Zhoumeng Lin,1– 4,* Wei-Chun Chou,1– 4,* Yi-Hsien Cheng,3,4 Chunla He,5 Nancy A Monteiro-Riviere,6,7 Jim E Riviere7,8 1Department of Environmental and Global Health, College of Public Health and Health Professions, University of Florida, Gainesville, FL, USA; 2Center for Environmental and Human Toxicology, University of Florida, Gainesville, FL, USA; 3Institute of Computational Comparative Medicine, Kansas State University, Manhattan, KS, USA; 4Department of Anatomy and Physiology, College of Veterinary Medicine, Kansas State University, Manhattan, KS, USA; 5Department of Biostatistics, College of Public Health and Health Professions, University of Florida, Gainesville, FL, USA; 6Nanoknow-how Innovation Center of Kansas State, Kansas State University, Manhattan, KS, USA; 7Center for Chemical Toxicology Research and Pharmacokinetics, North Carolina State University, Raleigh, NC, USA; 8 1Data Consortium, Kansas State University, Olathe, KS, USACorrespondence: Zhoumeng Lin, Department of Environmental and Global Health, College of Public Health and Health Professions, University of Florida, 1225 Center Dr., Gainesville, FL, 32610, USA, Tel +1 352-273-6160, Email [email protected]Background: Low supply effectivity of nanoparticles (NPs) to the tumor is a vital barrier within the subject of most cancers nanomedicine. Strategies on how to enhance NP tumor supply effectivity stay to be decided.Methods: This examine analyzed the roles of NP physicochemical properties, tumor fashions, and most cancers sorts in NP tumor supply effectivity utilizing a number of machine studying and synthetic intelligence strategies, utilizing knowledge from a lately printed Nano-Tumor Database that accommodates 376 datasets generated from a physiologically based mostly pharmacokinetic (PBPK) mannequin.Results: The deep neural community mannequin adequately predicted the supply effectivity of various NPs to totally different tumors and it outperformed all different machine studying strategies; together with random forest, help vector machine, linear regression, and bagged mannequin strategies. The adjusted willpower coefficients (R2) within the full coaching dataset had been 0.92, 0.77, 0.77 and 0.76 for the utmost supply effectivity (DEmax), supply effectivity at 24 h (DE24), at 168 h (DE168), and on the final sampling time (DETlast). The corresponding R2 values within the take a look at dataset had been 0.70, 0.46, 0.33 and 0.63, respectively. Also, this examine confirmed that most cancers sort was an essential determinant for the deep neural community mannequin in predicting the tumor supply effectivity throughout all endpoints (19– 29%). Among all physicochemical properties, the Zeta potential and core materials performed a higher position than different properties, corresponding to the kind, form, and focusing on technique.Conclusion: This examine offers a quantitative mannequin to enhance the design of most cancers nanomedicine with higher tumor supply effectivity. These outcomes assist to enhance our understanding of the causes of low NP tumor supply effectivity. This examine demonstrates the feasibility of integrating synthetic intelligence with PBPK modeling approaches to examine most cancers nanomedicine.Graphical Abstract: Keywords: synthetic intelligence, machine studying, physiologically based mostly pharmacokinetic modeling, nanomedicine, drug supply, nanotechnology
Introduction
Global Cancer Statistics estimated that just about 10 million deaths occurred worldwide due to most cancers in 2020.1 The therapy of most cancers stays a problem. Traditional small molecule-based chemotherapy has limitations; corresponding to antagonistic uncomfortable side effects, low therapeutic indices, low bioavailability, high-dose necessities, lack of specificity, and growth of multi-drug resistance.2 Recent advances in nanotechnology enabled nanoparticle (NP)-based drug formulations to be designed with superior properties in contrast to conventional small molecule chemotherapy, together with excessive drug loading, particular focusing on/supply, and the power to regulate the discharge of the anticancer drug in a managed or sustained method.3 During the final 30 years, many nanomedicines had been designed and examined to be efficient in lowering the dimensions of assorted tumor sorts in laboratory animals,4,5 however solely a restricted variety of nanomedicines have been translated to scientific success and accredited by the United States Food and Drug Administration (US FDA) or European Medicines Agency (EMA).6 Low scientific translation of animal outcomes to people has develop into a significant impediment within the subject of most cancers nanomedicine.
The low animal-to-human translation of most cancers nanomedicines have been, partly, attributed to two fundamental causes: (1) an absence of physiologically based mostly fashions that may extrapolate pharmacokinetic and biodistribution outcomes of NPs from animals to people,7–9 and (2) the low supply effectivity noticed for NPs to the tumor web site (~0.7% of median injected dose).4,5,10 To handle the previous situation, physiologically based mostly pharmacokinetic (PBPK) fashions had been developed to simulate the biodistribution of various NPs in wholesome rodents and people.11–20 Recently, considered one of these fashions developed by our group was extrapolated to tumor-bearing mice. This mannequin was calibrated with a whole lot of datasets obtained from various kinds of NPs in several tumor sorts.5 This mannequin was used to predict the supply efficiencies of various NPs at totally different occasions (eg, 24 and 168 h) after intravenous (IV) injection in tumor-bearing mice (this dataset is referred to as “Nano-Tumor Database” on this manuscript). To handle the problem of low tumor supply effectivity, it’s important to decide the connection between the physicochemical properties and the tumor supply effectivity of the NPs. Previous research tried to use easy a number of linear regression evaluation to decide the roles of the physicochemical properties in NP tumor supply effectivity.4,5 Although some statistically vital correlations had been recognized, the extent of those correlations had been low (eg, the willpower coefficients R2 values had been round 0.3 to 0.5). Therefore, it can be crucial to handle this vital limitation for the reason that position of the physicochemical properties in NP tumor supply effectivity are important for the right design to enhance the supply effectivity of most cancers nanomedicines.
