Development of Theranostic Perfluorocarbon Nanoemulsions as a Model Non-Opioid Pain Nanomedicine Using a Quality by Design (QbD) Approach

Abstract

Pain nanomedicine is an emerging field in response to current needs of addressing the opioid crisis in the USA and around the world. Our group has focused on the development of macrophage-targeted perfluorocarbon nanoemulsions as inflammatory pain nanomedicines over the past several years. We present here, for the first time, a quality by design approach used to design pain nanomedicine. Specifically, we used failure mode, effects, and criticality analysis (FMECA) which identified the process and composition parameters that were most likely to impact nanoemulsion critical quality attributes (CQAs). From here, we applied a unique combination approach that compared multiple linear regression, boosted decision tree regression, and partial least squares regression methods in combination with correlation plots. The presented combination approach allowed for in-depth analyses of which formulation steps in the nanoemulsification processes control nanoemulsion droplet diameter, stability, and drug loading. We identified that increase in solubilizer (transcutol) content increased drug loading and decreased nanoemulsion stability. This was mitigated by inclusion of perfluorocarbon oil in the internal phase. We observed negative correlation (R² = 0.4357, p value 0.0054) between the amount of PCE and the percent diameter increase (destabilization), and no correlation between processing parameters and percent diameter increase over time. Further, we identified that increased sonication time decreases nanoemulsion drug loading but does not significantly impact droplet diameter or stability. We believe the methods presented here can be useful in the development of various nanomedicines to produce higher-quality products with enhanced manufacturing and design control.

INTRODUCTION

As of 2017, 50 nanomedicines had been approved by the FDA for clinical use (1). Only three (6%) of these approved nanomedicines are indicated for pain treatment (1). This is disproportionate to the number of Americans living with chronic pain, which is greater than the number of Americans living with diabetes, heart disease, and cancer combined (2). Nanomedicines have the potential to improve pain treatment and reduce the need for opioids by providing targeted delivery to the affected area while reducing adverse systemic effects. However, there are many concerns facing not just pain nanomedicines, but all nanomedicines, including safety, toxicity, characterization, regulation, and cost (1). Additional challenges that may prevent the development of quality nanomedicines include (1) a lack of sufficient quality control in early development, (2) an incomplete understanding of process impact on physicochemical properties, (3) difficulty in reproducibility and scale-up, and (4) a lack of diagnostic approaches that assess disease progression and treatment efficacy. Adapting quality by design (QbD) approaches used in the pharmaceutical industry to nanomedicine development in the academic setting may help to overcome these challenges and result in higher-quality nanomedicines.

A QbD approach involves assignment of critical quality attributes (CQAs), measurable quality attributes that have a significant impact on final product quality. Risk analyses are used to strategically rank failure modes during product development, narrowing down the list of process parameters to those that are most likely to significantly impact the CQAs. These select process parameters can then be studied using a design of experiments (DoE), an approach to experimental design in which multiple factors are changed at a time. This minimizes the required number of experimental runs and increases efficiency compared to traditional one factor at a time approaches (3). When CQAs are measured for each run in a DoE, statistical regression approaches such as multiple linear regression, boosted decision tree regression, and partial least squares can be used to identify which potential critical process parameters have a significant impact on the CQAs.

A QbD approach emphasizes product quality and process understanding. When applied to nanomedicine, this may result in higher-quality products that are more likely to perform consistently and effectively in in vivo testing. QbD approaches are becoming more commonly used in nanomedicine development and have been applied to liposomes, emulsions, particles, micelles, and suspensions (4). Shah et al. used a QbD approach in the development of venlafaxine nanostructured lipid carriers (5). A full factorial 3² design was used to evaluate the impact of drug to lipid ratio and surfactant concentration on the CQAs of size, polydispersity index (PDI), and entrapment efficiency. Srinivas et al. used a QbD approach to develop gefitinib nanosuspensions (6). Polynomial equations and response surface plots were developed to predict particle size, PDI, and zeta potential as a function of polyvinylpyrrolidone content, polyvinyl alcohol content, and sonication time. This is an example of a mixture process variable design, which can be beneficial in nanomedicine development because it enables understanding of both process and composition on physicochemical properties (size, zeta potential, etc.). This can assist in nanomedicine optimization and scale-up.

