Utilizing machine learning and molecular dynamics for enhanced drug delivery in nanoparticle systems Authors: Alireza Jahandoost, Razieh Dashti, Mahboobeh Houshmand, Seyyed Abed Hosseini Published: 2024 ABSTRACT Materials data science and machine learning (ML) are pivotal in advancing cancer treatment strategies beyond traditional methods like chemotherapy. Nanotherapeutics, which merge nanotechnology with targeted drug delivery, exemplify this advancement by offering improved precision and reduced side effects in cancer therapy. The development of these nanotherapeutic agents depends critically on understanding nanoparticle (NP) properties and their biological interactions, often analyzed through molecular dynamics (MD) simulations. This study enhances these analyses by integrating ML with MD simulations, significantly improving both prediction accuracy and computational efficiency. We introduce a comprehensive three-stage methodology for predicting the solvent-accessible surface area (SASA) of NPs, which is crucial for their therapeutic efficacy. The process involves training an ML model to forecast the many-body tensor representation (MBTR) for future time steps, applying data augmentation to increase dataset realism, and refining the SASA predictor with both augmented and original data. Results demonstrate that our methodology can predict SASA values 299 time steps ahead with a 40-fold speed improvement and a 25% accuracy increase over existing methods. Importantly, it provides a 300-fold increase in computational speed compared to traditional simulation techniques, offering substantial cost and time savings for nanotherapeutic research and development. Introduction Cancer, a leading cause of mortality worldwide, manifests as a result of a complex interplay of environmental factors, lifestyle choices, and genetic predispositions1. This disease disrupts normal cellular function, displaying hallmark traits that promote its destructive progression, including uncontrolled proliferation, immune evasion, and support for metastasis through angiogenesis2. Current treatment options predominantly involve harsh chemotherapy, which can harm healthy cells and often face limitations due to cellular resistance3. Therefore, a comprehensive approach combining surgery, pharmacotherapy, or radiotherapy is commonly employed in clinical settings, albeit with inherent drawbacks4. In light of these challenges, nanotherapeutics emerge as a promising avenue for precise and less invasive treatment modalities in the realm of cancer therapy. These cutting-edge technologies offer the potential to overcome existing limitations by delivering targeted therapies directly to cancerous cells while minimizing harm to healthy tissues5. Nanoparticles (NPs), ranging from 1 to 100 nanometers, exhibit unique properties such as enhanced reactivity and optical behavior due to their small size. They are classified into various types, including metal, ceramic, polymeric, and carbon-based, and can take on shapes like spheres, rods, cubes, or irregular forms. Their applications are vast, from drug delivery to catalysis, electronics, and environmental remediation6. In medical applications, NPs offer significant advantages, such as improved targeting and efficiency, enhancing drug delivery by enabling precise, targeted therapy, minimizing side effects, and improving treatment efficacy7. Nanotherapeutics, a groundbreaking field at the intersection of nanotechnology and drug delivery, is reshaping the landscape of medical treatments8. The field of nanotechnology is considered to be among the most groundbreaking technologies of the 21st century. The unique physicochemical characteristics of materials at the nanoscale have been utilized across various industrial sectors, with anticipated significant advancements in biotechnology, medicine, and pharmaceutical endeavors. Nanotechnology indeed has the potential to offer resolutions to numerous unresolved challenges in contemporary medicine concerning disease prevention, diagnosis, and treatment, thereby presenting advantageous benefits for healthcare professionals, individual patients, and society at large9. NPs have garnered significant interest in biological science and biomedical technology as a promising toolset, owing to their controllable physicochemical properties and utility for studying living organisms. Specifically, the potential use of NPs in drug delivery systems, bioimaging, and theranostics has significantly escalated in the last decade. Moreover, substantial attention has been focused on exploring the interactions of NPs with biological systems to create highly effective nanomedicine for safe clinical application10. NP therapies offer several benefits over traditional small molecule therapies, particularly in the context of targeted effects. Unlike small molecules, which often lack specificity and can affect healthy cells along with cancerous ones, NPs can be engineered to target specific cancer cells with high precision. This targeted delivery enhances the therapeutic efficacy and minimizes off-target effects, leading to reduced side effects and improved patient outcomes11. Additionally, NPs can be designed to deliver drugs more efficiently, overcome biological barriers, and provide sustained release, further enhancing their effectiveness compared to traditional small-molecule therapies7,12. By skillfully utilizing nanoscale technology, nanotherapeutics precisely targets diseased cells while safeguarding healthy tissues, revolutionizing the efficacy and safety of treatments13. These advanced drug carriers not only reduce the need for high dosages but also improve treatment outcomes through controlled release mechanisms14. Tailoring NPs with specialized ligands enables researchers to direct them to specific tissues, enhancing drug concentration at the intended site and minimizing off-target effects15. The intricate design of NPs, encompassing their surface, shell, and core layers, is pivotal in influencing their behavior in the body16. Factors like size, shape, and composition are meticulously manipulated to optimize drug delivery efficiency, requiring a profound understanding of the interplay between chemical properties and biological responses17. Evaluating parameters such as solvent-accessible surface area (SASA) becomes crucial in predicting drug performance and refining delivery systems for maximal effectiveness. The pursuit of NPs with enhanced SASA values highlights the necessity of pushing the boundaries of nanotherapeutics research, despite the substantial time and financial investments involved18. While the challenges within nanotherapeutics development are formidable, the potential transformative impact on cancer treatment stresses the urgency and importance of these efforts. By overcoming these obstacles, researchers are paving the way for a future characterized by personalized and targeted therapies that promise to enhance patient outcomes and quality of life19. The use of molecular dynamics (MD) simulations has emerged as a robust and dependable approach for designing favorable structures. Initially conceptualized in the late 1970s, the MD simulation technique has undergone significant progression20, evolving from simulating small atomic systems to complex biological