Machine Learning–Enhanced Quantitative Structure-Activity Relationship Modeling for DNA Polymerase Inhibitor Discovery: Algorithm Development and Validation

Introduction

Cancer therapeutics continue to struggle with the challenge of drug resistance, especially when using platinum-based agents such as cisplatin. These drugs induce cytotoxicity by creating DNA cross-links that interfere with DNA replication and transcription, ultimately leading to apoptosis [1-6]. However, resistance often develops through enhanced DNA repair mechanisms, particularly translesion DNA synthesis (TLS) [7-9]. TLS allows cancer cells to bypass cisplatin-induced DNA damage by using specialized DNA polymerases—most notably human DNA polymerase η (hpol η)—which can accurately replicate damaged DNA. Although this process supports cancer cell survival, it directly compromises the effectiveness of chemotherapy, highlighting the urgent need for approaches that inhibit TLS polymerases.

Targeting hpol η has emerged as a promising approach to counteract resistance [10-13]. Small-molecule inhibitors such as PNR-7-02, as demonstrated by Zafar et al [14], selectively disrupt hpol η’s TLS activity by binding to its “little finger” domain, misorienting the DNA template and stalling lesion bypass. This compound exhibits specificity for hpol η (IC₅₀=8 μM), sparing replicative polymerases and minimizing off-target effects [14]. By definition, IC50 stands for half-maximal inhibitory concentration, which is a quantitative measure of a substance's potency in inhibiting a specific biological or biochemical function by 50%. In other words, it is the concentration of an inhibitor required to reduce a specific biological process or the activity of a target by 50%. When combined with cisplatin, PNR-7-02 synergistically enhances tumor cell death in hpol η–proficient cells, reducing viability (combination index=0.4‐0.6) and amplifying DNA damage markers such as γH2AX [14]. Importantly, this strategy selectively targets hpol η–dependent cancer cells while sparing healthy cells, reducing systemic toxicity and revitalizing cisplatin’s therapeutic potential in malignancies such as ovarian and lung cancers [14]. Despite this initial progress, no existing inhibitor achieves complete DNA polymerase η inhibition, underscoring the critical need for novel small molecules with improved potency and specificity [15-21].

The search for such inhibitors is complicated by challenges related to target specificity, resistance evolution, and off-target effects. Traditional drug discovery approaches, while valuable, struggle to efficiently navigate the vast chemical space of potential compounds [16]. This limitation has spurred interest in computational strategies, particularly machine learning (ML)–enhanced quantitative structure-activity relationship (QSAR) modeling, which predicts biological activity based on molecular descriptors that quantitatively represent physicochemical, structural, and electronic properties [15-21]. ML has provided the computational power and strength needed to tackle critical questions across diverse fields [22-24], ranging from drug discovery to precision medicine. Conventional QSAR methods, though instrumental in early drug discovery, often lack accuracy and scalability when applied to complex datasets [25-27].

In this study, we present a systematic framework to optimize the identification of DNA polymerase inhibitors through artificial intelligence (AI)–driven QSAR modeling. By leveraging a curated database of 220 molecular descriptors with known activity against DNA polymerases, we trained 17 distinct ML models (eg, random forests, gradient boosting machines, support vector machines, and deep neural networks) and evaluated them across 14 performance metrics (refer to Table 1 for a summary of ML algorithms used in this study).

AI-driven QSAR modeling enables the prediction of inhibitor efficacy and identifies critical molecular features for second-generation optimization. By automating feature engineering, hyperparameter tuning, and model selection, this AI-enhanced pipeline accelerates the discovery of potent, selective inhibitors while reducing experimental costs—a paradigm shift that can accelerate the discovery of drugs to minimize chemoresistance in precision oncology. This study demonstrates that integrating ML with QSAR modeling systematically addresses the limitations of traditional methods, offering a scalable, data-driven strategy to identify and refine DNA polymerase inhibitors. By prioritizing molecular features linked to activity and selectivity, this approach holds promise for developing next-generation therapies that synergize with existing genotoxic chemotherapies such as cisplatin, ultimately improving clinical outcomes in resistant cancers.