Mixture process variable design can also be applied to nanoemulsion development. Nanoemulsions are colloidal dispersions of two or more immiscible phases ranging typically from 100 to 400 nm in diameter (7). Nanoemulsions are kinetically stable emulsions that are typically produced using high-energy processes such as sonication and microfluidization (8). This makes nanoemulsions an attractive candidate for study using QbD approaches that include mixture process variable design of experiments. Nanoemulsions are also promising carriers for anti-inflammatory drugs and/or natural products (9–11) and potential future pain nanomedicines. Non-steroidal anti-inflammatory drugs (NSAIDs) are commonly prescribed to treat pain, as these drugs are non-addictive alternatives to opioids. However, NSAID use is associated with gastrointestinal irritation and cardiovascular toxicities (12–14). For example, celecoxib, a cyclooxygenase-2 (COX-2) selective NSAID, can increase the risk of gastrointestinal events and peripheral edema (15,16). Celecoxib is also poorly water soluble with a complex metabolic profile and excretion profile with significant interpatient variability (17). Thus, higher doses may be required to achieve sufficient pain relief in some patients, which may increase the risk of adverse side effects (18). Encapsulation of celecoxib into nanoemulsions for transdermal delivery was attempted to both increase overall bioavailability and overcome first-pass metabolism (19,20). Our work so far has focused on using nanoemulsions for parenteral delivery to both enhance bioavailability and improve tissue targeting. In previous studies, we showed that nanoemulsions can be designed to provide targeted celecoxib delivery to inflammatory macrophages, decreasing the effective analgesic dose by over 2000-fold compared to traditional oral administration (21). Further, developed celecoxib nanoemulsions can be produced with imaging agents (near infrared and/or magnetic resonance) that enable tracking of macrophage migration patterns in response to inflammation or injury. This combination therapeutic and diagnostic platform is referred to as a theranostic medicine. We have demonstrated that theranostic celecoxib nanoemulsions containing near infrared (NIR) dye reduce macrophage infiltration for 72 h in the footpad of mice that have been injected with complete Freund’s adjuvant (CFA) (22). Theranostic celecoxib nanoemulsions containing perfluorocarbon (magnetic resonance imaging (MRI) agent) were also used to simultaneously image and treat pain and neuroinflammation in a rat model of chronic constriction injury (23). These theranostic nanoemulsions have the potential to create a more personalized pain treatment through the combination of imaging and drug delivery properties into one nanosystem (21,22,24–31).

Janjic et al. demonstrated that celecoxib loaded nanoemulsions (0.185 mg/mL) can provide 4 days of pain relief (tested by mechanical allodynia) in rats with chronic constriction injury (21). Increasing nanoemulsion celecoxib concentration may prolong analgesia beyond 4 days (21) and reduce overall nanoemulsion dosing volume. This can reduce drug accumulation in off-target organs, further reducing the risk of adverse cardiovascular and gastrointestinal effects. The goal of the presented work was to develop a new celecoxib nanoemulsion capable of loading ten times more celecoxib than previously reported formulations (21,22). To achieve this goal, a QbD approach including a mixture process variable DoE was used to optimize celecoxib nanoemulsions to increase drug loading ten times compared to our previously reported formulations (21–23,31). Extensive risk analysis and quality control tests were conducted to identify stable celecoxib nanoemulsions with and without the MRI agent perfluoro-15-crown-5 ether (PCE). Statistical analyses (correlation plots) and multiple statistical regression approaches (multiple linear regression, gradient boosted decision tree regression, and partial least squares regression) were used to improve understanding of the processes that control nanoemulsion droplet diameter, stability, and drug loading. The approaches presented here can be adapted to other nanomedicine platforms to improve product quality and process understanding.

MATERIALS AND METHODS

Celecoxib was purchased from LC lab C-1502 Lot no. CXB-101. Miglyol 812N was generously donated by Cremer Oleo, product number 6330. Perfluro-15-crown-5-ether was purchased from Exfluor Research Company (Product number F15CROWN5). Transcutol HP 2-(2-ethoxyethoxy)-ethanol was purchased from Spectrum, catalog number E1022.

