding interpretability and maintaining competitive predictive accuracy. Unlike methods that predominantly depend on graph-based molecular representations, structure-derived inputs, or opaque latent embeddings, the proposed framework is constructed upon a unified feature representation that incorporates physicochemical protein descriptors, BLOSUM62-based evolutionary information, protein sequence motifs, and sequence-derived drug encodings within a singular modeling pipeline. Classical machine learning, deep learning, and ensemble models were all used to test the final version. During the whole experiment, a strict stratified cross-validation approach with fold-wise class imbalance correction was used. In addition to assessing predictive performance, the study explicitly addresses interpretability through feature-level analysis and substantiates the biological plausibility of the predictions via docking-based orthogonal assessment. By integrating sequence-only applicability, interpretability, and a broad spectrum of modeling approaches, this study aims to provide a practical and transparent alternative to existing drug–target interaction prediction frameworks. In this context, the present study makes four principal contributions to the existing body of literature. Firstly, it proposes a fully sequence-based drug-target interaction (DTI) prediction framework that does not necessitate the use of experimentally resolved or predicted three-dimensional (3D) structural representations during the model development process. Secondly, it introduces a unified and biologically interpretable feature space that integrates complementary protein and drug descriptors, including physicochemical properties of proteins, evolutionary substitution information, local protein sequence motifs, and sequence-derived drug encodings. Thirdly, it offers a systematic comparative evaluation of classical machine learning methods, deep learning models, and ensemble strategies within a consistent validation protocol that explicitly addresses class imbalance. Fourthly, it enhances the predictive evaluation with interpretability analysis and docking-based orthogonal assessments, thereby establishing a connection between predictive performance and biological plausibility. Collectively, these contributions position the proposed framework as an interpretable and practically applicable alternative to methods that rely more heavily on structural information or exhibit lesser transparency. The current study concentrates on the systematic integration and evaluation of complementary sequence-derived representations within a unified structure-agnostic DTI prediction framework, rather than the introduction of a fundamentally novel learning algorithm. The remainder of this paper is organized as follows. Section “Related work” reviews related work on computational DTI prediction, emphasizing methodological limitations and evaluation challenges. Section “Materials and methods” describes the dataset, feature engineering procedures, modeling framework, and validation protocols. Section “Results” presents the experimental results, including predictive performance, interpretability analyses, and docking-based validation. Section “Discussion” discusses the findings and their implications, and Sect. “Conclusion” concludes the paper with directions for future research. Related work Early studies on drug–target interaction (DTI) prediction were primarily based on similarity-driven assumptions, exploiting chemical fingerprints, protein sequence homology, and interaction network proximity to infer potential bindings8–11. While effective in mitigating data sparsity, these approaches depend heavily on predefined similarity metrics and are limited in their ability to model complex, nonlinear binding mechanisms. Matrix-factorization-based and graph-regularized approaches further improved DTI prediction by integrating chemical, genomic, and topological similarity information into latent interaction models12,13. Gradient-boosting-based methods such as SimBoost also demonstrated the effectiveness of engineered pairwise representations for drug–target binding prediction14. The advent of deep learning enabled automatic feature extraction from molecular sequences, with convolutional neural networks (CNNs) demonstrating improved predictive performance for DTI tasks15–17. Subsequent attention-based and transformer architectures further enhanced representation learning by capturing long-range dependencies between residues and ligand substructures18–20. However, these models typically rely on highly abstract latent embeddings, which hinders interpretability and limits biological insight into binding mechanisms. Graph-based methods introduced explicit structural representations for drug molecules and, in some cases, proteins, yielding strong empirical performance21–27. Despite their success, many graph-centric and hybrid architectures incorporate structural proxies—such as contact maps or inferred spatial features—which may not generalize reliably to proteins lacking structural annotation. Beyond modeling choices, recent studies have emphasized methodological concerns in DTI research, including information leakage through homologous sequences, inconsistent dataset partitioning, and biased evaluation protocols that can inflate reported performance6,7. In parallel, explainable artificial intelligence techniques have been explored to improve model transparency28–32, yet systematic interpretability analyses remain limited in fully sequence-based DTI pipelines. Representative studies published recently include a model integrating drug chemical structures with protein sequences for DTI network prediction33, a protein-feature and wrapper-selection framework34, and a fine-grained selection plus bidirectional random-walk approach35, illustrating the diversity of machine-learning strategies applied to DTI prediction. Machine learning has become an important component of modern drug discovery pipelines, spanning screening, optimization, and safety assessment36. Graph convolutional networks have also been applied directly to drug–target interaction prediction, demonstrating the value of relational molecular representations37. In contrast to existing approaches, the present study focuses exclusively on sequence-derived representations while integrating complementary, biologically interpretable features within a unified framework and enforcing rigorous evaluation and explainability protocols. Materials and methods This section describes the datasets, feature extraction procedures, learning models, and evaluation protocols employed in the proposed sequence-based drug–target interaction prediction framework. Figure 1 provides an overview of the end-to-end pipeline, illustrating the main stages of data preprocessing, feature extraction, model training, evaluation, and explainability analysis. Fig. 1End-to-end pipeline of the sequence-driven drug–target interaction prediction method. End-to-end