Methods

The study used a curated library of 85 indole thio-barbituric acid (ITBA) analogs with experimentally validated inhibition of hpol η activity, expressed as the mean percent reduction in activity [14]. In total, 6 compounds (PNR-7-02, PNR-7-01, PN9-66B, PNR-6-92, PNR-6-89, and PNR-6-97) were excluded due to absence of reported hpol η activity, and 3 outliers (PNR-5-88, PNR-3-50, and PNR-3-64) were identified via scatter plots and IQR analysis and removed to ensure dataset integrity. Chemical structures, initially drafted in ChemDraw (Revvity Signals) [75], were converted to Simplified Molecular Input Line Entry System (SMILES) format and then to SYBYL Mol2 files using MAESTRO (version 12.5; Schrödinger, Inc) [76] for 3D visualization. Ligand preprocessing involved energy minimization to optimize molecular geometries and structural alignment of conserved ITBA cores, thus standardizing the presentation of side-chain modifications and ensuring consistent descriptor computation [16].

Molecular descriptors, which encompass a wide range of molecular properties, were calculated using MAESTRO software [76]. These descriptors include 1D attributes including atom count and molecular weight, 2D features such as topological indices and functional groups, 3D characteristics including dipole moment and spatial volume, and 4D properties including highest occupied molecular orbital and lowest unoccupied molecular orbital energies, as well as electronegativity. These descriptors provide insights into the electronic behavior of molecules during interactions, facilitating a comprehensive analysis of molecular structure and properties [76]. Such descriptors allowed quantitative comparisons of physicochemical attributes (eg, hydration energy and polarizability) and quantum chemical behavior critical for DNA polymerase interactions [16]. The resulting database integrated 220 descriptors with experimental inhibition data, forming the basis for QSAR modeling (refer to Multimedia Appendix 1 for molecular descriptors computed in MAESTRO software [76]).

Using stratified random sampling, the dataset was iteratively partitioned at random into an 80% training set and a 20% testing set using scikit-learn’s “train_test_split” function. This split ensures a robust training dataset for learning and a significant test dataset for accurate performance evaluation, while also maintaining the distribution of activity classes to overcome bias [77]. Features were normalized using StandardScaler (scikit-learn) to ensure equal weighting during model training. A total of 17 ML algorithms were evaluated (Table 1), spanning linear models (linear regression, ridge, lasso, and ElasticNet), tree-based ensembles (decision trees, random forest, gradient boosting, and AdaBoost), kernel methods (support vector regression), instance-based learning (K-nearest neighbors), neural networks (multilayer perceptron), probabilistic approaches (Gaussian process regression), dimensionality reduction (partial least squares regression), nonparametric models (isotonic regression), and advanced gradient-boosting frameworks (XGBoost, light gradient boosting machines [LightGBM], and CatBoost) [78,79]. Hyperparameters were optimized via grid or random search with 5-fold cross-validation, prioritizing minimization of mean square error (MSE) and maximization of coefficient of determination (R²) and adjusted coefficient of determination (adjusted R²) metrics.

Model performance was rigorously assessed using 14 metrics: mean squared error (MSE), coefficient of determination (R²), mean absolute error (MAE), root mean squared error (RMSE), adjusted coefficient of determination (adjusted R²), mean absolute percentage error (MAPE), predictive squared correlation (Q²), concordance correlation coefficient (CCC), root mean squared logarithmic error (RMSLE), normalized mean squared error (NMSE), normalized root mean squared error (NRMSE), symmetric mean absolute percentage error (SMAPE), median absolute error (MedAE), and Pearson correlation coefficient (PCC) [80].

MSE quantifies the average squared difference between predictions and observations and is calculated as follows:

${\displaystyle {\displaystyle {\displaystyle \begin{array}{c}MSE=\frac{1}{n}\sum _{i=1}^n{\left(y_i-{\hat{y}}_i\right)}^2\end{array}}}}$(1)

where $y_i$ is the observed value and ${\hat{y}}_i$ is the predicted value. MSE is critical for identifying models prone to severe inaccuracies.