entities such as proteins in solution, proteins within membranes, and large macromolecular assemblies like nucleosomes or ribosomes21. Standard practices now include simulations involving tens of thousands to hundreds of thousands of atoms, with even larger simulations approaching half a million atoms being feasible with accessible computational resources22. MD simulations revolve around computer-based analyses of the dynamic motions of atoms and molecules, providing insights into the behavior of these systems over defined periods23. By numerically solving Newtons equations of motion, interactions among particles are simulated, characterized by forces and potential energies typically governed by interatomic potentials or molecular mechanical force fields. Widely applicable in chemical physics, materials science, and biophysics, MD simulations face challenges primarily related to the refinement of force fields and the substantial computational resources required, limiting the duration and exploration of conformational states in simulations24,25. Despite their effectiveness in predicting molecular movements accurately, MD simulations are constrained by the approximations inherent in the force fields, particularly when quantum phenomena are crucial, such as interactions involving transition metal atoms26. Addressing these limitations involves leveraging advanced hardware technologies like graphics processing units (GPUs) to enhance computational performance. However, adapting MD simulation algorithms to GPU architectures requires extensive refinement and tuning to maximize efficiency on these specialized platforms27-29. Machine learning (ML) techniques have attracted considerable attention in recent years as an innovative approach for analyzing the structure, performance, and synthesis of materials. ML methodologies are adept at processing large and complex datasets derived from experimental studies, including those characterized by variability or inconsistencies. ML algorithms are good at filtering out irrelevant information (interference) and handling variations in the data. This ability is essential for accurately predicting biological responses because biological systems are often complex and involve significant variability. By effectively managing these challenges, ML models can make more reliable predictions30. Furthermore, ML offers advantages in terms of speed and accuracy across a wide range of applications. For instance, Wang et al.31 proposed an active learning framework to streamline the optimization process for fabricating titanium dioxide nanotube micro-patterns (TNMs), thereby accelerating materials manufacturing and optimization. The integration of ML techniques in addressing challenges related to hardware architecture has demonstrated significant promise in the development of MD simulations and molecular modeling32. Wang and colleagues thorough review on the application of ML-driven approaches in analyzing and enhancing MD simulations highlights the pivotal role of artificial neural networks (ANNs) in processing high-dimensional data from MD simulations33,34. By leveraging various types of ANNs to generate latent vectors in a lower-dimensional feature space based on trajectory data, researchers have been able to more effectively evaluate the equilibrium and dynamic characteristics of systems. Moreover, the active integration of ML-based methodologies within simulation procedures has shown a potential to improve sampling efficiency and capacity35. Despite these advancements, a critical challenge remains in ensuring model interpretability and transferability to novel systems, essential for achieving dual objectives36. In addressing contemporary challenges, Kibria et al.18 presented an innovative approach that merged cutting-edge models for novel NP designs with enhanced interpretability using explainable Artificial Intelligence. Their methodological breakthrough involved a three-pronged strategy: commencing with data transformation via the versatile many-body tensor representation (MBTR) descriptor. MBTR captured structural nuances by incorporating an array of unique structural patterns, enabling comprehensive analysis of both finite and periodic systems37. Following this, a hybrid ML network was deployed to precisely forecast SASA values, with feature significance elucidated through Shapley additive explanations (SHAP) values. By fusing a time series model for MD interactions with a deep neural network (DNN)-based model for SASA determination, this investigation showcases the promise of ML techniques in cost-efficient simulations of NP designs. These findings hold substantial promise for navigating the NP design landscape and predicting pivotal properties. However, Kibria et al.18 exhibited elevated error and employed a window size of 40. In light of these limitations, Kibria et al. work is selected as a baseline, and this study aims to develop a method that enhances both computational efficiency (by reducing the window size) and prediction accuracy. Reducing window size offers significant advantages, as MD simulation is time-intensive and can require days of computation. Consequently, a reduced window size can substantially decrease simulation time, facilitating the acquisition of results more expeditiously. This investigation proposes a three-phased methodology shown in Fig. 1. In the initial phase, 72 linear regression (LR) models are trained to forecast the subsequent 72 MBTR values. The second phase introduces a data augmentation technique to generate more realistic data for the SASA model training. The third phase entails training an extra trees regressor (ETR) SASA model using both original and augmented data. The primary contributions of this study are enumerated below: LR emerges as the optimal ML algorithm for predicting the next MBTR value, particularly when data scarcity exists.The proposed data augmentation technique effectively enhances the prediction accuracy of the SASA predictor.Due to its inherent randomness, the ETR algorithm exhibits exceptional performance in SASA prediction and significantly reduces overfitting. LR emerges as the optimal ML algorithm for predicting the next MBTR value, particularly when data scarcity exists. LR emerges as the optimal ML algorithm for predicting the next MBTR value, particularly when data scarcity exists. The proposed data augmentation technique effectively enhances the prediction accuracy of the SASA predictor. The proposed data augmentation technique effectively enhances the prediction accuracy of the SASA predictor. Due to its inherent randomness, the ETR algorithm exhibits exceptional performance in SASA prediction and significantly reduces overfitting. Due to its inherent randomness, the ETR algorithm exhibits exceptional performance in SASA prediction and significantly reduces overfitting. This study is meticulously structured into distinct sections. Section Materials and Methods encompasses a comprehensive exposition of the materials and methodologies employed, alongside a precise definition of the problem under investigation. Section Proposed method elucidates the intricacies of the proposed method, providing a detailed account of its theoretical underpinnings and implementation. Section Results and Discussion presents and