In current years, due to the fast progress in computational energy, the provision of a considerable amount of knowledge in varied databases, and the event of refined knowledge evaluation algorithms, a number of machine studying (ML), and synthetic intelligence (AI) strategies can be found to assist predict the absorption, distribution, metabolism, and excretion (ADME) properties, in addition to the toxicity of chemical substances or NPs.21–24 A abstract of the essential traits of generally used strategies (additionally the strategies used on this examine) is supplied in Table 1. Among these strategies, a number of research counsel that synthetic neural networks are extra strong than linear regression as a result of they’ll course of giant datasets extra effectively, work with incomplete knowledge, deal with each linear and nonlinear processes and establish new relationships not inputted by the consumer.25–27 A current report confirmed {that a} deep neural community (DNN) mannequin (deep studying [DL] mannequin) might be used to predict the biodistribution of NPs based mostly on their floor chemistry in wholesome rats.28 However, using ML or AI strategies has not been utilized to predict the biodistribution of NPs in tumor-bearing animals or in most cancers sufferers.
Table 1 Summary of Modeling Algorithms Used in This Study
The goal of this examine was to decide the roles of the physicochemical properties, tumor fashions, and most cancers sorts within the tumor supply effectivity of NPs. We hypothesized that extra refined ML and DNN computational strategies outperform linear regression within the willpower of the connection between the physicochemical properties and the tumor supply effectivity of NPs, with the DNN mannequin having the very best efficiency. The current evaluation was based mostly on our printed Nano-Tumor Database that accommodates 376 datasets from various kinds of NPs in tumor-bearing mice.5 This examine extends our earlier work5 by offering a greater quantitative mannequin based mostly on the DNN approaches, which might assist within the design of latest most cancers nanomedicines with a better tumor supply effectivity. This examine additionally offers invaluable insights into the relative contributions of various physicochemical properties of NPs on tumor supply effectivity. Another novelty of this examine is that it demonstrates PBPK modeling will be built-in with ML and AI strategies, and the PBPK-based supervised DNN technique can be utilized to predict the supply effectivity of NPs to totally different tumors in animals and serves as a foundation for making use of this method to most cancers sufferers.
Materials and Methods
Datasets and Data Preprocessing
In our earlier examine,5 we estimated supply efficiencies (DE) of NPs to the tumor at 24 h (DE24), 168 h (DE168) and final sampling time factors (DETlast), in addition to the utmost DE (DEmax) after IV administration in tumor-bearing mice utilizing PBPK fashions based mostly on the Nano-Tumor Database consisting of 376 datasets protecting a variety of most cancers nanomedicines printed from 2005 to 2018. Multiple variables that may affect the tumor supply had been included on this database, together with the physicochemical properties of NPs [eg, log-transformed hydrodynamic diameter (Size), original value of Zeta potential (ZP), shape, core material (MAT), type of NPs (Type)], tumor remedy methods such because the focusing on methods (TS), most cancers sorts (CT) and tumor mannequin (TM), and many others. The hydrodynamic diameter was log-transformed as a result of its values had a variety and weren’t usually distributed. By leveraging the sooner work, the database was reorganized by filtering the lacking knowledge for the event of ML and DL fashions. The symbols and additional clarification for all of the variables within the Nano-Tumor Database are supplied in Table 2.
Table 2 List of Variables and Its Levels/Values within the Tumor-Nano Database
In the info preprocessing step, each categorical and numeric knowledge had been included within the database, so two totally different knowledge preprocessing strategies had been used together with one-hot encoding and have scaling had been utilized to the info to permit it to be acknowledged by the ML and DL fashions. For the specific variables, the info had been one-hot encoded by splitting variables into totally different columns with its personal encoded binary string to impose a synthetic ordering on the variables which will have implications for the ML and DL fashions.29 For numerical variables, the function scaling method was normalized to the worth to heart across the imply with a unit customary deviation to assist the mannequin optimization effectivity.30 The one-hot encoding and have scaling had been performed by the perform dummyVars and preProcess, respectively from R package deal caret.31
Feature Selection
To enhance the mannequin studying efficiency, the function choice is important earlier than the mannequin growth. In this examine, the low-variation function filtering algorithm and stepwise regression choice had been used to choose the enter options. The low-variation function filtering algorithm was used to filter out the an identical or virtually an identical options within the knowledge set by means of the set-up threshold values. This technique supplied process to eradicate the irrelevant and redundant variables with low variance and to have much less influence on the response variable. The low-variation function filtering algorithm was executed by the nearZeroVar perform in R package deal caret (model 6.0–86),31 adopted by the stepwise regression method by way of the perform step in R package deal stats. Stepwise regression is a statistical method to construct a mannequin by including or eradicating variables based mostly on a sequence of the take a look at statistics of the estimated coefficients, which was used to establish the smallest set of options that had a major influence to the response variable within the regression mannequin.
Model Development
A complete of 9 modeling algorithms had been utilized on this examine. These algorithms will be categorized into 4 courses together with traditional fashions, ensemble fashions, help vector machines (SVMs) and neural networks (Table 1). Two traditional fashions together with the straightforward linear regression (LR) and k-nearest neighbors (KNN) had been used as the straightforward ML algorithms. Three determination tree algorithms, together with Random Forest (RF), Bagged mannequin (Bag), and Gradient boosting mannequin (Gbm), had been used as ensemble fashions. For SVMs, three variations of the SVM fashions based mostly on the linear foundation kernel had been adopted, together with common SVM (R-SVM), least-squared (LS-SVM) and L2 Regularized (L2-SVM) fashions. These algorithms had been carried out in R software program (Version 4.02) with ML packages corresponding to kernlab (model 0.9–25),32 the randomForest (model 4.6–12),33 and the xgboost (model 0.4–4)34 for mannequin development. The random search method35 carried out within the R package deal caret was utilized to optimize the hyperparameters for every ML mannequin.