RMSE provides error magnitude in the same units as the response variable, enhancing interpretability and sensitivity to outliers. It is calculated as follows:

${\displaystyle {\displaystyle \begin{array}{c}RMSE=\sqrt{\mathrm{M}\mathrm{S}\mathrm{E}}\end{array}}}$(2)

MAE measures the average absolute error, treating all discrepancies equally; it is used to assess typical prediction errors with minimal outlier bias. It is calculated as follows:

${\displaystyle {\displaystyle {\displaystyle \begin{array}{c}MAE=\frac{1}{n}\sum _{i=1}^n\left|y_i-{\hat{y}}_i\right|\end{array}}}}$(3)

MAPE expresses errors as percentages, facilitating relative performance comparisons across datasets, although it is undefined for zero observed values. It is calculated as follows:

${\displaystyle {\displaystyle {\displaystyle \begin{array}{c}MAPE=\frac{100\mathrm{\%}}{n}\sum _{i=1}^n \frac{\left|y_i-{\hat{y}}_i\right|}{\left|y_i\right|}\end{array}}}}$(4)

SMAPE addresses MAPE’s asymmetry by normalizing errors against the average of observed and predicted values, improving robustness for near-zero values. It is calculated as follows:

${\displaystyle {\displaystyle {\displaystyle \begin{array}{c}SMAPE=\frac{100\mathrm{\%}}{n}\sum _{i=1}^n \frac{2\left|y_i-{\hat{y}}_i\right|}{\left|y_i\right|+\left|{\hat{y}}_i\right|}\end{array}}}}$(5)

MedAE is resistant to outliers and is calculated as follows:

${\displaystyle {\displaystyle \begin{array}{c}MedAE=median\left(\left|y_1-{\hat{y}}_1\right|,\dots ,\left|y_n-{\hat{y}}_n\right|\right)\end{array}}}$(6)

$R^2$ represents the proportion of variance explained by the model, with values closer to 1 indicating better fit. It is calculated as follows:

${\displaystyle {\displaystyle {\displaystyle \begin{array}{c}{R}^2=1-\frac{\sum _{i=1}^n {\left(y_i-{\hat{y}}_i\right)}^2}{\sum _{i=1}^n {\left(y_i-\bar{y}\right)}^2}\end{array}}}}$(7)

where $\bar{y}$ is the mean of observed values and ${\hat{y}}_i$ represents the predicted or fitted value of the dependent variable (y) for the i-th observation.

Adjusted R² adjusts for model complexity, preventing overfitting by penalizing unnecessary predictors. It is calculated as follows:

${\displaystyle {\displaystyle \text{Adjusted }{R}^2=1-(1-R^2)\times \frac{n-1}{n-k-1}}}$(8)

where R²=R-squared of the model

n=number of observations (data points)

k=number of predictors (independent variables) in the model.

CCC evaluates agreement between predictions and observations, combining precision (correlation) and accuracy (mean shift). It is calculated as follows:

${\displaystyle {\displaystyle \begin{array}{c}CCC=\frac{2\rho {\sigma }_x{\sigma }_y}{{\sigma }_x^2+{\sigma }_y^2+{\left({\mu }_x-{\mu }_y\right)}^2}\end{array}}}$(9)

where $\rho$ is Pearson correlation, ${\mu }_x\mathrm{a}\mathrm{n}\mathrm{d}{\sigma }_x$ are mean and SD of observed values, and ${\mu }_y\text{ and }{\sigma }_y$ are mean and SD of the predicted values, respectively.

NMSE scales MSE by dataset variance, enabling cross-study comparisons. It is calculated as follows:

${\displaystyle {\displaystyle \begin{array}{c}NMSE=\frac{\mathrm{M}\mathrm{S}\mathrm{E}}{\mathrm{V}\mathrm{a}\mathrm{r}\left(y\right)}\end{array}}}$(10)

NRMSE provides a scale-free error metric, which is useful for comparing models across different units. It is calculated as follows:

${\displaystyle {\displaystyle {\displaystyle \begin{array}{c}NRMSE=\frac{\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}}{\mathrm{R}\mathrm{a}\mathrm{n}\mathrm{g}\mathrm{e}\left(y\right)}=\frac{\sqrt{\frac{1}{n}\sum _{i=1}^n{\left(y_i-{\hat{y}}_i\right)}^2}}{y_{max}-y_{min}}\end{array}}}}$(11)

Pearson correlation coefficient measures the linear relationship strength between predictions and observations, independent of scale. It is calculated as follows:

${\displaystyle {\displaystyle {\displaystyle \begin{array}{c}r=\frac{\sum _{i=1}^n \left(y_i-\bar{y}\right)\left({\hat{y}}_i-\overline{\stackrel{ˆ}{y}}\right)}{\sqrt{\sum _{i=1}^n {\left(y_i-\bar{y}\right)}^2\sum _{i=1}^n {\left({\hat{y}}_i-\overline{\stackrel{ˆ}{y}}\right)}^2}}\end{array}}}}$(12)