analyzes the obtained results, offering insights into the efficacy and implications of the proposed approach. Section Limitations and Future Works demonstrates this studys limitations and future works. Finally, Sect. Conclusion concludes the study by summarizing the key findings, highlighting the contributions of the study, and outlining potential avenues for future research. Fig. 1Schematic representation of the proposed method. The proposed methodology comprises three distinct phases: (a) MBTR model training: in this phase, 72 LR models are trained to predict the subsequent MBTR values. (b) Data augmentation: this phase involves the generation of augmented data to enhance the realism of the training data for the SASA model. (c) SASA model training: the final phase entails training an ETR model, which is capable of predicting the SASA value corresponding to each MBTR value. Schematic representation of the proposed method. The proposed methodology comprises three distinct phases: (a) MBTR model training: in this phase, 72 LR models are trained to predict the subsequent MBTR values. (b) Data augmentation: this phase involves the generation of augmented data to enhance the realism of the training data for the SASA model. (c) SASA model training: the final phase entails training an ETR model, which is capable of predicting the SASA value corresponding to each MBTR value. Materials and methods Problem definition Consider a matrix, MBTR, with dimensions (timesteps, n features), representing MBTR for each NP design. Additionally, let SASA be a vector with length timesteps, representing SASA at each timestep. The objective of the model is to approximate the relationship defined in Eq. (1).1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$SAS{A^{ - 1}}=\mathcal{F}\left( {MBT{R^{0:window~size}};\theta } \right),$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$SAS{A^{ - 1}}$$\end{document} is the SASA value of the last timestep, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MBT{R^{0:window~size}}$$\end{document} the MBTR value of timesteps zero to window size, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta$$\end{document} is the parameters. Modeling such a function poses significant challenges. To address this complexity, the problem is decomposed into two more manageable sub-problems. The first sub-problem involves modeling a function, as represented by Eq. (2), capable of predicting the MBTR value for the subsequent timestep. The second sub-problem entails modeling a function, as represented by Eq. (3), capable of predicting the SASA value corresponding to MBTR at each timestep.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MBT{R^t}={\mathcal{F}_{MBTR}}\left( {MBT{R^{t - window~size:t}};\theta } \right),$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$SAS{A^t}={\mathcal{F}_{SASA}}\left( {MBT{R^t};\theta } \right).$$\end{document} After solving these two sub-problems, the main problem can be solved using the flowchart in Fig. 2. Fig. 2Flowchart of solving the main problem after addressing the sub-problems. Flowchart of solving the main problem after addressing the sub-problems. Dataset This research utilized the same dataset as presented by Kibria et al.18. The dataset comprises 64 NP designs simulated over 300 nanoseconds, 32 over 200 nanoseconds, and 23 over 120 nanoseconds. These designs were subsequently divided into training and test sets with an 80:20 ratio, resulting in 107 designs allocated for training and 12 designs for testing. As previously mentioned, each MBTR value possesses 72 features, and the ultimate objective is to predict the final SASA value for each design. LR LR is a fundamental regression model that establishes a linear relationship between independent and dependent variables. Its hypothesis function is mathematically represented by Eq. (4). To determine the optimal set of parameters, the model employs a cost function, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$J\left( \theta \right)$$\end{document}, defined in Eq. (5). Minimizing this cost function is achieved through gradient descent, as outlined in Eq. (6).4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h_\theta }\left( X \right)={\theta ^T}X,$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$J\left( \theta \right)=\frac{1}{2}\mathop \sum \limits_{{i=1}}^{m} {\left( {{h_\theta }\left( {{X^{\left( i \right)}}} \right) - {y^{\left( i \right)}}} \right)^2},$$\end{document}6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\theta _{n+1}}={\theta _n} - \alpha ~\nabla J\left( {{\theta _n}} \right),$$\end{document} where x is the input matrix (features), y is the output vector (targets), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document} is the learning rate. ETR ETR38 is a tree-based regression algorithm closely related to random forest regressor (RFR)39. However, ETR employs a distinct strategy in selecting features for node splitting. Rather than selecting the optimal split based on a random subset of features, ETR randomly selects a split among a random subset of features. This randomization technique effectively reduces variance in certain scenarios, enhancing the algorithms predictive performance. Notably, ETR prioritizes computational efficiency, resulting in a faster execution time compared to other tree-based methods38. Metrics In addition to the widely recognized metrics, mean absolute error (MAE) and root mean squared error (RMSE), defined in Eqs. (7) and (8) respectively, two additional metrics are introduced for more comprehensive analysis. The first metric, ErrorMBTR, represents the mean error in predicting the MBTR value of the last time step for test NP designs using the initial window size timesteps. Its mathematical definition is provided in Eq. (9). The second metric, ErrorSASA, signifies the mean error in predicting the SASA value of the last time step for test designs using the initial window size timesteps. Its mathematical definition is given in Eq. (10).7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MAE=\frac{1}{m}~\mathop \sum \limits_{{i=1}}^{m} \left| {{y^{\left( i \right)}} - {{\hat {y}}^{\left( i \right)}}} \right|,$$\end{document}8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$RMSE=\sqrt {\frac{1}{m}\mathop \sum \limits_{{i=1}}^{m} {{\left( {{y^{\left( i \right)}} - {{\hat {y}}^{\left( i \right)}}} \right)}^2}} ,$$\end{document}9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Erro{r_{MBTR}}=\frac{1}{k}\mathop \sum \limits_{{i=1}}^{k} \left| {MBT{R^{\left( i \right)~ - 1}} - {{\widehat {{MBTR}}}^{\left( i \right)~ - 1}}} \right|,$$\end{document}10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Erro{r_{SASA}}=\frac{1}{k}\mathop \sum \limits_{{i=1}}^{k} \left| {SAS{A^{\left( i \right)~ - 1}} - {{\widehat {{SASA}}}^{\left( i \right)~ - 1}}} \right|,$$\end{document} where m is the number of samples, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${y^{\left( i \right)}}$$\end{document} is the true value of ith sample, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat {y}^{\left( i \right)}}$$\end{document} is the predicted value of ith sample, k is the number of test set designs, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MBT{R^{\left( i \right)~ - 1}}$$\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}$$SAS{A^{\left( i \right)~ - 1}}$$\end{document} are the true values of MBTR and SASA in the last timestep of ith design, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat {{MBTR}}^{\left( i \right)~ - 1}}$$\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}$${\widehat {{SASA}}^{\left( i \right)~ - 1}}$$\end{document} are the predicted values of MBTR and SASA in the last timestep of ith design. Indeed, calculating ErrorSASA necessitates the prior computation of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\widehat {{MBTR}}^{\left( i \right)~ - 1}}$$\end{document}. Consequently, with reference to Eqs. (3) and (10), ErrorSASA can be reformulated as presented in Eq. (11).11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Erro{r_{SASA}}=\frac{1}{k}\mathop \sum \limits_{{i=1}}^{k} \left| {SAS{A^{\left( i \right) - 1}} - {\mathcal{F}_{SASA}}\left( {{{\widehat {{MBTR}}}^{\left( i \right) - 1}};\theta } \right)} \right|.