pipeline of the sequence-driven drug–target interaction prediction method. Dataset description Data on drug–target interactions has been collected from evaluated repositories with experimentally validated connections. Small-molecule compounds were obtained from DrugBank, and their canonical SMILES representations were utilized to generate sequence-like encodings and physical characteristics. Protein targets were depicted by their primary amino acid sequences sourced from verified protein databases. Binary interaction labels were established based on reported binding evidence. For orthogonal structural validation, experimentally determined three-dimensional protein structures were sourced from the Protein Data Bank (PDB). Ligand structures were derived from canonical SMILES strings and transformed into three-dimensional conformations before docking. Molecular docking simulations were performed utilizing AutoDock Vina with defined parameters. The binding affinity estimations were compared with machine-learning projected interaction probabilities to evaluate the consistency between sequence-based inference and structure-based assessment. All datasets were manually processed to eliminate incomplete or inconsistent information. Data preprocessing, scaling, and oversampling techniques were meticulously used within each training fold during cross-validation to avert information leakage and provide an impartial performance assessment. The drug–target interaction dataset used in this study consists of experimentally validated binding and non-binding drug–protein pairs collected from publicly available sources. Each interaction is represented as a binary label indicating the presence or absence of binding. To enhance dataset transparency, comprehensive dataset statistics were included in the revised manuscript. Positive samples are empirically confirmed drug-target interaction pairs sourced from reviewed public repositories, while negative samples were generated from drug-protein pairs. Protein targets are provided as amino acid sequences, while drug compounds are represented using SMILES strings. To ensure consistency and reproducibility, sequences containing ambiguous characters or incomplete annotations are excluded during preprocessing. Given the inherent imbalance commonly observed in DTI datasets—where non-interacting pairs significantly outnumber confirmed interactions—the dataset exhibits a skewed class distribution. This imbalance is explicitly addressed during model training and evaluation, as described in subsequent sections. All preprocessing steps are performed prior to feature extraction to prevent unintended information leakage across training and test splits. Table 1 illustrates the quantity of medicines, proteins, positive interactions, negative interactions, and the corresponding class distribution employed in the study. All preprocessing and filtering techniques were applied before feature extraction and cross-validation prevent potential information loss. Table 1Dataset statistics and sample construction summary.StatisticValueNumber of Drugs303Number of Proteins5067Positive drug-target interaction pairs524Negative drug-target interaction pairs524Total Interaction Pairs1048Positive/Negative Ratio524:524Removed Incomplete Entries1 Table 1Dataset statistics and sample construction summary.StatisticValueNumber of Drugs303Number of Proteins5067Positive drug-target interaction pairs524Negative drug-target interaction pairs524Total Interaction Pairs1048Positive/Negative Ratio524:524Removed Incomplete Entries1 Dataset statistics and sample construction summary. Sequence-based feature engineering Protein representation Protein sequences are transformed into numerical representations using a combination of complementary sequence-derived feature classes. First, physicochemical descriptors are computed to capture global properties of amino acid sequences, including hydrophobicity, polarity, charge distribution, and molecular weight–related characteristics. These descriptors provide biologically interpretable summaries of protein composition relevant to molecular recognition. Second, evolutionary similarity information is incorporated through BLOSUM62-based encoding, which captures substitution preferences reflecting evolutionary conservation patterns across protein families. This representation enables the model to leverage sequence homology without relying on explicit structural annotations. Finally, protein sequences are encoded using n-gram motif representations, specifically 3-gram patterns, to capture local sequence contexts and short-range residue interactions. The resulting motif frequency vectors provide a balance between expressive capacity and interpretability, allowing the identification of sequence motifs potentially associated with binding behavior. Let the protein sequence be defined as P = (a₁, a₂, …, aₙ) where ai denotes the amino acid residue at position i and n represents the sequence length. A set of overlapping 3-gram motifs was generated using a sliding window of size three: gi = (ai, ai₊₁, ai₊₂), i = 1,2,…,n − 2. Thus, each consecutive triplet of amino acids forms a unique local sequence motif. The complete collection of extracted 3-grams is represented as G = {g₁, g₂, …, gₙ₋₂}. The occurrence frequency of each distinct 3-gram motif was then calculated to construct the protein motif feature vector: v₃g = [f₁, f₂, …, fₘ] where m denotes the total number of unique 3-gram motifs retained after preprocessing and fⱼ corresponds to the normalized frequency of the j-th motif: fⱼ = cⱼ/\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sum\:_{k=1}^{m}{c}_{k}$$\end{document} where cⱼ represents the raw occurrence count of the j-th 3-gram motif in the protein sequence. Drug representation Drug molecules are represented using sequence-like numerical encodings derived from their SMILES strings. Each SMILES sequence is converted into a fixed-length representation that preserves the sequential ordering of chemical symbols, enabling compatibility with both classical machine learning models and deep learning architectures. This representation allows the framework to operate without explicit molecular graph construction while retaining essential chemical information relevant to ligand binding. Let the canonical SMILES representation of a drug molecule be defined as S = (s₁, s₂, …, sₖ) where si denotes the i-th token in the SMILES sequence and k represents the total sequence length. Overlapping subsequence motifs were generated from the SMILES sequence using a sliding window strategy. For a motif length of l, each molecular subsequence was defined as mi = (si, si₊₁, …, si₊ₗ₋₁), i = 1,2,…,k − l+1. The complete set of extracted molecular motifs is represented as M = {m₁, m₂, …, mₖ₋ₗ₊₁}. The occurrence frequencies of the resulting motifs were then computed to generate the drug feature vector: vdrug = [d₁, d₂, …, dₚ] where p denotes the total number of unique molecular motifs retained after