For the synthetic neural networks, a DL neural community was constructed to predict the supply efficiencies of most cancers nanomedicines. DL was performed by means of the R package deal “h2o” (Version 3.32.0.5), which is a R interface for a multilayer feedforward neural community mannequin.36 Five dense layers (three hidden layers) had been included within the structure of the DL mannequin. Three hidden layers comprised [512, 256, 128], [480, 240, 120], [512, 128, 64] and [180, 90, 45] nodes had been used for the predictions of DEmax, DE24, DE168 and DETlast, respectively. The ReLu was used as an activation perform to carry out non-linear transformations.37 Using the coaching dataset, the training charge and the regulation perform with L1 [ie, Lasso Regression] and L2 [ie, Ridge Regression] was optimized. The Adam and root imply sq. error (RMSE) had been used as optimizers and loss features to compile the DL mannequin on this examine. The dropout function38 and early stopping rule39 had been utilized to scale back overfitting and to enhance the generalization error within the mannequin. In addition, the variable significance for the DL mannequin was calculated based mostly on the tactic from Gedeon.40
Model Performance Evaluation
The analysis of mannequin efficiency was performed utilizing the interior validation and exterior validation strategies generally utilized in ML and DL modeling research.41–45 The unique dataset was randomly cut up right into a coaching set (80% of the info), as inside validation knowledge by means of 5-fold cross-validation for coaching the mannequin, and a take a look at set (20% of information) for exterior validation of the mannequin. In the 5-fold cross-validation evaluation, the coaching set was additional partitioned into 5 equal sized subsets. Of the 5 subsets, 4 subsets had been used for constructing the mannequin and a single subset was retrained as validation knowledge for evaluating the developed mannequin. This cross-validation course of was then repeated 5 occasions till all subsets had been used for validation as soon as. The efficiency of every mannequin for the 5-fold cross-validation and exterior validation was evaluated by root imply sq. error (RMSE), imply absolute error (MAE) and adjusted willpower coefficient (R2). These evaluated metrics are outlined as beneath:
(1)
(2)
(3)
the place y is the noticed response variable worth derived from the Nano-Tumor Database based mostly on a PBPK mannequin,5 is its imply, is the corresponding predicted worth, and n is the variety of knowledge units. R2 was used to consider the goodness-of-fit of the mannequin, and RMSE and MAE had been used to consider the error between noticed and predicted values. The decrease values of MAE and RMSE indicate larger accuracy of the mannequin. However, a better worth of R2 is taken into account fascinating.
Results
Study Workflow
The total examine framework is depicted in Figure 1. All knowledge on the physicochemical properties of NPs, tumor mannequin, and most cancers sort had been obtained from our printed Nano-Tumor Database.5 The mixture of 4 forms of response variables (supply effectivity at totally different time factors after IV injection, together with DE24, DE168, DEmax, and DETlast) together with 9 modeling algorithms, 36 fashions had been developed. Specifically, the physicochemical properties of NPs (Type, Size, ZP, form, MAT), and the parameters associated to the outline of tumor research (TM, TS, and CT) had been used as enter options to predict the tumor supply efficiencies of various NPs derived from the PBPK mannequin from our earlier examine.5 After knowledge preprocessing (ie, function choice and scaling), the fashions had been developed utilizing the totally different modeling algorithms introduced above (Table 1). During mannequin coaching, the optimization of hyperparameters was performed to select a set of optimum hyperparameters for ML and DL mannequin algorithms. Final fashions had been evaluated by the 5-fold cross-validation technique and the standard of every mannequin was evaluated with each coaching and take a look at knowledge units, individually utilizing a number of indicators, together with R2, RMSE, and MAE.
Figure 1 Overview of the examine framework to develop machine studying and deep studying fashions to predict supply effectivity of nanoparticles to the tumor web site in tumor-bearing mice. X represents the preliminary enter variables and W represents the variables after function picks.Abbreviations: R2, adjusted coefficient of willpower; RMSE, root imply sq. error; MAE, imply absolute error.
Nano-Tumor Database
Figure 2 is an outline of the info units on this examine. The tumor supply effectivity knowledge had been categorized based mostly on the most cancers therapeutic situations or the physicochemical properties of the NPs, together with most cancers TS (Figure 2A), CT (Figure 2B), TM (Figure 2C), Type (Figure 2D), form (Figure 2E), MAT (Figure 2F), Size (Figure 2G), and ZP (Figure 2H). For the parameter of focusing on methods, nearly all of datasets had been from research that used passive focusing on (68%). The knowledge had been associated to a variety of various most cancers sorts, together with breast (30%), liver (17%), colon (8%), cervix (7%), and lung (6%) and others (32%). The mouse tumor fashions included Allograft Heterotopic (AH, 38%), Allograft Orthotopic (AO, 12%), Xenograft Heterotopic (XH, 38%), and Xenograft Orthotopic (XO, 12%). For the physicochemical properties of NPs, the bulk had been natural NPs (71%). The core supplies had been additionally very various, together with polymeric (40%), gold (17%), liposomes (9%), hydrogels (6%), silica (6%), iron oxide (2%), dendrimers (2%), and others (18%). The values of ZP, which described the cost on the interface between the NP floor and its liquid medium, ranged from −59.4 mV to 71.30 mV. The “Size” of the studied NPs, which represented the log10-transformed hydrodynamic diameters of the NPs, ranged from 0.43 nm to 2.66 nm (ie, the vary of the unique hydrodynamic diameter values was 2.51 nm to 457 nm).