$$\end{document} Proposed method The proposed method consists of three main phases, which are described below. Training the MBTR model During this phase, the MBTR model is trained to address the initial sub-problem, which involves predicting the subsequent time step of MBTR using the preceding time steps. In this particular undertaking, a total of 72 LR models are individually trained, with each model dedicated to a specific feature of MBTR. This methodology closely mirrors the approach outlined in the work by Kibria et al.18; however, instead of an extreme gradient boosting machine (XGBM), LR is employed. The illustration of its framework is presented in Fig. 3. While the baseline research did not employ shuffling of the training set, this research incorporates shuffling to enhance the generalization ability of the model. Fig. 3The architecture of the proposed MBTR model. The architecture of the proposed MBTR model. Data augmentation The SASA model should predict the SASA value for each MBTR time step. However, training the SASA model solely on the original MBTR values as features will result in inaccurate predictions, as the training set does not adequately represent real-world data, which is the prediction of the MBTR model. Consequently, data augmentation is necessary to mitigate this issue. To address this challenge, the MBTR models predictions for all designs are incorporated into the training set of the SASA model. The pseudo-code for this phase is outlined in Algorithm 1. It is important to note that this algorithm should be executed independently for each design, and the newly generated data from all NP designs should be subsequently merged. Training the SASA model In the final phase, the augmented data is merged with the original data, and the combined dataset is shuffled. Subsequently, the SASA model, which is an ETR, is trained on the obtained data. Finally, after training the MBTR and SASA models, the final prediction can be made based on Fig. 2. Results and discussion In this study, a diverse array of 119 NP designs was employed, to provide insights that are not limited to a specific type of cancer but are applicable across various cancer types, thereby enhancing the generalizability of the model. To address the limitations of MD simulations, which are constrained by time and length scales as well as computational resource demands, advanced computational methods were integrated with pre-screened NP designs. This strategy not only improves the stability and predictive efficiency of NPs for cancer treatment but also contributes to practical benefits in the drug development process by reducing both timelines and costs. The broad representation of NP designs in the study further supports the relevance of the findings across different types of cancer, ensuring that the insights gained are widely applicable. Implementation environment All implementations were conducted within the Google Colab environment. Python version 3.10.12 was utilized for development. Except for XGBM, all ML algorithms employed for training were sourced from the scikit-learn library, version 1.2.2. For XGBM, the xgboost Python library version 2.0.3 was employed. In total, nine ML algorithms were examined: LR, XGBM40, adaptive boosting (AdaBoost)41, gradient boosting machine (GBM)42, ETR, support vector regressor (SVR)43, k-neighbors regressor (KNR)44, decision tree regressor (DTR)45, and RFR39. Training and evaluating the robustness and effectiveness of multiple ML algorithms is a common practice. For example, Wang et al.31examined five classification algorithms and six regression algorithms in their proposed framework. Notably, in this study, DNNs were excluded due to the challenges posed by data scarcity, which can hinder accurate predictions18,31. Training the MBTR model The initial phase of the proposed method involves training a model capable of predicting subsequent MBTR values. To achieve this, various ML algorithms were employed across different window sizes. As indicated in Table 1, LR exhibited the highest performance, outperforming other methods across all window sizes. While its MAE was comparable to certain algorithms, such as ETR, in specific window sizes (e.g., 20 and 40), its RMSE was significantly lower. The RMSE metric penalizes large errors, demonstrating LRs superior ability to predict without substantial noise. The observed superiority of LR over other experimented methods can be attributed to two potential factors: Linear Relationship between Subsequent MBTR Values: This possibility suggests that LR performs well because the relationship between consecutive MBTR values is inherently linear, aligning with LRs inherent assumption. However, this explanation is unlikely, as the observed error rate is non-zero, indicating a more complex relationship between MBTR values.Reduced Overfitting due to Simplicity: This possibility posits that LRs inherent simplicity mitigates overfitting, particularly in scenarios where data is scarce. This explanation is supported by the fact that other experimented ML methods exhibited signs of overfitting, failing to generalize well to unseen data. Linear Relationship between Subsequent MBTR Values: This possibility suggests that LR performs well because the relationship between consecutive MBTR values is inherently linear, aligning with LRs inherent assumption. However, this explanation is unlikely, as the observed error rate is non-zero, indicating a more complex relationship between MBTR values. Linear Relationship between Subsequent MBTR Values: This possibility suggests that LR performs well because the relationship between consecutive MBTR values is inherently linear, aligning with LRs inherent assumption. However, this explanation is unlikely, as the observed error rate is non-zero, indicating a more complex relationship between MBTR values. Reduced Overfitting due to Simplicity: This possibility posits that LRs inherent simplicity mitigates overfitting, particularly in scenarios where data is scarce. This explanation is supported by the fact that other experimented ML methods exhibited signs of overfitting, failing to generalize well to unseen data. Reduced Overfitting due to Simplicity: This possibility posits that LRs inherent simplicity mitigates overfitting, particularly in scenarios where data is scarce. This explanation is