preprocessing and dⱼ represents the normalized occurrence frequency of the j-th motif: dⱼ = cⱼ/\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sum\:_{t=1}^{p}{c}_{t}$$\end{document} where cⱼ corresponds to the raw occurrence count of the associated molecular motif within the SMILES sequence. Feature integration Protein and drug feature vectors are concatenated to form joint representations for each drug–target pair. Prior to model training, all features are standardized to zero mean and unit variance to ensure numerical stability and to prevent dominance of features with larger scales. Feature integration is performed independently within each training fold to avoid data leakage. For each drug–target pair, the final joint representation vector was constructed by concatenating the protein-derived and drug-derived feature vectors into a unified numerical representation: x = [vphys || vBLOSUM || v3g || vdrug] where || denotes vector concatenation, vphys represents the physicochemical protein descriptor vector, vBLOSUM corresponds to the evolutionary similarity representation derived from BLOSUM62 encoding, v3g denotes the normalized protein 3-gram motif vector, and vdrug represents the sequence-derived molecular descriptor vector extracted from canonical SMILES strings. Assuming dimensionalities of vphys ∈ Rᵈ¹, vBLOSUM ∈ Rᵈ², v3g ∈ Rᵈ³, vdrug ∈ Rᵈ⁴,the final feature representation satisfies x ∈ \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}^{({d}^{1}+{d}^{2}+{d}^{3}+{d}^{4})}$$\end{document}. Prior to model training, all feature dimensions were standardized using z-score normalization: x’i = (xi − µi)/σi. where xi denotes the original feature value, µi and σi represent the mean and standard deviation estimated exclusively from the corresponding training fold, and x’i denotes the normalized feature value. This normalization procedure was applied independently within each cross-validation fold to prevent information leakage. Handling class imbalance To mitigate the effects of class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) is applied exclusively to the training subsets within each cross-validation fold38. SMOTE generates synthetic samples of the minority class by interpolating between existing positive instances in feature space, thereby improving class balance without duplicating samples. Importantly, no resampling is performed on validation or test data to preserve unbiased performance estimation. To address class imbalance during model training, the SMOTE was applied independently within each training fold. Let xi ∈ Rᵈ denote a minority-class feature vector in a d-dimensional feature space, and let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{nn}$$\end{document} ∈ Rᵈ represent one of its k-nearest minority-class neighbors. A synthetic sample was generated by linear interpolation between these two samples according to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{new}$$\end{document} = \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}$$\end{document}+ λ\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:({x}_{nn}-\:{x}_{i})$$\end{document} where λ ~ U(0,1) is a random interpolation coefficient sampled from a uniform distribution over the interval [0,1]. Thus, each synthetic instance lies on the line segment connecting = \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{nn}$$\end{document}in feature space. This procedure increases minority-class density while preserving the local geometric structure of the minority-class distribution. To avoid information leakage, SMOTE was applied exclusively to the training partition within each cross-validation fold, while validation and test partitions remained untouched. The preprocessing and evaluation pipeline applied within each cross-validation fold is summarized below: Split the dataset using stratified 5-fold cross-validation.Select one fold as the validation/test partition.Apply feature normalization using statistics computed only from the training fold.Apply SMOTE exclusively to the training fold.Train the classifier on the resampled training data.Evaluate model performance on the untouched validation/test fold.Repeat the procedure for all folds and aggregate performance metrics. Split the dataset using stratified 5-fold cross-validation. Split the dataset using stratified 5-fold cross-validation. Select one fold as the validation/test partition. Select one fold as the validation/test partition. Apply feature normalization using statistics computed only from the training fold. Apply feature normalization using statistics computed only from the training fold. Apply SMOTE exclusively to the training fold. Apply SMOTE exclusively to the training fold. Train the classifier on the resampled training data. Train the classifier on the resampled training data. Evaluate model performance on the untouched validation/test fold. Evaluate model performance on the untouched validation/test fold. Repeat the procedure for all folds and aggregate performance metrics. Repeat the procedure for all folds and aggregate performance metrics. This fold-wise strategy ensured that no information from the validation/test partitions was used during normalization or synthetic sample generation, thereby preventing information leakage. Learning models A diverse set of predictive models is employed to evaluate the effectiveness of the proposed feature representations across different learning paradigms. Classical machine learning classifiers include Logistic Regression, Support Vector Machine (SVM), and Random Forest, which provide strong baselines and varying degrees of interpretability. Random Forest was included as a robust ensemble baseline because it models nonlinear feature interactions via aggregated decision trees39. In addition, a Multilayer Perceptron (MLP) and a Convolutional Neural Network (CNN) are used to assess the performance of deep learning approaches on the sequence-derived features. To further enhance predictive robustness, advanced ensemble methods—including Extra Trees, Gradient Boosting, Histogram-Based Gradient Boosting, and a stacking classifier—are evaluated. The stacking ensemble integrates predictions from Random Forest, SVM, and Logistic Regression models via a meta-learner, enabling complementary decision patterns to be combined into a unified prediction. Training and evaluation protocol Model training and evaluation are conducted using stratified five-fold cross-validation to ensure consistent class proportions across folds. All preprocessing steps, including feature scaling and SMOTE-based resampling, are performed within each training fold to prevent information leakage. Hyperparameters are optimized using grid-based search strategies applied exclusively to the training data. Performance is assessed using a comprehensive set of evaluation metrics, including accuracy, precision, recall, F1-score, receiver operating characteristic area under the curve (ROC-AUC), Matthews correlation coefficient (MCC), Cohen’s kappa, and specificity. In addition to accuracy and F1-score, MCC was included because it is particularly informative for binary classification under class-imbalanced conditions40. This multi-metric evaluation strategy provides a balanced assessment of model behavior, particularly under class imbalance conditions. The complete set of hyperparameter configurations used for training and optimizing all models is summarized in Table 2. Table 2Comparative evaluation of sequence-based DTI prediction models using multiple classification metrics.ModelAccuracyPrecisionRecallF1-ScoreROC-AUCSpecificityMCCKappaBalanced AccuracyStacking (RF + SVM+LR)0.840.880.790.830.910.890.690.680.84Random Forest0.840.890.770.820.910.900.680.680.84ExtraTrees0.840.910.760.830.900.920.690.690.84HistGradientBoosting0.830.840.810.830.900.850.670.660.83GradientBoosting0.840.850.820.830.900.850.680.670.84CNN0.790.800.780.790.880.810.590.590.79Logistic Regression0.790.780.810.790.860.770.580.580.79SVM0.790.780.790.790.840.780.580.580.79MLP0.740.720.770.740.800.700.480.480.74 Table 2Comparative evaluation of sequence-based DTI prediction models using multiple classification metrics.ModelAccuracyPrecisionRecallF1-ScoreROC-AUCSpecificityMCCKappaBalanced AccuracyStacking (RF + SVM+LR)0.840.880.790.830.910.890.690.680.84Random Forest0.840.890.770.820.910.900.680.680.84ExtraTrees0.840.910.760.830.900.920.690.690.84HistGradientBoosting0.830.840.810.830.900.850.670.660.83GradientBoosting0.840.850.820.830.900.850.680.670.84CNN0.790.800.780.790.880.810.590.590.79Logistic Regression0.790.780.810.790.860.770.580.580.79SVM0.790.780.790.790.840.780.580.580.79MLP0.740.720.770.740.800.700.480.480.74 Comparative evaluation of sequence-based DTI prediction models using multiple classification metrics. Model interpretability To enhance transparency and biological interpretability, model predictions are analyzed using explainable artificial intelligence techniques. Permutation importance is employed to estimate the global contribution of individual features to model performance, while SHAP (SHapley Additive exPlanations) values are used to quantify feature-level contributions to individual predictions through additive attribution analysis41. These analyses enable the identification of protein sequence motifs and physicochemical properties that are most influential in determining binding likelihood. Docking-based orthogonal validation To provide an independent validation of the sequence-based predictions, molecular docking experiments are conducted on a selected subset of predicted drug–target pairs. Figure 2 illustrates the docking-based validation workflow used to evaluate the biological plausibility of the predicted interactions. Fig. 2Docking-based orthogonal validation workflow. Docking-based orthogonal validation workflow. Protein structures generated using AlphaFold2 are first filtered based on confidence metrics, including predicted Local Distance Difference Test (pLDDT) scores and Predicted Aligned Error (PAE). Protein structures predicted by AlphaFold2 were considered for docking-based validation, with predicted models obtained from the AlphaFold Protein Structure Database when appropriate42,43. Drug molecules are converted from SMILES strings into three-dimensional conformations, protonated, and energy-minimized prior to docking. Docking simulations are performed using AutoDock Vina to estimate binding affinities and generate plausible binding conformations. The resulting docking scores are compared with the predicted binding probabilities produced by the sequence-based models. Consistency between these independent assessments serves as an orthogonal validation of the proposed framework. Results Overall model performance The overall predictive performance of the evaluated models was assessed using stratified five-fold cross-validation and a comprehensive set of evaluation metrics. The baseline performance of all classifiers without class imbalance correction is summarized in Table 3, providing a quantitative comparison across accuracy, precision, recall, F1-score and ROC-AUC. Table 3Baseline model performance before SMOTE (5-fold CV).ModelAccuracy (mean ± std)F1-Score (mean ± std)ROC-AUC (mean ± std)MLP0.7876 ± 0.0170.7908 ± 0.01870.8665 ± 0.0166CNN0.8133 ± 0.01140.8086 ± 0.01020.8919 ± 0.0038Random Forest0.8422 ± 0.0160.8327 ± 0.01890.9128 ± 0.0164Logistic Regression0.8184 ± 0.01130.8176 ± 0.01010.8750 ± 0.0039SVM0.8137 ± 0.01590.8082 ± 0.01520.8634 ± 0.0036 Table 3Baseline model performance before SMOTE (5-fold CV).ModelAccuracy (mean ± std)F1-Score (mean ± std)ROC-AUC (mean ± std)MLP0.7876 ± 0.0170.7908 ± 0.01870.8665 ± 0.0166CNN0.8133 ± 0.01140.8086 ± 0.01020.8919 ± 0.0038Random Forest0.8422 ± 0.0160.8327 ± 0.01890.9128 ± 0.0164Logistic Regression0.8184 ± 0.01130.8176 ± 0.01010.8750 ± 0.0039SVM0.8137 ± 0.01590.8082 ± 0.01520.8634 ± 0.0036 Baseline model performance before SMOTE (5-fold CV). As reported in Table 3, ensemble-based methods generally achieve stronger overall discrimination capability compared to individual classifiers, as reflected by higher ROC-AUC and MCC values. Deep learning models also demonstrate competitive baseline performance, particularly in terms of recall and F1-score, indicating an effective ability to identify positive drug–target interactions. In contrast, classical linear models, such as Logistic Regression, exhibit comparatively lower recall under baseline conditions, suggesting greater sensitivity to class imbalance when resampling is not applied. Overall, the baseline results summarized in Table 3 establish a reference point for subsequent analyses and highlight the influence of model choice on predictive performance prior to explicit class imbalance handling. Impact of class imbalance handling The effect of class imbalance handling on model performance was systematically evaluated by comparing results obtained before and after the application of the Synthetic Minority Over-sampling Technique (SMOTE). Quantitative performance metrics after resampling are summarized in Table 4, while a comprehensive visual comparison across evaluation metrics is provided in Fig. 3. As shown in Table 4, the impact of SMOTE on predicted performance differed among models and assessment criteria. Although resampling typically enhanced sensitivity and recall for minority classes across many classifiers, its effect on Accuracy, F1-score, and ROC-AUC was contingent upon the specific model employed. In many instances, slight decreases in overall discrimination scores were noted upon resampling, albeit enhanced class balance. These data indicate that SMOTE may increase minority-class detection while concurrently modifying decision boundaries in a manner that does not consistently enhance all performance metrics. The findings suggest that the efficacy of addressing class imbalance is contingent upon the classifier architecture and the assessment metric employed. Table 4Model performance after SMOTE (5-fold CV).ModelAccuracy (mean ± std)F1-Score (mean ± std)ROC-AUC (mean ± std)MLP0.7343 ± 0.04920.7424 ± 0.04730.8020 ± 0.0623CNN0.7952 ± 0.01280.7868 ± 0.01970.8818 ± 0.0149Random Forest0.8343 ± 0.02180.8219 ± 0.02680.9101 ± 0.0113Logistic Regression0.7952 ± 0.01200.7991 ± 0.01210.8637 ± 0.0122SVM0.7838 ± 0.02340.7881 ± 0.02580.8454 ± 0.0126 Table 4Model performance after SMOTE (5-fold CV).ModelAccuracy (mean ± std)F1-Score (mean ± std)ROC-AUC (mean ± std)MLP0.7343 ± 0.04920.7424 ± 0.04730.8020 ± 0.0623CNN0.7952 ± 0.01280.7868 ± 0.01970.8818 ± 0.0149Random Forest0.8343 ± 0.02180.8219 ± 0.02680.9101 ± 0.0113Logistic Regression0.7952 ± 0.01200.7991 ± 0.01210.8637 ± 0.0122SVM0.7838 ± 0.02340.7881 ± 0.02580.8454 ± 0.0126 Model performance after SMOTE (5-fold CV). The heatmap visualization in Fig. 3 further illustrates the impact of SMOTE by highlighting relative changes across multiple evaluation metrics simultaneously. In particular, recall and F1-score exhibit the most notable increases after resampling, while ROC-AUC values remain stable or improve slightly for most models. This indicates that improved classification balance is achieved without compromising overall discrimination capability. Taken together, these results demonstrate that explicit handling of class imbalance is a critical factor in reliable drug–target interaction prediction and provides a necessary foundation for subsequent model comparison and interpretability analyses. Fig. 3Heatmap comparison of model performance metrics before and after SMOTE for CNN, Logistic Regression, MLP, Random Forest, and SVM models. The figure summarizes Accuracy, F1-Score, and ROC-AUC values obtained under 5-fold stratified cross-validation. SMOTE produced model-dependent effects across evaluation metrics, generally improving minority-class sensitivity while introducing variable changes in overall discrimination performance. Heatmap comparison of model performance metrics before and after SMOTE for CNN, Logistic Regression, MLP, Random Forest, and SVM models. The figure summarizes Accuracy, F1-Score, and ROC-AUC values obtained under 5-fold stratified cross-validation. SMOTE produced model-dependent effects across evaluation metrics, generally improving minority-class sensitivity while introducing variable changes in overall discrimination performance. Comparison of classical, deep learning, and ensemble models A comparative analysis of classical machine learning, deep learning, and ensemble-based models was conducted to assess their relative strengths under the same experimental protocol. The comparative performance trends across these model families are summarized in Fig. 4, which provides a consolidated view of discrimination capability across key evaluation metrics. As illustrated in Fig. 4, ensemble-based models consistently achieve superior performance compared to both classical and deep learning approaches, particularly in terms of ROC-AUC and MCC. This trend indicates a stronger ability of ensemble methods to capture complex, non-linear relationships within the integrated sequence-derived feature space. Among the ensemble strategies, tree-based ensembles exhibit stable performance profiles with reduced variance across cross-validation folds. Deep learning models demonstrate competitive performance, especially with respect to recall and F1-score, reflecting their effectiveness in identifying positive drug–target interactions. However, their overall discrimination capability remains slightly lower than that of ensemble models under baseline conditions. Classical machine learning models, while providing solid baseline performance, generally exhibit lower recall and higher sensitivity to class imbalance, resulting in reduced performance on metrics that emphasize minority class detection. Fig. 4Comparison of classical, deep learning, and ensemble model performances across evaluation metrics. Comparison of classical, deep learning, and ensemble model performances across evaluation metrics. These trends are further reflected in the detailed metric breakdown provided in Table 2. Table 2 presents an in-depth analysis of all evaluation measures across the analyzed models, facilitating a common comparison of classical machine learning, deep learning, boosting-based, and ensemble-based methodologies under the same cross-validation technique. Beside typical classification measures, additional performance indicators—such as Matthews correlation coefficient (MCC), Cohen’s Kappa, balanced accuracy, and specificity—were investigated to deliver a more dependable evaluation in conditions of class imbalance. Ensemble-based techniques exhibited superior discrimination capabilities, particularly in ROC-AUC, MCC, and specificity metrics, although individual models, notably CNN and Random Forest, maintained competitiveness in recall and F1 measurements. Of the analyzed methodologies, the stacking classifier attained the superior ROC-AUC performance, demonstrating the efficacy of combining various base learners within a cohesive decision framework. The findings demonstrate that boosting and stacking techniques enhance predictive stability and alignment between anticipated and actual labels beyond random chance, as evidenced by increased MCC and Cohen’s Kappa values. All tabulated values represent mean performance across stratified cross-validation folds. Overall, the results shown in Fig. 4 highlight clear performance stratification among model families, with ensemble methods providing the most robust and consistent predictive behavior, followed by deep learning models and classical classifiers. This comparison establishes a quantitative foundation for subsequent analyses of learning dynamics, error patterns, and interpretability presented in the following sections. Learning dynamics of deep learning models The learning behavior of the deep learning models was examined to assess convergence characteristics, stability across training epochs, and consistency across cross-validation folds. The evolution of training and validation accuracy and loss for the convolutional neural network (CNN) is illustrated in Fig. 5(a), providing insight into model convergence and generalization behavior. As shown in Fig. 5, the CNN exhibits a steady decrease in training loss accompanied by a corresponding increase in training accuracy, indicating effective optimization over successive epochs. Validation curves follow similar trends, suggesting that the model achieves convergence without pronounced divergence between training and validation performance. This behavior indicates stable learning dynamics under the selected training configuration. An alternative visualization of the loss evolution is presented in Fig. 5(b), which further highlights the convergence pattern of the CNN across epochs. The loss curves demonstrate a gradual stabilization after an initial learning phase, with no abrupt oscillations or divergence observed. This consistency supports the robustness of the training process and suggests that the