Nanoemulsion Formulation

Developed nanoemulsions were composed of celecoxib, two solubilizers (miglyol 812 and transcutol), two surfactants (cremophor EL and pluronic P105), and water. Additionally, some nanoemulsions contained the perfluorocarbon perfluoro-15-crown-5-ether (PCE). Our previously published celecoxib nanoemulsions used miglyol as the sole solubilizer for celecoxib (23,31). To increase drug loading ten times, it was necessary to introduce a co-solubilizer into the formulation. We chose transcutol as the co-solubilizer because it is a commonly used FDA-approved solubilizer that has been used in microemulsion-based hydrogels containing celecoxib (32). PCE is a perfluorocarbon that was incorporated into some presented nanoemulsions as a magnetic resonance (MR) imaging agent. In previous publications, we have demonstrated that perfluorocarbon nanoemulsions allow for in vivo tracking of nanoemulsion biodistribution using MR (23). In these complex formulations, perfluorocarbon (PCE in the presented work) makes up the core of the nanoemulsion droplet. PCE is surrounded by a hydrocarbon shell consisting of celecoxib, miglyol, and any additional solubilizers (transcutol in the presented work). Nanoemulsion droplets are stabilized in water with a mixture of two surfactants, cremophor EL, and pluronic P105 (23,31).

Nanoemulsion Manufacture

Celecoxib was dissolved in a solution of miglyol and transcutol by stirring for 24 h at 350 rpm (33). PCE was added to the celecoxib solution and vortexed on high for 30 s. Surfactant solution (3% w/v Cremophor EL, 2% w/v Pluronic P105) was added to the mixture and vortexed for approximately 30 s to form a coarse emulsion. The coarse emulsion then underwent sonication using a Sonic Dismembrator (Model 500, Fisher Scientific). Sonication frequency was set to 29% amplitude, and the emulsion was chilled on ice during sonication. Emulsions were then processed on a microfluidizer M110S (Microfluidics Corporation, Westwood MA). Microfluidization pressure was set to 80 psi. The microfluidization chamber was iced for a minimum of 1 h prior to nanoemulsion manufacture.

Dynamic Light Scattering

Diameter and polydispersity index measurements were performed using dynamic light scattering (Zetasizer NanoZS, Malvern, UK). Nanoemulsions were diluted 1:80 v/v in de-ionized water. Measurements were made at 25°C at a light scattering angle of 173°.

Nanoemulsion Drug Loading

Celecoxib content was measured in nanoemulsions using high-performance liquid chromatography (DIONEX Ultimate 3000, Thermo Scientific^™). Nanoemulsions were diluted in methanol to a theoretical concentration of 6.67 μg/mL. If the sample contained PCE, it was then centrifuged at 3000 rpm for 10 min, allowing the undissolved perfluorocarbon to form a pellet. Methanol supernatant was then diluted in water to a theoretical concentration of 5 μg/mL. An isocratic method of 75% methanol, 25% water was used to analyze samples. A celecoxib peak would appear at approximately 7 min at wavelength 255 nm.

Risk Assessment

A failure mode, effects, and criticality analysis (FMECA) was used as a risk assessment tool to strategically rank methods of failure during nanoemulsion production and identify potential critical process parameters. Failure modes were assigned values between 1 and 5, with a value of 5 being the most severe, in each of the following three categories: severity (impact on final product), frequency of occurrence, and detectability. The product of these three values was defined as the risk priority number (RPN). RPNs were used to rank failure modes and identify potential critical process parameters.

Mixture Process Variable Design of Experiments

JMP Pro13 software was used to develop a custom D-optimal mixture process variable design of experiments. Three mixture factors were studied: miglyol (1.9, 3.8 mL), transcutol (0.2, 0.8 mL), and PCE (0, 1.9 mL). Mixture variables were defined such that the sum of the three components equaled 4.0 mL. Two process factors were also studied: microfluidization pulses (15, 30, 45) and sonication time (0, 1, 2 min). In addition to these five factors, six interaction variables were studied: miglyol*transcutol, miglyol*PCE, transcutol*PCE, microfluidization pulses*sonication time, microfluidization pulses*microfluidization pulses, and sonication time*sonication time.