supported by the fact that other experimented ML methods exhibited signs of overfitting, failing to generalize well to unseen data. Therefore, the superiority of LR can primarily be attributed to its ability to effectively handle data scarcity by reducing overfitting. This contrasts with the other ML methods, which suffered from overfitting due to their greater complexity and inability to generalize well. It is noteworthy that while increasing window sizes generally improved results for models like LR and ETR, some models, such as RFR, performed best with a window size in between, in this case10. This can again be attributed to data scarcity, leading to overfitting. Similar behavior was observed with XGBM, which achieved optimal performance with window sizes of 10 and 20. Conversely, models like SVR exhibited decreased accuracy with larger window sizes. Table 1Test sets MBTR errors for different models and window sizes.ModelsWindow sizes12510203040LR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1.4455 1.8539 1.3104 1.6791 1.2393 1.5923 1.2199 1.5684 1.2090 1.5535 1.1997 1.5424 1.1953 1.5374 XGBM MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1.82892.86681.65072.55391.68032.75101.57552.43781.57242.40871.69172.73401.70622.8399AdaBoost MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 7.899410.64868.283610.86178.973411.73987.636010.25788.916411.28308.417110.85039.188611.6853GBM MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1.87212.80581.85722.78801.81272.71981.84722.77131.85882.82521.87392.83641.86342.8529ETR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1.71242.33841.41521.94011.24891.7394 1.2094 1.6797 1.1968 1.66461.20061.65741.20251.6508SVR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 5.381311.57635.317111.59445.341011.68495.459311.88335.697312.29785.939912.74206.182213.2056KNR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1.73132.61401.62022.47981.57782.46381.61392.54461.69752.69351.77322.81071.83432.9259DTR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 2.05313.12841.94923.00741.89642.95751.90062.97931.94773.10041.95963.11392.04063.2891RFR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1.83412.78041.56892.43591.44562.28071.41922.26081.43972.33961.45072.36471.48102.4448The bold values represent the lowest error of each window size and metric. Table 1Test sets MBTR errors for different models and window sizes.ModelsWindow sizes12510203040LR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1.4455 1.8539 1.3104 1.6791 1.2393 1.5923 1.2199 1.5684 1.2090 1.5535 1.1997 1.5424 1.1953 1.5374 XGBM MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1.82892.86681.65072.55391.68032.75101.57552.43781.57242.40871.69172.73401.70622.8399AdaBoost MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 7.899410.64868.283610.86178.973411.73987.636010.25788.916411.28308.417110.85039.188611.6853GBM MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1.87212.80581.85722.78801.81272.71981.84722.77131.85882.82521.87392.83641.86342.8529ETR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1.71242.33841.41521.94011.24891.7394 1.2094 1.6797 1.1968 1.66461.20061.65741.20251.6508SVR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 5.381311.57635.317111.59445.341011.68495.459311.88335.697312.29785.939912.74206.182213.2056KNR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1.73132.61401.62022.47981.57782.46381.61392.54461.69752.69351.77322.81071.83432.9259DTR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 2.05313.12841.94923.00741.89642.95751.90062.97931.94773.10041.95963.11392.04063.2891RFR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1.83412.78041.56892.43591.44562.28071.41922.26081.43972.33961.45072.36471.48102.4448The bold values represent the lowest error of each window size and metric. Test sets MBTR errors for different models and window sizes. The bold values represent the lowest error of each window size and metric. Training the SASA model without data augmentation The accessible surface area (ASA) or SASA represents the total area of a biomolecule that can be reached by a solvent. As indicated in Table 2, various ML models demonstrate impressive performance in predicting the SASA value for a particular MBTR time step. However, upon examining Table 3, it becomes evident that the ErrorSASA values for all models are considerably higher. For instance, LR exhibits its best performance in predicting the SASA value with a window size of 40, achieving an MAE of 895.87. However, its corresponding ErrorSASA value is 2659.49, indicating a threefold increase. This discrepancy can be attributed to the inherent difficulty in perfectly predicting the last MBTR time step for each design, inevitably introducing noise into the MBTR models predictions. Models that are sensitive to this noise experience a significant increase in ErrorSASA compared to their MAE and RMSE values. This phenomenon affects other ML algorithms as well, leading to a noticeable decline in their performance when predicting the last SASA value. Among these algorithms, ETR with a window size of 40 and RFR with a window size of 40 exhibit the best performance. It is noteworthy that larger window sizes are generally required for accurate SASA value prediction using this approach, as smaller window sizes yield significantly lower accuracy. Reducing the window size is crucial in this context, as it can substantially decrease simulation time. Consequently, the training process must be optimized to achieve this goal, which is addressed in the subsequent section through data augmentation. Table 2Test sets SASA errors for different models and window sizes without data augmentation.ModelsWindow sizes12510203040LR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 965.57 1206.86 961.59 1200.93 952.66 1189.75 941.53 1174.37 920.05 1147.57 909.81 1138.85 895.87 1121.10 XGBM MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1663.872524.271845.802907.471873.182852.951904.182809.381646.342344.081519.872117.511889.672677.43AdaBoost MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 4560.605656.234518.765496.604653.595586.744587.345607.264146.945100.014151.815049.044161.105044.81GBM MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 2535.613339.622526.693315.202483.683272.812363.973073.332294.132949.012195.522831.192114.732739.65ETR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1313.151986.921225.811881.021196.711824.491137.031655.181121.071641.891054.081550.201011.061465.72SVR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 6652.128093.426641.668065.116612.907988.866567.877878.686495.877729.306434.267620.796376.387535.49KNR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 2052.803156.162008.793049.261892.382883.151763.442674.241650.612463.671531.722290.811484.362219.91DTR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 3482.226964.653137.045748.643339.696399.972976.875358.601751.522776.671688.742581.331712.102607.60RFR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 2789.225664.692653.685379.792668.705445.092569.774992.521422.352155.881352.902098.661252.731869.75The bold values represent the lowest error of each window size and metric. Table 2Test sets SASA errors for different models and window sizes without data augmentation.ModelsWindow