Figure 2 Overview of the Nano-Tumor Database. (A–C) Percentages of information associated to most cancers remedy situations based mostly on the Targeting Strategy (TS), Cancer Type (CT), and Tumor Model (TM). (D–H) Percentages of information that had been categorized based mostly on the physicochemical properties of the studied nanoparticles, together with the kind (D), the form (E), the core (F), the log-transformed hydrodynamic diameter (G), and Zeta potential (H). In Panel H, Zeta potential values are introduced utilizing field whisker plots throughout 2.fifth, twenty fifth, fiftieth, seventy fifth, and 97.fifth percentiles. Data within the Nano-Tumor Database are from Cheng et al.5
Development and Validation of NP Tumor Delivery Efficiency Models
The R2, RMSE and MAE values based mostly on 5-fold cross-validation and testing outcomes from the developed ML and DL fashions are summarized in Table 3. Among the chosen ML mannequin algorithms, the RF mannequin confirmed a greater predicting efficiency for every of the endpoints with larger R2 and decrease RMSE or MAE values than different classes of ML algorithms. The comparatively weakest modeling algorithm was the KNN mannequin with R2 values for all endpoints beneath 0.1. The R2 and RMSE values for RF ranged from 0.11 to 0.29 and from 3.17 to 7.15 throughout every of endpoints within the take a look at set, whereas the values within the coaching set ranged from 0.15 to 0.19 and from 2.06 to 3.72, respectively. Although the L2-SVM mannequin barely outperformed the others within the take a look at set for DE168, the outcomes is probably not dependable as a result of the values of R2 and RMSE between coaching and take a look at units had vital variations. For the predicting efficiency of the DL mannequin (Table 3; Figure 3), the outcomes outperformed all ML strategies with the very best R2 values and considerably decrease RMSE and MAE values throughout all endpoints in contrast to these from different algorithms. The R2 values had been 0.70, 0.46, 0.33 and 0.63 for DEmax, DE24, DE168 and DETlast within the take a look at set, respectively, whereas these values had been 0.92, 0.77, 0.77 and 0.76 within the full coaching set (Figure 3). In addition, the same ranges of R2, RMSE and MAE between 5-fold cross-validation leads to the coaching and take a look at units within the DL mannequin outcomes counsel that there have been minimal or no overfitting issues.
Table 3 Five-Fold Cross-Validation and Testing Results for Tumor Delivery Efficiency Using Different Machine Learning and Deep Learning Models
Figure 3 Correlation between values from the Nano-Tumor Database and the deep neural community model-predicted values for (A) DEmax, (B) DE24, (C) DE168 and (D) DETlast. Root imply sq. error (RMSE) and coefficient of willpower (R2) within the coaching set and take a look at set are proven. RMSE_train and R2_train characterize the very best efficiency of RMSE and R2 values within the coaching set, whereas RMSE_test and R2_test characterize the values for exterior validation. DEmax, DE24, DE168 and DETlast characterize the utmost tumor supply effectivity (DE), DE at 24 h, 168 h, and the final sampling time, respectively.
To additional assess and to affirm whether or not the performances of the DL mannequin was extra superior than different strategies, the normal easy linear regression mannequin (LM) and RF mannequin had been used to predict supply effectivity (Figures 4–5) after which the outcomes had been in contrast with these from the DL mannequin (Figure 3). These outcomes confirmed that the performances of the DL mannequin within the predictions of DEmax, DE24, DE168 and DETlast had been higher than the LM mannequin based mostly on R2 and RMSE (Figure 4). Although comparable values of R2 and RMSE had been discovered between the DL and RF fashions within the coaching dataset, the RF mannequin didn’t predict as nicely within the take a look at dataset in contrast with the DL mannequin (Figure 5). These outcomes counsel that of all of the developed ML and DL fashions, the DL mannequin had the very best predicting efficiency amongst all endpoints within the coaching and take a look at units. The code of all of the developed ML and DL fashions is supplied within the GitHub (https://github.com/UFPBPK/Nano-ML-AI) to allow different researchers to reproduce our outcomes and to apply our optimum DL mannequin for the design of latest nanomedicines.
Figure 4 Correlation between values from the Nano-Tumor Database and the expected values based mostly on the straightforward linear regression mannequin for (A) DEmax, (B) DE24, (C) DE168 and (D) DETlast. Root imply sq. error (RMSE) and coefficient of willpower (R2) within the coaching set and take a look at set are proven. RMSE_train and R2_train characterize the very best efficiency of RMSE and R2 values within the coaching set, whereas RMSE_test and R2_test characterize the values for exterior validation. DEmax, DE24, DE168 and DETlast characterize the utmost tumor supply effectivity (DE), DE at 24 h, 168 h, and the final sampling time, respectively.
Figure 5 Correlation between values from the Nano-Tumor Database and the expected values based mostly on the random forest mannequin for (A) DEmax, (B) DE24, (C) DE168 and (D) DETlast. Root imply sq. error (RMSE) and coefficient of willpower (R2) within the coaching set and take a look at set are additionally proven. RMSE_train and R2_train characterize the very best efficiency of RMSE and R2 values within the coaching set, whereas RMSE_test and R2_test characterize the values for exterior validation. DEmax, DE24, DE168 and DETlast characterize the utmost tumor supply effectivity (DE), DE at 24 h, 168 h, and the final sampling time, respectively.
Feature Importance
To perceive the influence of the enter options on mannequin predictions, we estimated the variable significance based mostly on the tactic from Gedeon.40 As proven in Figure 6, the most cancers sort had a higher influence on the DL mannequin throughout all endpoints (19% ~ 29%). For the variables associated to the physicochemical properties of NPs, MAT (22%) considerably contributed to the DEmax, whereas ZP and Size had influence on the DE24, DE168 and DETlast. Overall, the significance of options associated to the most cancers remedy methods was larger than the parameters associated to the physicochemical properties of NPs. Among the physicochemical parameters of NPs, the ZP and MAT had been a very powerful components to the ultimate mannequin than different parameters.