Regression Modeling and Statistical Analyses

All regression modeling and statistical analyses were performed using JMP Pro13 statistical software. Multiple linear regression (MLR) models were generated using standard least squares approach. A backward stepwise regression approach was used to eliminate variables that did not significantly impact the output of interest. All main effects terms, mixture-mixture interaction terms, and process-process interaction terms were studied. Terms were only included in the model if they had a p value < 0.05. Correlation trends were deemed statistically significant if the F-test of the fitted regression line yielded a p value less than 0.05. Boosted models were developed using only main effects. Model parameters included a maximum of 200 layers, learning rate ranging from 0.1 to 0.8, number of splits per tree ranging from 1 to 8, and minimum split size between 3 and 4. Partial least squares (PLS) regression models were developed using all main effects, mixture-mixture interactions, and process-process interactions. Two sequential PLS models were developed and compared for each studied CQA specification. The first PLS model included all studied variables. The number of latent variables in each PLS model was selected to minimize the prediction error sum of squares (PRESS). Variable importance for projection (VIP) vs. coefficient plots were used to identify VIP variables with a threshold above 0.8. These VIP variables were then used to develop a second PLS model. The PLS model with the lowest PRESS, highest R², and lowest root average square error (RASE) was selected as the final PLS model. To enable direct comparison between MLR, boosted tree, and PLS models, all models were trained on a portion of the experimental runs and validated using the runs which were withheld from the training.

Filtration Stability

Nanoemulsions were filtered through a Millex-GS syringe filter with a pore size of 0.22 μm, as this sterilization step is necessary for in vitro biological testing. Filtered nanoemulsion was diluted 1:80 v/v in de-ionized water, and diameter and polydispersity index were measured using dynamic light scattering. A same-day measurement of unfiltered nanoemulsion was used as a control.

Centrifugation Stability

Filtered nanoemulsions were diluted 1:80 v/v in de-ionized water, Dulbecco’s modified Eagle’s medium (DMEM), 10% fetal bovine serum (FBS) in DMEM, and 20% FBS in DMEM to a total volume of 2 mL in a 15 mL plastic centrifuge tube. Nanoemulsion dilutions were centrifuged at 1620×g for 30 min. Nanoemulsion diameter and polydispersity index were then measured without additional sample dilution. Same-day measurements of non-centrifuged nanoemulsion were used as a control.

Serum Stability

One week after production, nanoemulsion was diluted in three biological media (DMEM, 10% FBS in DMEM, and 20% FBS in DMEM) at 1:80 v/v to a total volume of 4 mL in a 15-mL sterile centrifuge tube. Droplet diameter and PDI were measured using DLS immediately (0 h time point) after dilution. Remaining dilutions were incubated at 37°C. Samples were removed and measured every 24 h for 72 h. Nanoemulsion dilution in de-ionized water was used as a control.

Thermal Cycling

Undiluted, filtered nanoemulsion (5.0 mL) was aliquoted to a 15-mL glass vial. The vials were sealed with parafilm and stored at 4°C. After 24 h, vials were moved to an incubator set to 50°C. Every 24 h, vials were moved between 4 and 50°C for a total of four thermal cycles (8 days). Upon completion, diameter and PDI of thermal cycling study samples were measured after samples were allowed to equilibrate to room temperature for 1 h. Same-day measurements of filtered nanoemulsion continuously stored at ambient temperature were used as a control.

Sterility Test

pH measurements were performed using an Orion^™ 8172BNWP ROSS^™ Sure-Flow^™ (Thermo Scientific^™). Methods were adopted from USP 34 microbiological tests/71 sterility test (34). Nanoemulsion was diluted 1:80 v/v in (1) fluid thioglycolate medium (Millipore, Catalog STBMCTM12) or (2) trypcase soy broth (Millipore, Catalog STBMTSB12). pH was recorded immediately after dilution. Fluid thioglycolate medium dilutions were stored in a water bath at 35°C for 14 days. Trypcase soy broth dilutions were incubated at 25°C for 14 days. After 14 days’ incubation, pH was recorded. A visual test was also performed, with increased cloudiness indicating microorganism growth. For each test, medium or broth incubated without nanoemulsion was used as a control.