sizes12510203040LR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 965.57 1206.86 961.59 1200.93 952.66 1189.75 941.53 1174.37 920.05 1147.57 909.81 1138.85 895.87 1121.10 XGBM MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1663.872524.271845.802907.471873.182852.951904.182809.381646.342344.081519.872117.511889.672677.43AdaBoost MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 4560.605656.234518.765496.604653.595586.744587.345607.264146.945100.014151.815049.044161.105044.81GBM MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 2535.613339.622526.693315.202483.683272.812363.973073.332294.132949.012195.522831.192114.732739.65ETR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1313.151986.921225.811881.021196.711824.491137.031655.181121.071641.891054.081550.201011.061465.72SVR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 6652.128093.426641.668065.116612.907988.866567.877878.686495.877729.306434.267620.796376.387535.49KNR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 2052.803156.162008.793049.261892.382883.151763.442674.241650.612463.671531.722290.811484.362219.91DTR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 3482.226964.653137.045748.643339.696399.972976.875358.601751.522776.671688.742581.331712.102607.60RFR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 2789.225664.692653.685379.792668.705445.092569.774992.521422.352155.881352.902098.661252.731869.75The bold values represent the lowest error of each window size and metric. Test sets SASA errors for different models and window sizes without data augmentation. The bold values represent the lowest error of each window size and metric. Table 3ErrorSASA of different algorithms and window sizes without data augmentation.ModelsWindow sizes12510203040LR5886.025246.844663.974607.13842.93281.752659.49XGBM4109.93 2552.11 2462.042879.552948.91 2291.77 2557.45AdaBoost5231.824987.825052.045059.684755.234340.074450.13GBM3686.752721.54 2454.91 2403.22 2911.032615.492685.33ETR 3285.28 3201.972936.023041.922733.562373.53 1985.04 SVR5427.615424.915417.295407.35408.085409.255410.63KNR3943.733696.173379.813485.183070.542520.332345.58DTR5445.94802.234656.564677.262550.882422.122175.32RFR5100.793609.623382.113781 2506.79 2381.792055.64The bold values represent the lowest error of each window size. Table 3ErrorSASA of different algorithms and window sizes without data augmentation.ModelsWindow sizes12510203040LR5886.025246.844663.974607.13842.93281.752659.49XGBM4109.93 2552.11 2462.042879.552948.91 2291.77 2557.45AdaBoost5231.824987.825052.045059.684755.234340.074450.13GBM3686.752721.54 2454.91 2403.22 2911.032615.492685.33ETR 3285.28 3201.972936.023041.922733.562373.53 1985.04 SVR5427.615424.915417.295407.35408.085409.255410.63KNR3943.733696.173379.813485.183070.542520.332345.58DTR5445.94802.234656.564677.262550.882422.122175.32RFR5100.793609.623382.113781 2506.79 2381.792055.64The bold values represent the lowest error of each window size. ErrorSASA of different algorithms and window sizes without data augmentation. The bold values represent the lowest error of each window size. Training the SASA model with data augmentation Training SASA models solely on original MBTR values is not optimal, as the final predictions of MBTR models deviate significantly from the original MBTR values. To address this challenge, a data augmentation technique is proposed and outlined in Algorithm 1. This technique involves utilizing not only the original MBTR values and their corresponding SASA values as training data, but also the predicted MBTR values from the MBTR model. By comparing Tables 2 and 4, it can be observed that the MAE and RMSE of SASA models trained with data augmentation have marginally decreased. However, upon comparing Tables 3 and 5, it becomes evident that the ErrorSASA values of algorithms trained with data augmentation have improved considerably in comparison to those trained without data augmentation. This improvement can be attributed to the fact that SASA models trained with data augmentation are exposed to a more realistic dataset. As a result, they become more proficient in predicting the last SASA value for each NP design. Table 4Test sets SASA errors for different models and window sizes with data augmentation.ModelsWindow sizes12510203040LR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1493.211991.871427.161953.621351.031873.9212931762.141170.1 1550.37 1063.79 1395.24 1041.43 1387.48 XGBM MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1719.432629.751957.32967.281659.932443.941764.62402.11503.412093.231838.662662.191585.582183.41AdaBoost MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 4568.065643.244780.915781.694707.945754.384266.75312.614260.715213.154025.844939.183716.54655.68GBM MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 2566.083399.172607.433438.872509.473340.292543.673292.062464.453175.882355.343016.2322822905.21ETR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1210.51 1828.58 1145.77 1742.99 1124.04 1666.21 1141.62 1701.04 1091.54 1622.981092.091611.02 991.19 1444.22SVR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 6538.847983.936528.327955.946500.037881.946455.387771.96386.147631.026328.317527.536274.067447.89KNR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 2252.863501.712158.663355.692041.293114.461982.282992.281821.772646.21764.812529.71650.612389.79DTR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 2476.223648.312737.94504.223301.636221.42146.183111.361961.772852.581750.862574.451649.522489.37RFR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1354.52036.211991.823264.652448.684544.571393.622019.061406.42003.441477.252079.951507.732300.14The bold values represent the lowest error of each window size and metric. Table 4Test sets SASA errors for different models and window sizes with data augmentation.ModelsWindow sizes12510203040LR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1493.211991.871427.161953.621351.031873.9212931762.141170.1 1550.37 1063.79 1395.24 1041.43 1387.48 XGBM MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1719.432629.751957.32967.281659.932443.941764.62402.11503.412093.231838.662662.191585.582183.41AdaBoost MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 4568.065643.244780.915781.694707.945754.384266.75312.614260.715213.154025.844939.183716.54655.68GBM MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 2566.083399.172607.433438.872509.473340.292543.673292.062464.453175.882355.343016.2322822905.21ETR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1210.51 1828.58 1145.77 1742.99 1124.04 1666.21 1141.62 1701.04 1091.54 1622.981092.091611.02 991.19 1444.22SVR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 6538.847983.936528.327955.946500.037881.946455.387771.96386.147631.026328.317527.536274.067447.89KNR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 2252.863501.712158.663355.692041.293114.461982.282992.281821.772646.21764.812529.71650.612389.79DTR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 2476.223648.312737.94504.223301.636221.42146.183111.361961.772852.581750.862574.451649.522489.37RFR MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE MAE RMSE 1354.52036.211991.823264.652448.684544.571393.622019.061406.42003.441477.252079.951507.732300.14The bold values represent the lowest