Figure 6 Importance share within the deep studying mannequin for every function variable. (A–D) characterize outcomes for the function variable of DEmax, DE24, DE168 and DETlast, respectively. Individual significance is represented by colours within the stacked bars.Abbreviations: ZP, zeta potential; Type, sort of nanoparticles; TS, focusing on technique; TM, tumor mannequin; Size, log-transformed worth of the hydrodynamic measurement; Shape, form of nanoparticles; MAT, core materials of nanoparticles; CT, most cancers sort; AH, allograft heterotopic; AO, allograft orthotopic; XH, xenograft heterotopic; XO, xenograft orthotopic.
Discussion
This examine stories a PBPK-based DL neural community mannequin that may adequately predict the supply effectivity of NPs to tumors in mice based mostly on the NP physicochemical properties, tumor mannequin and most cancers sort. This mannequin can function a predictive device to help within the design of latest NP-based drug formulations for most cancers remedy. This device is anticipated to facilitate nanomedicine growth sooner or later by stopping NPs of low tumor supply effectivity from coming into preclinical trials, thereby serving to scientists to make higher knowledgeable choices and scale back and refine animal research. This examine additionally represents a strategy advance by integrating ML and AI approaches with PBPK modeling within the subject of most cancers nanomedicine.
Low supply effectivity of NPs to tumors has for many years been a barrier within the subject of most cancers nanomedicine.3–5 We acknowledge {that a} nanomedicine’s specificity and efficacy are vital components which will overcome low supply effectivity. However, for NP with comparable pharmacodynamic exercise, enhanced supply will enhance efficacy, lower antagonistic results in non-target tissues, and doubtlessly decrease whole dose wanted for therapy. Previous research on supply effectivity used conventional easy multivariate regression evaluation to decide the connection between NP physicochemical properties and tumor supply effectivity.4,5 While some vital correlations had been recognized, the extent of the correlation was usually low with willpower coefficients ranging solely from 0.3 to 0.5. These simplified fashions had been additionally not rigorously validated. In the current examine, we carried out 9 ML and DL fashions (Table 1) and in contrast their outcomes (Table 3). We discovered that the RF mannequin had higher predictive efficiency than different ML fashions, and the DL mannequin had the very best predictive efficiency in contrast to all different strategies. The DL mannequin was additionally rigorously validated with inside cross-validation and exterior validation. This DL mannequin is a extra superior mannequin than the straightforward multivariate linear regression fashions reported in earlier research,4,5 and it represents the very best mannequin presently obtainable to predict NP tumor supply effectivity on this subject.
Compared to earlier research,4,5 the current examine additionally improves our understanding on the relative contributions of various physicochemical properties to the tumor supply effectivity of NPs. It is well-known that pharmacokinetics and tissue distribution (together with tumor supply), depend upon a number of components, together with the dose, measurement, ZP, floor coating, MAT, form and TS.8,18 Earlier research utilizing conventional multivariate linear regression evaluation and stepwise inclusion/exclusion strategies had been in a position to establish just one or a number of main components within the willpower of tumor supply effectivity for particular NPs. The limitation of this conventional linear regression technique is that the ultimate optimum mannequin is proscribed to a number of main parameters, whereas different comparatively minor parameters have to be excluded. Thus, the linear regression mannequin solely considers the roles of some parameters, however ignores others. As a end result, the relative contributions of every of the components in a linear regression mannequin are unknown.4,5 In this examine, the DL mannequin thought-about all parameters and evaluated the relative contributions of every of the studied components on every of the chosen tumor supply effectivity metrics (Figure 6). These outcomes counsel that most cancers sort was an essential contributor to the DL mannequin in predicting tumor supply effectivity. Among all of the physicochemical properties, the ZP and MAT performed a higher position than different properties, together with the kind, form, and TS. These findings are according to earlier research utilizing easy linear regression fashions.4,5 Different forms of cancers have totally different most cancers cells with totally different tumor microenvironments (eg, totally different stromal cells and extracellular matrix parts), leading to totally different dynamic interactions of most cancers cells with their native microenvironment and totally different blood circulate charges; these components can in flip lead to totally different tumor supply efficiencies.46–48 ZP and MAT have been proven to be key components in tumor supply.3,49,50 Internalization of NPs into the tumor cells is a vital step in tumor supply as the method of binding of NPs to cell membrane is generally affected by ZP.49 Also, various kinds of NPs have vastly totally different physicochemical properties, which might all contribute to totally different tumor supply efficiencies.50
This examine has a number of limitations. While the ultimate DL mannequin was evaluated extensively with inside cross-validation and exterior validation, the coaching and testing datasets had been essentially restricted to the Nano-Tumor Database printed two years in the past.5 The mannequin could be extra rigorous if it had been evaluated with newer datasets that weren’t included on this database on the time of its compilation. We weren’t in a position to broaden it additional as a result of it will have required further animal testing and PBPK simulations, which is out of the scope of this examine; which was performed solely to decide if ML and AI strategies will be built-in with PBPK-derived outcomes to develop a greater mannequin to enhance predictions from this knowledge set. To facilitate mannequin purposes, sooner or later it will likely be helpful to convert the ultimate mannequin to a web-based interface. This is comparable with a number of lately printed interactive physiologically based mostly pharmacokinetic (iPBPK) interfaces.51–53 Our outcomes are solely based mostly on knowledge from tumor-bearing mice. The implications of those outcomes to people stay to be investigated. This will be addressed by extrapolating the PBPK mannequin from tumor-bearing mice to most cancers sufferers, and a PBPK mannequin is a perfect mannequin to carry out this animal-to-human extrapolation.9,13,20 However, further knowledge on pharmacokinetics of NPs in human most cancers sufferers are wanted so as to validate the human mannequin to be used to generate knowledge to practice ML and AI fashions.