selected hyperparameters provide adequate control over the optimization trajectory. To evaluate training stability across different data partitions, fold-wise learning dynamics are illustrated in Fig. 6, which summarizes performance trends across cross-validation folds for the deep learning models. As shown, learning curves remain largely consistent across folds, with limited variance in convergence behavior. This observation indicates that the deep learning models exhibit stable training dynamics and are not overly sensitive to specific train–validation splits. Overall, the learning dynamics presented in Fig. 5 demonstrate that the deep learning models converge reliably under the adopted training protocol, providing a sound basis for subsequent analyses of classification behavior and interpretability. Fig. 5Learning dynamics of the CNN model across training epochs: (a) training and validation accuracy curves, and (b) training and validation loss curves. The figure illustrates the convergence behavior, optimization stability, and generalization performance of the CNN architecture during model training. Learning dynamics of the CNN model across training epochs: (a) training and validation accuracy curves, and (b) training and validation loss curves. The figure illustrates the convergence behavior, optimization stability, and generalization performance of the CNN architecture during model training. Fig. 6Training and validation loss curves of the CNN model across training epochs. The figure demonstrates the fold-wise learning dynamics and loss convergence behavior of the CNN architecture during model training. Training and validation loss curves of the CNN model across training epochs. The figure demonstrates the fold-wise learning dynamics and loss convergence behavior of the CNN architecture during model training. Confusion matrix analysis of key models To further analyze classification behavior beyond aggregate performance metrics, confusion matrices were examined for representative models from different model families. The confusion matrices for the convolutional neural network (CNN), Logistic Regression, Random Forest, and Support Vector Machine (SVM) are presented in Fig. 7, respectively. As illustrated in Fig. 7, the CNN demonstrates a balanced distribution between true positive and true negative predictions, indicating effective detection of interacting drug–target pairs while maintaining reasonable control over false positives. The corresponding confusion matrix suggests that the CNN achieves a favorable trade-off between sensitivity and specificity under the adopted training protocol. The confusion matrix for Logistic Regression reveals a comparatively higher number of false negatives, reflecting reduced sensitivity toward the minority class under baseline conditions. This behavior is consistent with the performance trends observed in earlier sections and highlights the limitations of linear decision boundaries when class imbalance is present. In contrast, the Random Forest confusion matrix exhibits improved classification balance, with a higher proportion of true positives and fewer misclassifications overall. This pattern reflects the model’s ability to capture non-linear feature interactions and to aggregate multiple decision paths effectively. Similarly, the SVM confusion matrix demonstrates enhanced discrimination compared to linear models, with a more favorable distribution of true positives and true negatives. However, some degree of misclassification remains, indicating that model performance is influenced by both feature representation and decision boundary complexity. Overall, the confusion matrix analysis provides complementary insight into model behavior by revealing error patterns that are not fully captured by scalar performance metrics alone. These results help contextualize the comparative performance of different model families and motivate further analysis of feature contributions and decision mechanisms. Fig. 7CNN, logistic regression, random forest and SVM confusion matrices respectively. CNN, logistic regression, random forest and SVM confusion matrices respectively. Feature importance and explainability results To investigate the factors driving model predictions and to enhance transparency, feature importance and explainability analyses were conducted using model-agnostic and model-specific techniques. The global feature importance scores derived from permutation-based analysis are summarized in Fig. 8, providing an overview of the most influential features contributing to binding prediction. As shown in Fig. 8, protein-derived features consistently rank among the most important contributors to model performance. In particular, physicochemical descriptors related to hydrophobicity, charge distribution, and residue composition exhibit high importance scores, indicating their substantial influence on predictive outcomes. In addition to global descriptors, several sequence-derived features, including specific 3-gram protein motifs, are identified as prominent contributors, highlighting the relevance of local sequence context in determining binding likelihood. The feature importance analysis further reveals that drug-derived sequence encodings contribute meaningfully to prediction, although their relative influence is generally lower than that of protein-derived features. This observation suggests that, within the proposed unified feature representation, protein sequence characteristics play a dominant role in shaping model decisions. Overall, the results presented in Fig. 8 demonstrate that the proposed framework leverages biologically interpretable features to drive prediction, offering insight into the relative contributions of different feature classes. These findings provide a quantitative foundation for subsequent discussion of biological relevance and model interpretability. Fig. 8Permutation-based feature importance analysis showing the relative contribution of the most informative protein and drug sequence-derived features. The x-axis represents normalized importance scores computed from the decrease in model performance after feature permutation. Permutation-based feature importance analysis showing the relative contribution of the most informative protein and drug sequence-derived features. The x-axis represents normalized importance scores computed from the decrease in model performance after feature permutation. ROC curve analysis Receiver operating characteristic (ROC) curve analysis was performed to evaluate the discrimination capability of the proposed models across different decision thresholds. ROC curves provide a threshold-independent assessment of model performance and are particularly informative under class imbalance conditions. The ROC curves for the deep learning models are presented in Figs. 9 and 10, corresponding to the convolutional neural network (CNN) and the multilayer perceptron (MLP), respectively. As shown in these figures, both models