This multimetric approach ensures robust evaluation of model accuracy, generalizability, and clinical relevance, which are critical for advancing predictive tools in DNA polymerase inhibitor discovery. Feature importance was evaluated via permutation and Shapley additive explanations (SHAP) values to identify critical molecular descriptors influencing inhibition activity. The computational pipeline, implemented in Python (version 3.8; Python Software Foundation) [81], combined pandas for data manipulation, scikit-learn for model building, XGBoost, LightGBM, and CatBoost for gradient boosting, and SHAP for interpretability. Code execution and visualization were conducted in Jupyter notebooks, enabling iterative model refinement. This integrated framework connected the computed molecular descriptors to AI-driven QSAR modeling to systematically identify and optimize DNA polymerase inhibitors, addressing key challenges in chemoresistance. Figure 1 displays a graphical abstract for the methodology adopted for this study.

‎

Results

Overview of ML Performance Evaluation

The 17 ML models all led to robust predictions of compounds’ specific inhibition of hpol η, as evidenced by their training and testing performance metrics across all algorithms. Table 2 presents validation results for the training dataset, highlighting the models’ ability to learn from the data, while Table 3 displays results for the test datasets, providing insights into their generalization capabilities. Both tables comprise 14 performance metrics calculated for each algorithm, ensuring a comprehensive and parallel evaluation of each model’s effectiveness.

Model Performance Evaluation

In total, 17 ML models demonstrated robust predictive capabilities for DNA polymerase η (hpol η) inhibition activities, validated through comprehensive performance metrics (Table 2 and Table 3). Ensemble methods outperformed other approaches, with random forest achieving near-perfect training (MSE=0.0002; R²=0.9999) and testing performance (MSE=0.0003; R²=0.9998). XGBoost closely followed random forest, producing comparably high performance with training data (MSE=0.0001; R² = 0.9999) and testing data (MSE=0.0002; R²=0.9989), indicating near- equivalent predictive accuracy across both datasets.

Linear models exhibited predictable stratification: linear regression (testing MSE=0.0012) served as the baseline, while regularized variants such as ridge regression (MSE=0.0022) and lasso regression (MSE=0.0032) improved multicollinearity handling at the expense of accuracy. Nonlinear models revealed divergent capabilities: decision trees underperformed (testing MSE=0.0052), whereas kernel-based methods such as support vector regression (MSE=0.0006) surpassed neural networks (MSE=0.0008). Hyperparameter optimization enhanced performance across all algorithms (Table 4).

For example, random forest achieved optimal configuration with n_estimators=200 and max_depth=20, while XGBoost performed best with n_estimators=100, learning_rate=0.1, and max_depth=3. Model robustness was confirmed through CCC (CCC>0.9988) and low error ranges (MAE=0.0088‐0.051; RMSE=0.0141‐0.0721).

Feature Importance via SHAP Analysis

The SHAP summary plot identified r_desc_PEOE6 (electronic properties) as the most influential descriptor, with a mean absolute SHAP value 23% higher than the next best feature (Figure 2).

The second and third top-ranked features were r_qp_QPlogPC16 (partition coefficients) and i_desc_Sum_of_topological_distances_between_O.Cl (atom spacing), respectively. Secondary contributors included r_qp_PISA (polar surface area) and solvation indices such as r_desc_Solvation_connectivity_index_chi-4, which stabilized interactions within the polymerase active site. Lower-impact descriptors such as r_qp_FOSA (hydrophobic surface area) and r_qp_mol_MW (molecular weight) provided structural insights but contributed minimally to predictive reliability.