RESULTS

Identification of Critical Quality Attributes

The primary goals of this work were to (1) quickly and efficiently identify stable nanoemulsions that solubilized celecoxib at a concentration of 1.85 mg/mL and (2) improve understanding of nanoemulsion composition and manufacturing processes and their impact on nanoemulsion droplet diameter, stability, and drug loading. This was accomplished by adapting quality by design (QbD) methodology to nanoemulsion production. Critical quality attributes (CQAs), defined as measurable quality attributes that have a significant impact on final product quality, were identified and categorized into two phases (Table I). Basic physicochemical properties were categorized as phase 1 CQAs, and all nanoemulsion formulations were evaluated using these parameters. Phase 1 CQA specifications were used to identify the most promising formulations, and these select nanoemulsions underwent additional stability testing (phase 2 CQAs).

Table I.

Critical Quality Attributes (CQAs), Specifications, and Brief Testing Method Descriptions. All Nanoemulsions Were Evaluated for Phase 1 CQAs Outlined in the Top Half of the Table (Droplet Diameter, PDI, pH, and Drug Loading). Select Nanoemulsions that Passed All CQA Specifications Underwent Additional Stability and Sterility Testing (Phase 2, Bottom Half of Table)

Phase 1—Evaluated on all formulations
Droplet diameter	100–150 nm	DLSa
PDIb	<0.20	DLS
90 day diameter change	< 10%	DLS
90 day PDI	<0.20	DLS
pH	4.5–8.0	pH meter
Drug loading	> 80%	HPLCc
Phase 2—Evaluated on select formulations
Filtration stability	Droplet diameter change < 10% PDI < 0.20	0.22 um pore sizer filter
Centrifugation stability	Droplet diameter change < 10% PDI < 0.20	1620rcf for 30 min
Serum stability	Droplet diameter change < 20 nm PDI < 0.20	72 h incubation in biological medium (37 °C)
Thermal cycling stability	Droplet diameter change < 10% PDI < 0.20	Alternate between 4 and 50 °C every 24 h for 8 days
Sterility	No significant pH change	14-day incubation in thioglycolate broth (35 °C) and trypticase soy broth (25 °C)

^aDynamic light scattering

^bPolydispersity index

^cHigh-performance liquid chromatography

Risk Assessment and Selection of Design Parameters

A failure mode, effects, and criticality analysis (FMECA) was used to rationally identify the parameters that were most likely to impact nanoemulsion CQAs (Table II). This was accomplished by identifying and ranking potential failure modes using risk priority numbers (RPNs). In the presented work, the highest RPN scores were associated with the number of microfluidization pulses, temperature during microfluidization and sonication, sonication time, and nanoemulsion composition (amount of solubilizer, perfluorocarbon to oil ratio, etc.). Since temperature cannot be accurately controlled during these processes, we chose to focus solely on nanoemulsion composition, number of microfluidization pulses, and sonication time. These parameters were studied using a 16-run mixture process variable design of experiments (Table III). All 16 of these nanoemulsions were evaluated using phase 1 CQA specifications (Table I). All formulations met the CQA specifications for day 2 diameter, day 2 PDI, 90-day PDI, and pH (Table III). However, five formulations failed to meet the drug loading CQA specification and 12 formulations failed to meet the 90-day percent diameter increase CQA specification. Therefore, potential causes for these failures were investigated.

Table II.

Abridged Failure Modes, Effects, and Criticality Analysis (FMECA). Potential Failure Modes Were Identified for Each Phase 1 CQA and Ranked Using a Risk Priority Number (RPN). RPNs Were Used to Identify That Nanoemulsion Composition, Microfluidization Pulses, and Sonication Time Were Most Likely to Influence the Phase 1 CQAs. Therefore, These Parameters Were Studied in a Mixture Process Variable Design of Experiments (see Table III). Color legend: Green = low risk; orange = medium risk; red = high risk

S	F	D	RPN	CQA Impacted	Unit Operation	Cause of Failure
4	1	5	20	Droplet diameter and/or PDI	Microfluidization	High interaction chamber temperature
4	3	5	60			Too many/too few pulses
4	2	4	32		Sonication	Frequency too high/low
4	3	4	48			Time too long/short
4	1	4	16			Temperature too high
4	3	4	48		Dissolving and Dispensing	High PFC:oil ratio
4	3	4	48			Too much solubilizer
4	2	4	32			High oil:surfactant ratio
4	2	5	40	90-day diameter change and/or 90-day PDI	Dissolving and Dispensing	High oil:surfactant ratio
4	3	5	60			Too much solubilizer
4	2	5	40			Too much perfluorocarbon
4	3	5	60			Large initial PDI (day 2)
4	3	3	36		Storage	Room temperature fluctuations
5	1	5	25	pH	Microfluidization	Overheating causes drug/excipient degradation
5	1	5	25		Sonication
5	1	2	10		Storage	Light exposure causes drug/excipient degradation
4	1	5	20	Drug Loading	Microfluidization	High shear rate causes drug degradation
4	3	5	60			Cold interaction chamber causes drug precipitation
4	3	5	60		Sonication	High frequency and/or long sonication time causes drug precipitation
4	3	5	60			Cold temperature causes drug precipitation
4	1	2	8		Storage	Light exposure causes drug degradation

Table III.