achieve strong discrimination capability, with smooth ROC profiles indicating stable performance across thresholds. Fig. 9ROC curve of the CNN classifier obtained from a representative evaluation fold. The displayed ROC-AUC value corresponds to a single cross-validation fold and may therefore differ slightly from the mean ROC-AUC value reported in Table 2. ROC curve of the CNN classifier obtained from a representative evaluation fold. The displayed ROC-AUC value corresponds to a single cross-validation fold and may therefore differ slightly from the mean ROC-AUC value reported in Table 2. Fig. 10ROC curve of the MLP classifier obtained from a representative evaluation fold. The displayed ROC-AUC value corresponds to a single cross-validation fold and may therefore differ slightly from the mean ROC-AUC value reported in Table 2. ROC curve of the MLP classifier obtained from a representative evaluation fold. The displayed ROC-AUC value corresponds to a single cross-validation fold and may therefore differ slightly from the mean ROC-AUC value reported in Table 2. A comparative ROC analysis of classical machine learning models, including Random Forest, Logistic Regression, and Support Vector Machine, is shown in Fig. 11. The figure highlights differences in discrimination behavior among these models, with ensemble-based and kernel-based approaches generally exhibiting higher true positive rates at lower false positive rates. Fig. 11ROC curves of random forest, logistic regression, and SVM evaluated on the SMOTE-balanced test set. The displayed AUC values correspond to this specific evaluation split and may differ from the mean cross-validation ROC-AUC values reported in Table 2. ROC curves of random forest, logistic regression, and SVM evaluated on the SMOTE-balanced test set. The displayed AUC values correspond to this specific evaluation split and may differ from the mean cross-validation ROC-AUC values reported in Table 2. To provide an integrated view across all evaluated models, Fig. 12 presents ROC curves aggregated over the stratified five-fold cross-validation procedure. This figure illustrates consistent performance trends across folds and confirms the robustness of the evaluated models under repeated resampling. Fig. 12ROC curves for all models (5-Fold Cross-Validation). ROC curves for all models (5-Fold Cross-Validation). Model-specific ROC curves are further detailed in Figs. 13, 14 and 15, which show individual ROC profiles for Logistic Regression, Random Forest, and Support Vector Machine, respectively. These figures enable closer inspection of discrimination behavior at the model level and complement the comparative analyses presented earlier. Fig. 13ROC curve of the logistic regression classifier evaluated on a representative cross-validation fold. The figure illustrates the discrimination capability of the model between interacting and non-interacting drug–target pairs. ROC curve of the logistic regression classifier evaluated on a representative cross-validation fold. The figure illustrates the discrimination capability of the model between interacting and non-interacting drug–target pairs. Fig. 14ROC curve of the random forest model classifier evaluated on a representative cross-validation fold. The figure illustrates the discrimination capability of the model between interacting and non-interacting drug–target pairs. ROC curve of the random forest model classifier evaluated on a representative cross-validation fold. The figure illustrates the discrimination capability of the model between interacting and non-interacting drug–target pairs. Fig. 15ROC curve of the SVM model classifier evaluated on a representative cross-validation fold. The figure illustrates the discrimination capability of the model between interacting and non-interacting drug–target pairs. ROC curve of the SVM model classifier evaluated on a representative cross-validation fold. The figure illustrates the discrimination capability of the model between interacting and non-interacting drug–target pairs. Overall, the ROC analyses presented in Figs. 9, 10, 11, 12, 13, 14 and 15 demonstrate that the proposed framework achieves strong and consistent discrimination performance across a diverse set of learning models, providing further evidence of its effectiveness for sequence-based drug–target interaction prediction. To further examine discrimination behavior at a finer granularity, additional ROC analyses are presented in Fig. 16, which illustrates model-specific and fold-level performance variations. These extended curves demonstrate consistent discrimination trends across cross-validation folds and provide further support for the robustness of the proposed framework under different data partitions. Fig. 16ROC curves (5-fold CV, Stage 4 final models) for Stacking (RF + SVM+LR), Gradient Boosting, ExtraTrees, and HistGradientBoosting. ROC curves (5-fold CV, Stage 4 final models) for Stacking (RF + SVM+LR), Gradient Boosting, ExtraTrees, and HistGradientBoosting. Docking-based case illustration To further illustrate the biological plausibility of the proposed sequence-based predictions, a representative docking-based case analysis was conducted. Figure 17 presents a comparative visualization of structure-based docking results and sequence-only machine-learning predictions for a selected drug–target pair. The comparison demonstrates consistency between high predicted binding probability and favorable docking configuration, providing qualitative support for the learned sequence-derived representations. While this analysis does not constitute quantitative validation, it offers a mechanistic illustration of how sequence-based predictions align with structure-informed binding assessments. This illustrative case complements the docking-based validation described in Sect. “Docking-based orthogonal validation”. Fig. 17Representative comparison between sequence-based binding prediction and docking-derived interaction assessment. Left: AutoDock Vina docking pose with top-r anked affinities (–6.4 and − 6.2 kcal/mol), showing the ligand fitted in the EGFR active site. Right: Binding probability estimated by the sequence-only model and the corresponding feature contributions (SHAP), illustrating how the ML framework infers interaction likelihood without requiring 3D structural inputs. Representative comparison between sequence-based binding prediction and docking-derived interaction assessment. Left: AutoDock Vina docking pose with top-r anked affinities (–6.4 and − 6.2 kcal/mol), showing the ligand fitted in the EGFR active site. Right: Binding probability estimated by the sequence-only model and the corresponding feature contributions (SHAP), illustrating how the ML framework infers interaction likelihood without requiring 3D structural inputs. Discussion The experimental results presented in Sect. Results demonstrate that sequence-derived representations, when combined with appropriately selected learning strategies, can yield