‎

Discussion

Principal Findings

The exceptional predictive performance of ensemble methods, particularly random forest and XGBoost, underscores their suitability for modeling the complex, nonlinear relationships inherent in hpol η inhibition [82-92]. Random forest achieved near-perfect testing metrics (MSE=0.0003; R²=0.9998), demonstrating robust generalization through feature space partitioning and aggregation of decision trees. This finding aligns with prior studies in which ensemble methods excelled for biological datasets, such as cancer transcriptome prediction of cell survival, due to their capacity to handle high-dimensional, sparse molecular descriptors [83,85-87]. The minimal performance gap between training and testing (ΔMSE=0.0001 [%]²) highlights effective overfitting mitigation, a critical advantage given the multicollinearity observed in QSAR datasets. XGBoost’s superior performance over neural networks (testing MSE=0.0002 vs 0.0008) further emphasizes gradient-boosted trees’ adaptability to sparse feature spaces, a finding consistent with their success in predicting protein-DNA binding affinity [23,84,92-96]. In contrast, linear models such as lasso regression (testing MSE=0.0032) revealed the necessity of regularization to manage sparsity, although at the cost of predictive accuracy—a trade-off well documented in drug discovery applications [84-93].

SHAP analysis identified electronic properties (r_desc_PEOE6) as the most critical determinant of inhibition activity, with a mean absolute SHAP value 23% higher than the second-ranked descriptor. This aligns with crystallographic evidence showing that charge distribution governs ligand binding stabilization in polymerase active sites [80,97]. The prominence of partition coefficients (r_qp_QPlogPC16) underscores lipophilicity’s dual role in cellular permeability and target engagement, a principle central to antiviral drug design [98,99]. Structural descriptors such as i_desc_Sum_of_topological_distances_between_O.Cl further emphasize steric complementarity requirements, mirroring findings in DNA polymerase β inhibition studies where atomic spacing dictated binding specificity [82,100-102]. Secondary features, including polar surface area (r_qp_PISA) and solvation indices (r_desc_Solvation_connectivity_index_chi-4) [83], elucidate how compounds stabilize aqueous-phase interactions, consistent with enzyme-substrate kinetic models [103-105]. While lower-impact descriptors (r_qp_FOSA, r_qp_mol_MW) provided auxiliary structural insights, their minimal contributions suggest prioritization of electronic and topological optimization in rational drug design [106,107].

The models’ consistent error distribution (MAPE: 0.89%‐5.1%) across activity ranges indicates reliability for moderate-activity compounds but exposes limitations in predicting extreme potencies. This mirrors similar challenges observed in solubility modeling, where outlier compounds often defy linear or ensemble-based predictions [108,109]. The clustering of MedAE around 0.01 suggests that while the models capture general trends, they struggle with highly potent inhibitors—a critical gap in drug discovery pipelines. This limitation likely stems from insufficient representation of extreme-activity compounds in training data, a common issue for biochemical datasets. Future work can address this limitation through synthetic minority oversampling or adversarial training techniques.

Methodologically, the integration of SHAP values bridges the interpretability-accuracy divide. While simpler models such as linear regression underperformed by 2 orders of magnitude, SHAP’s ability to deconvolute feature contributions enables actionable insights without sacrificing predictive power [82,83,110]. For instance, the identification of r_desc_PEOE6 as a top predictor provides a direct optimization target for medicinal chemists: tuning electronic properties to enhance binding affinity. Similarly, r_qp_QPlogPC16’s influence offers a pathway to balancing lipophilicity and solubility—a strategy validated in recent hpol η inhibitor development [83]. Integrating molecular-dynamic simulations may enhance predictive accuracy for structurally flexible compounds.

While our models emphasize solvation indices, Salgado et al [111] prioritized hydrogen-bonding descriptors in their polymerase inhibition studies. Discrepancies between these approaches may reflect hpol η’s uniquely hydrophobic active site, suggesting the need for crystallographic validation of descriptor-activity relationships. Conversely, consistency with the solvation models by Gupta et al [112] emphasizes the importance of aqueous-interaction stabilization in enzyme kinetics [104,113-115]. Such contrasts highlight the critical role of target-specific descriptor selection in QSAR workflows.

Translating these findings into drug discovery requires balancing multiparameter optimization. For example, improving r_desc_PEOE6 (electronic distribution) might conflict with r_qp_QPlogPC16 (lipophilicity) adjustments, necessitating Pareto front analysis to identify optimal compound profiles. Additionally, the moderate impact of r_qp_QPlogHERG (cardiac toxicity risk) implies the necessity for parallel absorption, distribution, metabolism, excretion, and toxicity profiling during lead optimization—a practice increasingly adopted in computational drug design.