This examine considerably advances the sector of most cancers nanomedicine by integrating ML and AI applied sciences with PBPK modeling to examine most cancers nanomedicine. ML and AI fashions are helpful to predict PBPK-related parameters, corresponding to in vitro dissolution charge, hepatic clearance, and membrane permeability of small molecular medication;54,55 and these parameters are in flip useful to help PBPK mannequin growth. While ML and AI strategies have been utilized to help PBPK mannequin growth of small molecular medication,54,55 they haven’t been utilized to develop PBPK fashions of nanomedicines. ML and AI fashions are identified to be comparatively complicated and require a considerable amount of knowledge to practice and consider. There is a problem to combine ML and AI fashions with pharmacokinetic knowledge as a result of pharmacokinetic knowledge are sometimes derived from present animal trials which have totally different experimental designs, that are normally restricted when it comes to time factors, matrix, dosing teams, and many others. In this regard, as soon as a PBPK mannequin is validated, it may be used to predict chemical or NP plasma and tissue concentrations, in addition to pharmacokinetic parameters (eg, most focus [Cmax] and space underneath the focus curve [AUC]) at totally different time factors in several matrices after totally different publicity situations; thereby optimizing obtainable datasets which weren’t designed to have constant designs amenable to such a meta-analysis. This was properly illustrated in our earlier examine the place a PBPK mannequin was used to predict pharmacokinetic knowledge and tumor supply efficiencies of various kinds of NPs, producing the Nano-Tumor Database that accommodates 376 datasets.5 This database made it attainable to practice and consider the ML and DL fashions within the current examine. This examine demonstrates that it’s attainable to use PBPK simulation outcomes to inform ML and AI mannequin growth. Therefore, earlier research and our current examine collectively counsel that ML and AI and PBPK will be nicely built-in to help one another. This AI-PBPK integrative method is new and offers an development to the sector of nanomedicine. This conclusion has large implications contemplating that PBPK modeling has quite a lot of biomedical purposes, together with drug discovery and growth (each small molecular medication and nanodrugs),56 environmental Health danger evaluation,57,58 and animal-derived meals security evaluation.59
Conclusion
Overall, this examine extends our earlier study5 by creating a extra strong quantitative mannequin based mostly on DL and PBPK approaches that can be utilized to predict tumor supply effectivity of various NPs based mostly on the physicochemical properties, most cancers sort and tumor mannequin. The current examine additionally demonstrates the feasibility of integrating ML/AI with PBPK fashions to help most cancers nanomedicine analysis and growth. These findings characterize a methodological development within the subject of most cancers nanomedicine. The remaining DL mannequin can function a platform to assist sooner or later design of latest most cancers nanomedicines and assist scientists to make knowledgeable choices as to which NPs ought to enter preclinical trials, thereby, lowering and refining animal research. This framework will be prolonged to different purposes of PBPK modeling, together with small molecular drug growth, environmental well being danger evaluation, and animal-derived meals security evaluation.
Acknowledgments
The authors would really like to acknowledge funding help from the National Institute of Biomedical Imaging and Bioengineering of US National Institutes of Health (NIH) (Grant numbers: R01EB031022 and R03EB026045).
Disclosure
The authors report no conflicts of curiosity on this work.
References
1. Sung H, Ferlay J, Siegel RL, et al. Global most cancers statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 international locations. CA Cancer J Clin. 2021;71(3):209–249. doi:10.3322/caac.21660
2. Senapati S, Mahanta AK, Kumar S, Maiti P. Controlled drug supply autos for most cancers therapy and their efficiency. Signal Transduct Target Ther. 2018;3(1):7. doi:10.1038/s41392-017-0004-3
3. Rosenblum D, Joshi N, Tao W, Karp JM, Peer D. Progress and challenges in the direction of focused supply of most cancers therapeutics. Nat Commun. 2018;9(1):1410. doi:10.1038/s41467-018-03705-y
4. Wilhelm S, Tavares AJ, Dai Q, et al. Analysis of nanoparticle supply to tumours. Nat Rev Mater. 2016;1(5):16014. doi:10.1038/natrevmats.2016.14
5. Cheng YH, He CL, Riviere JE, Monteiro-Riviere NA, Lin ZM. Meta-analysis of nanoparticle supply to tumors utilizing a physiologically based mostly pharmacokinetic modeling and simulation method. ACS Nano. 2020;14(3):3075–3095. doi:10.1021/acsnano.9b08142
6. Anselmo AC, Mitragotri S. Nanoparticles within the clinic: an replace. Bioeng Transl Med. 2019;4(3):e10143. doi:10.1002/btm2.10143
7. Riviere JE. Of mice, males and nanoparticle biocoronas: are in vitro to in vivo correlations and interspecies extrapolations lifelike? Nanomedicine. 2013;8(9):1357–1359. doi:10.2217/nnm.13.129
8. Lin Z, Monteiro-Riviere NA, Riviere JE. Pharmacokinetics of metallic nanoparticles. Wiley Interdiscip Rev Nanomed Nanobiotechnol. 2015;7(2):189–217. doi:10.1002/wnan.1304
9. Valic MS, Zheng G. Research instruments for extrapolating the disposition and pharmacokinetics of nanomaterials from preclinical animals to people. Theranostics. 2019;9(11):3365–3387. doi:10.7150/thno.34509