reliable and robust predictions of drug–target interactions without reliance on three-dimensional structural information. Across all evaluated scenarios, model performance is strongly influenced by both the choice of learning paradigm and the handling of dataset imbalance, underscoring the importance of methodological design in sequence-based DTI prediction. The baseline results reported in Sect. “Overall model performance” establish that ensemble-based models achieve stronger discrimination capability than individual classifiers under imbalanced conditions, even before explicit resampling is applied. This observation suggests that aggregating complementary decision boundaries enables ensemble methods to partially compensate for class imbalance, a trend that becomes more pronounced following resampling. Model-dependent effects ofSMOTE application in Sect. “Impact of class ımbalance handling” further confirm that explicit imbalance handling is critical for enhancing sensitivity toward true interactions while preserving overall discrimination performance. The comparative analysis across model families in Sect. “Comparison of classical, deep learning, and ensemble models” highlights clear performance stratification between classical machine learning, deep learning, and ensemble approaches. Ensemble models consistently demonstrate superior robustness, while deep learning models exhibit competitive recall-oriented performance. These findings indicate that different model families capture complementary aspects of the underlying sequence–binding relationship, motivating the use of ensemble strategies for large-scale virtual screening tasks. Interpretability analyses presented in Sect. “Feature ımportance and explainability results” demonstrate that the proposed framework leverages biologically meaningful features to drive prediction. The dominance of protein-derived physicochemical descriptors and local sequence motifs aligns with established principles of molecular recognition, such as hydrophobic complementarity and residue-level conservation. Importantly, these findings indicate that sequence-based models need not rely on opaque latent representations to achieve strong performance, and that interpretable feature engineering remains a viable and informative strategy. This study only examines a framework that is entirely sequence-based and architecturally agnostic. In the future, researchers may employ graph neural representations, contrastive learning techniques, and multimodal feature fusion methods to enhance the accuracy and comprehensibility of predictions in biological contexts. Furthermore, comparing the proposed framework with recently introduced large-scale DTI architectures may provide a more comprehensive understanding of its competitiveness across various biological contexts. While layered cross-validation allows for the prediction of model stability with limited data availability, the lack of an independent test set represents a limitation of the current study. Specifically, sequence similarity between cross-validation layers can introduce a bias in performance prediction. Therefore, future studies will aim to further evaluate the ability to generalize the proposed method across various protein families and ligand classes by focusing on assessing it on externally compiled datasets and sequence splitting settings. The docking-based analysis presented in this study should be interpreted as an orthogonal qualitative validation rather than a comprehensive quantitative benchmark across the entire dataset. Docking simulations were performed on representative drug-target pairs to assess the consistency between biological probability and sequence-based predictions versus structure-based interaction assessment. While the observed agreement supports the validity of the proposed framework, a formal statistical correlation analysis between predicted binding probabilities and docking affinity scores was not performed. Therefore, future studies will focus on large-scale docking assessment and quantitative agreement analysis using broader structurally validated benchmark datasets. The current assessment technique employs stratified cross-validation instead of cold-start validation, despite the fact that fold-wise preprocessing and resampling are implemented to prevent direct information leakage during model training. Consequently, the emergence of structurally or sequentially comparable proteins and chemically related compounds across folds may lead to optimistic performance estimations. Consequently, the framework’s generalization in biological situations that are becoming more challenging will be evaluated through the investigation of sequence-identity-aware partitioning, scaffold-based drug splitting, and target-level cold-start evaluation protocols. Conclusion This study presented a fully sequence-based framework for drug–target interaction prediction that operates independently of three-dimensional structural information while achieving robust and competitive predictive performance. By integrating physicochemical protein descriptors, sequence-derived n-gram motifs, and sequence-like drug representations within a unified modeling pipeline, the proposed approach effectively captures informative patterns underlying molecular binding. Comprehensive experimental evaluation across classical machine learning, deep learning, and ensemble models demonstrated that ensemble-based strategies provide the most consistent and robust performance, particularly under class imbalance conditions. The systematic analysis of class imbalance handling confirmed that explicit resampling plays a critical role in improving sensitivity toward true interactions without compromising overall discrimination capability. Learning dynamics, confusion matrix analyses, and ROC evaluations further illustrated the stability and reliability of the proposed models across cross-validation folds and decision thresholds. Beyond predictive accuracy, the framework explicitly addressed model interpretability by identifying biologically meaningful sequence-derived features that contribute to binding predictions. Feature importance analyses highlighted the dominant role of protein-derived physicochemical properties and local sequence motifs, supporting alignment with established principles of molecular recognition. In addition, docking-based orthogonal validation provided independent, structure-informed support for the biological plausibility of the sequence-based predictions. Overall, the proposed framework offers a scalable, interpretable, and computationally efficient alternative to structure-dependent drug–target interaction prediction methods, making it well suited for large-scale virtual screening and early-stage drug discovery applications. Future work will explore the integration of pretrained sequence embeddings, uncertainty-aware learning strategies, and more stringent generalization benchmarks to further enhance robustness and applicability.