error of each window size and metric. Test sets SASA errors for different models and window sizes with data augmentation. The bold values represent the lowest error of each window size and metric. Table 5ErrorSASA of different algorithms and window sizes with data augmentation.ModelsWindow sizes12510203040LR2663.092520.012294.072640.192328.262372.992271.38XGBM2521.361894.311945.521850.382198.281782.351654.78AdaBoost5356.855031.884742.54811.45005.114295.473986.78GBM2751.072295.182224.862796.42533.522389.742200.03ETR1448.69 1464.8 1301.75 1569.21 1447.59 1591.1 1415.21 SVR5426.075423.685417.85396.465397.845400.015402.65KNR2408.772682.972878.242808.822383.672465.992657.4DTR2027.712094.522259.182686.423176.33324.331987.95RFR 1431.23 2158.091943.011754.762036.961898.741845.35The bold values represent the lowest error of each window size. Table 5ErrorSASA of different algorithms and window sizes with data augmentation.ModelsWindow sizes12510203040LR2663.092520.012294.072640.192328.262372.992271.38XGBM2521.361894.311945.521850.382198.281782.351654.78AdaBoost5356.855031.884742.54811.45005.114295.473986.78GBM2751.072295.182224.862796.42533.522389.742200.03ETR1448.69 1464.8 1301.75 1569.21 1447.59 1591.1 1415.21 SVR5426.075423.685417.85396.465397.845400.015402.65KNR2408.772682.972878.242808.822383.672465.992657.4DTR2027.712094.522259.182686.423176.33324.331987.95RFR 1431.23 2158.091943.011754.762036.961898.741845.35The bold values represent the lowest error of each window size. ErrorSASA of different algorithms and window sizes with data augmentation. The bold values represent the lowest error of each window size. Similar to the case of training without data augmentation, ETR demonstrates the best performance in this scenario as well. Notably, ETR exhibits high prediction accuracy even with low window sizes, achieving an ErrorSASA of 1448.69 with a window size of one. While RFR exhibits slightly better performance than ETR with a window size of one, ETR is considered a more suitable choice for the SASA model. This is because RFRs low ErrorSASA value with a window size of one may be attributed to chance given the limited amount of data, since it is much lower than its ErrorSASA with other window sizes. In contrast, ETR maintains a consistently high performance across all window sizes, indicating its superior generalization capability for SASA prediction. The superior performance of ETR relative to both simpler models (e.g., LR and DTR) and more complex models (e.g., RFR) suggests that it possesses advantages in both capturing complex relationships and mitigating overfitting. The ETRs superiority over simpler models can be attributed to its greater capacity for feature exploration and its more sophisticated learning process. In contrast, its outperformance of RFR stems primarily from its greater randomness, which includes randomly selecting split points rather than the optimal split point employed by RFR. This randomness helps to reduce overfitting, as the limited nature of the available data makes selecting the absolute optimal split point less reliable, potentially leading to a solution that is suboptimal in the broader context. Furthermore, the ETR demonstrates notable performance compared to the state-of-the-art boosting method, XGBoost. While the architectural differences between these methods make a direct comparison of their underlying reasons for performance difficult, the ETRs superiority to boosting algorithms is clearly evident. Comparison with the baseline The objective of this research was to enhance the efficacy of SASA prediction and reduce overall computational time. Notably, both objectives were successfully achieved. As evidenced in Table 6, the proposed methodology demonstrated superior ErrorSASA performance in comparison to the baseline approach. This improvement in predictive capability exceeded 25%. Additionally, this enhancement was accompanied by a significant reduction in overall execution time, as the proposed method utilized a window size of one, while the baseline employed a window size of 40. Table 6Overall comparison of the baseline and the current research.MBTR MAESASA MAEErrorMBTRErrorSASAXGBM+DNN (window size = 40)1.72941928.44*6.7042*1956.93LR+DA+ETR (window size = 1) 1.4455 1210.51 7.4388 1448.69 The bold values represent the lowest error of each window size.The starred values are not presented in the mentioned study. Instead, they are computed based on the provided models in the study. Table 6Overall comparison of the baseline and the current research.MBTR MAESASA MAEErrorMBTRErrorSASAXGBM+DNN (window size = 40)1.72941928.44*6.7042*1956.93LR+DA+ETR (window size = 1) 1.4455 1210.51 7.4388 1448.69 The bold values represent the lowest error of each window size.The starred values are not presented in the mentioned study. Instead, they are computed based on the provided models in the study. Overall comparison of the baseline and the current research. The bold values represent the lowest error of each window size. The starred values are not presented in the mentioned study. Instead, they are computed based on the provided models in the study. It is noteworthy that full NP simulations require several days of computational effort utilizing high-performance computing resources, while the execution time of ML models for this task is typically less than ten minutes, even with average-performance computing resources. Therefore, the execution time of ML models can be considered negligible. Accordingly, the baseline achieved a time savings factor of 7.5, due to its use of a window size of 40, corresponding to a 40 ns simulation time. This enabled the prediction of a 300 ns simulation, resulting in a 7.5-fold increase in speed. Given that the proposed method employed a window size of one, it is approximately 300 times faster than the traditional approach and 40 times faster than the baseline. Another significant observation is that while the ErrorMBTR of the proposed method with a window size of one exceeds that of the baseline, the ErrorSASA of the proposed method is lower. This discrepancy may be attributed to the suitability of ETR for this task, given the limited dataset and the potential for overfitting in DNNs. Additionally, the employed ETR exhibits strong noise resistance, as evidenced by its ability to accurately predict SASA for the final time step despite a relatively high ErrorMBTR. This may be due to the proposed methods SASA model being able to learn the errors and limitations of the MBTR model through the use of the data augmentation technique. Additionally, the inherent nature of ETR and its randomness in selecting the employed features have contributed to this high performance. Figure 4a displays the MAE of the LR models using window sizes of 1 and 40, alongside the baselines MBTR model, while Fig. 4b portrays the MAE of these models for the top ten highest errors. Notably, the baseline model employed a window size of 40. The results clearly demonstrate that LR, utilizing the same window size as the baseline, outperformed the baselines model in the prediction of each feature. Intriguingly, LR with a window size of one, which is 40 times shorter than the baselines window size, still exhibited commendable performance. This empirical evidence suggests that LR proves more