10. Price LSL, Stern ST, Deal AM, Kabanov AV, Zamboni WC. A reanalysis of nanoparticle tumor supply utilizing classical pharmacokinetic metrics. Sci Adv. 2020;6(29):eaay9249. doi:10.1126/sciadv.aay9249
11. Lee HA, Leavens TL, Mason SE, Monteiro-Riviere NA, Riviere JE. Comparison of quantum dot biodistribution with a blood-flow-limited physiologically based mostly pharmacokinetic mannequin. Nano Lett. 2009;9(2):794–799. doi:10.1021/nl803481q
12. Mager DE, Mody V, Xu C, et al. Physiologically based mostly pharmacokinetic mannequin for composite nanodevices: impact of cost and measurement on in vivo disposition. Pharm Res. 2012;29(9):2534–2542. doi:10.1007/s11095-012-0784-7
13. Bachler G, von Goetz N, Hungerbuhler Ok. A physiologically based mostly pharmacokinetic mannequin for ionic silver and silver nanoparticles. Int J Nanomedicine. 2013;8:3365–3382. doi:10.2147/IJN.S46624
14. Chen WY, Cheng YH, Hsieh NH, et al. Physiologically based mostly pharmacokinetic modeling of zinc oxide nanoparticles and zinc nitrate in mice. Int J Nanomedicine. 2015;10:6277–6292. doi:10.2147/IJN.S86785
15. Carlander U, Li D, Jolliet O, Emond C, Johanson G. Toward a basic physiologically-based pharmacokinetic mannequin for intravenously injected nanoparticles. Int J Nanomedicine. 2016;11:625–640. doi:10.2147/IJN.S94370
16. Li D, Morishita M, Wagner JG, et al. In vivo biodistribution and physiologically based mostly pharmacokinetic modeling of inhaled contemporary and aged cerium oxide nanoparticles in rats. Part Fibre Toxicol. 2016;13(1):45. doi:10.1186/s12989-016-0156-2
17. Lin Z, Monteiro-Riviere NA, Riviere JE. A physiologically based mostly pharmacokinetic mannequin for polyethylene glycol-coated gold nanoparticles of various sizes in grownup mice. Nanotoxicology. 2016;10(2):162–172. doi:10.3109/17435390.2015.1027314
18. Yuan D, He H, Wu Y, Fan J, Cao Y. Physiologically based mostly pharmacokinetic modeling of nanoparticles. J Pharm Sci. 2019;108(1):58–72. doi:10.1016/j.xphs.2018.10.037
19. Price E, Gesquiere AJ. Animal simulations facilitate good drug design by means of prediction of nanomaterial transport to particular person tissue cells. Sci Adv. 2020;6(4):eaax2642. doi:10.1126/sciadv.aax2642
20. Lin Z, Monteiro-Riviere NA, Kannan R, Riviere JE. A computational framework for interspecies pharmacokinetics, publicity and toxicity evaluation of gold nanoparticles. Nanomedicine. 2016;11(2):107–119. doi:10.2217/nnm.15.177
21. Baskin II. Machine studying strategies in computational toxicology. Methods Mol Biol. 2018;1800:119–139.
22. Basile AO, Yahi A, Tatonetti NP. Artificial intelligence for drug toxicity and security. Trends Pharmacol Sci. 2019;40(9):624–635. doi:10.1016/j.ideas.2019.07.005
23. Bhhatarai B, Walters WP, Hop C, Lanza G, Ekins S. Opportunities and challenges utilizing synthetic intelligence in ADME/Tox. Nat Mater. 2019;18(5):418–422. doi:10.1038/s41563-019-0332-5
24. Singh AV, Ansari MHD, Rosenkranz D, et al. Artificial intelligence and machine studying in computational nanotoxicology: unlocking and empowering nanomedicine. Adv Healthc Mater. 2020;9(17):e1901862. doi:10.1002/adhm.201901862
25. Tu JV. Advantages and drawbacks of utilizing synthetic neural networks versus logistic regression for predicting medical outcomes. J Clin Epidemiol. 1996;49(11):1225–1231. doi:10.1016/S0895-4356(96)00002-9
26. Bourquin J, Schmidli H, van Hoogevest P, Leuenberger H. Advantages of Artificial Neural Networks (ANNs) as various modelling method for knowledge units displaying non-linear relationships utilizing knowledge from a galenical examine on a strong dosage type. Eur J Pharm Sci. 1998;7(1):5–16. doi:10.1016/S0928-0987(97)10028-8
27. Lee S, Jung S, Lee J. Prediction mannequin based mostly on a synthetic neural community for user-based constructing vitality consumption in South Korea. Energies. 2019;12(4):608. doi:10.3390/en12040608
28. Lazarovits J, Sindhwani S, Tavares AJ, et al. Supervised studying and mass spectrometry predicts the in vivo destiny of nanomaterials. ACS Nano. 2019;13(7):8023–8034. doi:10.1021/acsnano.9b02774
29. Hancock JT, Khoshgoftaar TM. Survey on categorical knowledge for neural networks. J Big Data. 2020;7(1):28. doi:10.1186/s40537-020-00305-w
30. Alshdaifat E, Alshdaifat D, Alsarhan A, Hussein F, El-Salhi SMFS. The impact of preprocessing strategies, utilized to numeric options, on classification algorithms’ efficiency. Data. 2021;6(2):11. doi:10.3390/data6020011
31. Kuhn M. Building predictive fashions in R utilizing the caret package deal. J Stat Softw. 2008;28(5):1–26. doi:10.18637/jss.v028.i05
32. Karatzoglou A, Feinerer I. Kernel-based machine studying for quick textual content mining in R. Comput Stat Data Anal. 2010;54(2):290–297. doi:10.1016/j.csda.2009.09.023
33. Liaw A, Mathew W. Classification and regression by randomForest. R News. 2002;2:18–22.
34. Chen T, Guestrin C. XGBoost: a scalable tree boosting system. In Proceedings of the twenty second ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. New York, NY, USA: ACM; 2016: 785–794.
35. Bergstra J, Bengio Y. Random seek for hyper-parameter optimization. J Mach Learn Res. 2012;13:281–305.
36. LeDell E, Aiello NGS, Fu A, et al. h2o: r Interface for the ‘H2O’ scalable machine studying platform. R package deal model 3.30.0.1; 2020: 13.
37. Schmidhuber J. Deep studying in neural networks: an outline. Neural Netw. 2015;61:85–117. doi:10.1016/j.neunet.2014.09.003
38. Srivastava N, Hinton G, Krizhevsky A, Sutskever I, Salakhutdinov R. Dropout: a easy method to forestall neural networks from overfitting. J Mach Learn Res. 2014;15:1929–1958.