This study establishes a predictive framework for hpol η inhibitors by combining ensemble methods (for accuracy) and SHAP analysis (for interpretability). The models prioritize electronic distribution, topological alignment, and solvation properties as critical descriptors, directly guiding rational drug design. The integration of these features underscores the need for multidimensional optimization in QSAR workflows, aligning with modern computational approaches.

Limitations and Future Directions

This study faced challenges in accurately predicting extreme values, highlighting the need for improved methodologies to address outlier prediction. Additionally, the absence of 3D conformational data limits the ability to model dynamic molecular interactions, which is crucial for capturing the full spectrum of polymerase-targeted binding events. Incorporating such structural information in future models will enhance the realism and predictive power of dynamic interaction analyses.

While SHAP analysis effectively identifies key molecular features, mechanistic interpretations, such as the role of r_desc_PEOE6 in binding pocket interactions, require validation through molecular dynamics simulations. This integration will strengthen the biological relevance of feature importance findings [1-5].

Furthermore, the current model’s applicability domain does not extend to metalloenzyme inhibitors, despite structural similarities among DNA polymerases. Expanding the training set to include these compounds could improve model generalizability and utility across a broader range of enzyme targets. Finally, future studies should explore hybrid modeling architectures that combine ensemble learning methods with graph neural networks. Such approaches may better capture both topological and electronic molecular effects, thereby refining QSAR methodologies for diverse enzyme systems. A key limitation of this study is the relatively small sample size compared to the large number of descriptors, which can increase the risk of overfitting despite the application of robust feature selection, regularization, and algorithmic strategies. Although our SHAP analysis identifies key molecular features, this study does not systematically assess pairs of structurally similar compounds with divergent activities, which is essential for fully evaluating model reliability and understanding potential activity cliffs. Addressing this limitation through focused analyses and validation in future studies will enhance the robustness and interpretability of our QSAR models. While our study claims superiority over linear models, it does not include direct comparisons with recent ML–based QSAR (MLQSAR) approaches, such as deep learning–based models. This limits our ability to fully contextualize our results within the broader scope of state-of-the-art MLQSAR studies. Although the high performance across our algorithms suggests model trustworthiness, future work will address this gap by benchmarking our models against advanced MLQSAR and deep learning methodologies. Additionally, this study evaluated model performance using only an internal 20% test split and did not include external validation with independent datasets or prospective testing. This limits the ability to fully assess the generalizability and real-world applicability of the models. Future work will incorporate validation using independent external datasets and prospective testing to rigorously evaluate model robustness, confirm generalizability, and strengthen confidence in their predictive performance in practical applications. Finally, critical toxicity descriptors (eg, r_qp_QPlogHERG) were identified but not optimized in this study. Future work will optimize these key toxicity descriptors to strengthen absorption, distribution, metabolism, excretion, and toxicity profiling and predictive safety. The current models show reduced performance for outliers and extreme inhibition values, which may impact predictive reliability. Future work will explore strategies such as data augmentation, robust loss functions, and uncertainty estimation to improve the models’ resilience to extreme values and enhance prediction accuracy across the full activity range. While this study achieves strong predictive performance, there is a lack of explicit analysis of the model’s applicability domain and chemical space coverage for novel ITBA analogs. Without thorough assessment of the regions in descriptor space where the model is most reliable, the generalizability to structurally diverse or previously unseen compounds remains uncertain. To address this, future work will incorporate formal applicability domain evaluation, such as leverage and distance-based techniques, to more precisely define the confidence boundaries of predictions and ensure robust extrapolation to new ITBA scaffolds and analogs. This will strengthen the practical utility of the model for prospective inhibitor discovery and design.

Conclusions

The ML-driven QSAR framework presented in this study overcomes cisplatin resistance challenges by identifying hpol η inhibitors with unprecedented precision. Ensemble methods (random forest and XGBoost) outperformed traditional models, capturing nonlinear relationships between molecular features and activity. SHAP analysis prioritized electronic distribution (r_desc_PEOE6), lipophilicity (r_qp_QPlogPC16), and structural topology (i_desc_Sum_of_topological_distances_between_O.Cl) as critical for efficacy, consistent with biochemical binding principles. While limitations persist in predicting extreme-potency compounds, the study provides actionable strategies to optimize inhibitor design. Future integration of dynamic 4D descriptors, experimental validation, and generative AI could accelerate development of next-generation therapies, revitalizing cisplatin-based treatments for resistant cancers through computationally guided precision.