suitable than the baseline model (XGBM) for MBTR prediction within this specific scenario. Fig. 4MAE comparison of LR with window sizes of 1 and 40 and the baseline for (a) all 72 features of MBTR and (b) the top 10 features with the highest errors for LR with a window size of 40. MAE comparison of LR with window sizes of 1 and 40 and the baseline for (a) all 72 features of MBTR and (b) the top 10 features with the highest errors for LR with a window size of 40. Furthermore, Table 7; Fig. 5 illustrate the predictions of the baseline and proposed methods for each design. The results indicate that the proposed method outperforms the baseline for the majority of designs. Notably, the highest percentage error of the proposed method (5.06 for PAN31_3) is lower than the highest percentage error of the baseline (5.46 for OQL13v2_3). This indicates greater stability and generalization compared to the baseline, rendering the proposed method a more suitable model for practical applications. Table 7Comparison of the baseline and the current research for each test design.Design nameActual SASAXGBM+DNNLR+DA + ETR (window size = 1)Predicted SASAPrediction percentage errorPredicted SASAPrediction percentage errorGEM1174,784.6878,390.124.82 74,596.45 0.25 GEM4180,337.0783,150.553.50 81,439.73 1.37 NCL1163,675.27 62,458.62 1.91 66,554.324.52NHQ5160,745.8559,342.532.31 61,125.06 0.62 OQL11_366,567.9567,595.281.54 67,332.05 1.14 OQL13v2_375,432.5171,308.365.46 74,074.29 1.80 PAN11v2_3b74,523.9377,509.724.00 73,945.33 0.77 PAN14v2_376,080.1475,104.631.28 75,171.69 1.19 PAN31_382,504.89 83,197.89 0.83 78,324.765.06S1_11R2_368,690.32 67,374.12 1.91 66,968.942.50S1_11R4_364,980.5167,374.273.68 67,220.58 3.44 S1_15_375,070.25 74,141.69 1.23 73,986.011.44The bold values represent the lowest error of each window size. Table 7Comparison of the baseline and the current research for each test design.Design nameActual SASAXGBM+DNNLR+DA + ETR (window size = 1)Predicted SASAPrediction percentage errorPredicted SASAPrediction percentage errorGEM1174,784.6878,390.124.82 74,596.45 0.25 GEM4180,337.0783,150.553.50 81,439.73 1.37 NCL1163,675.27 62,458.62 1.91 66,554.324.52NHQ5160,745.8559,342.532.31 61,125.06 0.62 OQL11_366,567.9567,595.281.54 67,332.05 1.14 OQL13v2_375,432.5171,308.365.46 74,074.29 1.80 PAN11v2_3b74,523.9377,509.724.00 73,945.33 0.77 PAN14v2_376,080.1475,104.631.28 75,171.69 1.19 PAN31_382,504.89 83,197.89 0.83 78,324.765.06S1_11R2_368,690.32 67,374.12 1.91 66,968.942.50S1_11R4_364,980.5167,374.273.68 67,220.58 3.44 S1_15_375,070.25 74,141.69 1.23 73,986.011.44The bold values represent the lowest error of each window size. Comparison of the baseline and the current research for each test design. The bold values represent the lowest error of each window size. Fig. 5Prediction comparison of (a) the proposed method and (b) the baseline18on the test sets designs. Prediction comparison of (a) the proposed method and (b) the baseline18on the test sets designs. Practical usages Computational advancements of the proposed method significantly accelerate the evaluation of NP designs, bridging the gap between in silico simulations and in vitro/in vivo experiments. Specifically, this approach reduces NP simulation time from days to mere minutes, generating substantial time and cost savings during the early stages of drug development. This method is particularly impactful during preclinical stages, where extensive screening of potential drug candidates is crucial. By pre-screening and eliminating ineffective designs early, the selection process is streamlined, ensuring that only the most promising candidates proceed to experimental validation. This optimizes resource allocation and enhances the overall efficiency of drug development, ultimately accelerating the delivery of novel treatments to the clinic46,48. The proposed SASA prediction method offers a significant advancement in the design of cancer nanomedicine. By rapidly predicting NP surface interactions with tumor cells, this method provides a substantial acceleration compared to traditional experimental approaches. Through accurate and expeditious prediction of SASAs, the method facilitates the customization of NPs to enhance their binding affinity and specificity to cancerous tissues. This precision eliminates the need to explore ineffective NP designs, streamlining the development process. Consequently, the development of more effective targeted therapies is facilitated, potentially increasing therapeutic efficacy while minimizing off-target effects. Furthermore, the enhanced accuracy in NP design can lead to more efficient drug delivery systems, ultimately contributing to improved treatment outcomes and reduced side effects in cancer therapy46,49. Limitations and future works The primary limitation of this study was the scarcity of available data. This constraint led to overfitting in many ML methods. To address this issue, future work could explore the use of a wider range of NP designs for training. Additionally, employing generative AI methods, such as generative adversarial networks (GANs) and variational autoencoders (VAEs), could facilitate the creation of synthetic training data. Furthermore, enhancing the interpretability and explainability of the proposed model through the integration of explainable AI (XAI) techniques would contribute to a more reliable system. The models performance could also be rigorously evaluated on novel, unseen NP designs. Another limitation lies in the requirement for model retraining with new data. This could be addressed by implementing reinforcement learning to develop self-updating models, enhancing practicality. Future research could also focus on developing more effective data augmentation techniques to further improve model performance. Conclusion This study presented a novel methodology for predicting the SASA value of the final timestep of various NP designs using an initial MBTR value. The proposed methodology comprised three distinct phases: MBTR model training (an LR model), data augmentation, and SASA model training (an ETR model). In the first phase, comparative analysis revealed that the LR model outperformed other ML methods evaluated in predicting MBTR values. This suggested that the relationship between MBTR values at different timesteps was predominantly linear, or that the LR model exhibited superior performance due to data scarcity, as more intricate models are prone to overfitting when data is limited. In the second and third phases, the employed data augmentation technique significantly enhanced SASA prediction accuracy by incorporating predictions from the MBTR models for various designs into the SASA model training dataset alongside the original data. Comparative analysis against a baseline method revealed that the proposed methodology achieved a significant accuracy improvement of over 25%. Moreover, the proposed method exhibited a 40-fold speed advantage over the baseline and a 300-fold speed advantage over the conventional MD simulation approach. The proposed methodology provided a highly accurate and efficient approach for predicting SASA values, offering significant advantages over the existing method. This methodology has potential applications in various fields, including drug design and materials science.