39. Prechelt L. Early stopping – But when? Neural Netw. 1998;1524:55–69.
40. Gedeon TD. Data mining of inputs: analysing magnitude and practical measures. Int J Neural Syst. 1997;8(2):209–218. doi:10.1142/S0129065797000227
41. Yan XL, Sedykh A, Wang WY, Yan B, Zhu H. Construction of a web-based nanomaterial database by huge knowledge curation and modeling pleasant nanostructure annotations. Nat Commun. 2020;11(1):2519. doi:10.1038/s41467-020-16413-3
42. Vitsios D, Dhindsa RS, Middleton L, Gussow AB, Petrovski S. Prioritizing non-coding areas based mostly on human genomic constraint and sequence context with deep studying. Nat Commun. 2021;12(1):1504. doi:10.1038/s41467-021-21790-4
43. Chen R, Zhang Y, Sahneh FD, et al. Nanoparticle floor characterization and clustering by means of concentration-dependent floor adsorption modeling. ACS Nano. 2014;8(9):9446–9456. doi:10.1021/nn503573s
44. Aiken EL, Nguyen AT, Viboud C, Santillana M. Toward using neural networks for influenza prediction at a number of spatial resolutions. Sci Adv. 2021;7(25):eabb1237. doi:10.1126/sciadv.abb1237
45. Yu F, Wei C, Deng P, Peng T, Hu X. Deep exploration of random forest mannequin boosts the interpretability of machine studying research of sophisticated immune responses and lung burden of nanoparticles. Sci Adv. 2021;7(22):eabf4130. doi:10.1126/sciadv.abf4130
46. Baghban R, Roshangar L, Jahanban-Esfahlan R, et al. Tumor microenvironment complexity and therapeutic implications at a look. Cell Commun Signal. 2020;18(1):59. doi:10.1186/s12964-020-0530-4
47. Zhang M, Gao S, Yang D, et al. Influencing components and methods of enhancing nanoparticles into tumors in vivo. Acta Pharm Sin B. 2021;11(8):2265–2285. doi:10.1016/j.apsb.2021.03.033
48. Fan F, Xie B, Yang L. Promoting nanoparticle supply effectivity to tumors by domestically rising blood circulate there. ACS Appl Bio Mater. 2021;4(10):7615–7625. doi:10.1021/acsabm.1c00871
49. Honary S, Zahir F. Effect of Zeta potential on the properties of nano-drug supply techniques – A assessment (Part 1). Trop J Pharm Res. 2013;12(2):255–264.
50. Mitchell MJ, Billingsley MM, Haley RM, Wechsler ME, Peppas NA, Langer R. Engineering precision nanoparticles for drug supply. Nat Rev Drug Discov. 2021;20(2):101–124. doi:10.1038/s41573-020-0090-8
51. Riad MH, Baynes RE, Tell LA, et al. Development and utility of an interactive Physiologically Based Pharmacokinetic (iPBPK) mannequin to predict oxytetracycline tissue distribution and withdrawal intervals in market-age sheep and goats. Toxicol Sci. 2021;183(2):253–268. doi:10.1093/toxsci/kfab095
52. Li M, Cheng YH, Chittenden JT, et al. Integration of Food Animal Residue Avoidance Databank (FARAD) empirical strategies for drug withdrawal interval willpower with a mechanistic population-based interactive physiologically based mostly pharmacokinetic (iPBPK) modeling platform: instance for flunixin meglumine administration. Arch Toxicol. 2019;93(7):1865–1880. doi:10.1007/s00204-019-02464-z
53. Punt A, Pinckaers N, Peijnenburg A, Louisse J. Development of a web-based toolbox to help Quantitative In-Vitro-to-In-Vivo Extrapolations (QIVIVE) inside nonanimal testing methods. Chem Res Toxicol. 2021;34(2):460–472. doi:10.1021/acs.chemrestox.0c00307
54. Gao H, Wang W, Dong J, Ye Z, Ouyang D. An built-in computational methodology with data-driven machine studying, molecular modeling and PBPK modeling to speed up strong dispersion formulation design. Eur J Pharm Biopharm. 2021;158:336–346. doi:10.1016/j.ejpb.2020.12.001
55. Schneckener S, Grimbs S, Hey J, et al. Prediction of oral bioavailability in rats: transferring insights from in vitro correlations to (Deep) machine studying fashions utilizing in silico mannequin outputs and chemical construction parameters. J Chem Inf Model. 2019;59(11):4893–4905. doi:10.1021/acs.jcim.9b00460
56. Jones H, Rowland-Yeo Ok. Basic ideas in physiologically based mostly pharmacokinetic modeling in drug discovery and growth. CPT Pharmacomet Syst Pharmacol. 2013;2(8):e63. doi:10.1038/psp.2013.41
57. Lin Z, Fisher JW. Chapter 1 – A historical past and up to date efforts of chosen physiologically based mostly pharmacokinetic modeling matters. In: Fisher JW, Gearhart J, Lin Z, editors. Physiologically Based Pharmacokinetic (PBPK) Modeling: Methods and Applications in Toxicology and Risk Assessment. 1st ed. Elsevier, Academic Press; 2020:1–26.
58. Tan YM, Worley RR, Leonard JA, Fisher JW. Challenges related to making use of physiologically based mostly pharmacokinetic modeling for public well being decision-making. Toxicol Sci. 2018;162(2):341–348. doi:10.1093/toxsci/kfy010
59. Lin Z, Gehring R, Mochel JP, Lave T, Riviere JE. Mathematical modeling and simulation in animal well being – Part II: ideas, strategies, purposes, and worth of physiologically based mostly pharmacokinetic modeling in veterinary drugs and meals security evaluation. J Vet Pharmacol Ther. 2016;39(5):421–438. doi:10.1111/jvp.12311
https://www.dovepress.com/predicting-nanoparticle-delivery-to-tumors-using-machine-learning-and–peer-reviewed-fulltext-article-IJN