A 16-Run Mixture Process Variable Design of Experiments Was Developed to Study the Impact of Nanoemulsion Composition and Manufacturing Process on Phase 1 CQA Specifications. Main Effects, Mixture-Mixture Interactions, and Process-Process Interactions Were Studied. Measurements that Satisfied the CQA Specification Are Italicized. Nanoemulsions 3, 4, and 16 Satisfy All Phase 1 CQA Specifications

Nanoemulsion	Miglyol (mL)	Transcutol (mL)	PCE (mL)	Sonication time (min.)	Microfluidization pulses	Droplet diameter (nm)	90-day diameter Change (%)	PDI	90 day PDI	pH	Drug loading (%)
	Design parameters					CQA measurements
1	3.20	0.80	0	0	30	116.70	19.77	0.101	0.071	5.98	87.16
2	1.90	0.80	1.30	0	15	127.77	14.77	0.162	0.119	6.00	90.84
3	1.95	0.46	1.6	0	45	116.63	8.57	0.138	0.099	5.60	82.72
4	2.85	0.20	0.95	0	45	120.47	6.78	0.108	0.085	6.05	84.68
5	3.20	0.80	0	2	15	125.23	16.32	0.106	0.078	5.96	84.18
6	3.20	0.80	0	1	45	113.70	18.67	0.097	0.082	5.73	80.98
7	1.90	0.20	1.90	2	45	122.33	6.76	0.125	0.122	5.82	72.94
8	1.90	0.20	1.90	1	15	133.27	10.08	0.144	0.154	6.05	76.56
9	1.90	0.80	1.30	1	45	115.33	12.14	0.148	0.111	5.48	83.73
10	3.80	0.20	0	2	15	131.70	12.30	0.135	0.082	6.06	80.02
11	3.80	0.20	0	2	45	122.93	13.50	0.110	0.058	6.08	78.73
12	3.80	0.20	0	1	30	128.87	11.23	0.116	0.072	6.12	78.81
13	2.73	0.51	0.76	1	15	128.47	12.04	0.153	0.115	5.93	79.10
14	3.04	0.20	0.76	0	15	136.10	11.29	0.130	0.099	6.00	82.45
15	2.22	0.63	1.15	2	30	122.50	10.37	0.122	0.107	5.57	82.52
16	2.66	0.20	1.14	2	30	126.43	6.43	0.101	0.091	5.73	83.31

Evaluating Nanoemulsion Diameter, Diameter Increase, and Drug Loading

To the best of our knowledge, this is the first literature report that extensively investigates the parameters that impact stability and drug loading of complex nanoemulsions with more than two immiscible phases (perfluorocarbon, hydrocarbon, and aqueous). Therefore, correlation plots were developed as a basic starting point to provide us with a preliminary understanding of potential trends between process parameters and CQAs. Correlation plots were developed for the CQA specifications of nanoemulsion diameter, percent diameter increase, and drug loading (Supplemental Fig. 1). Nanoemulsion parameters (miglyol, transcutol, PCE, microfluidization pulses, and sonication time) were plotted as a function of the CQA specification of interest. R² and p values were calculated for each correlation plot (Table IV). Correlations trends with p value < 0.05 were considered statistically significant (gray highlights, Table IV). Transcutol content and the number of microfluidization pulses were found to significantly impact nanoemulsion diameter. However, miglyol and PCE vs. nanoemulsion diameter correlation trends were not significant. Percent diameter increase correlation plots showed significant correlation trends for transcutol content and PCE content. A significant correlation trend was not observed between miglyol and percent diameter increase. Finally, drug loading showed significant correlation trends for transcutol content and sonication time.

Table IV.

Correlation R² and p Values Calculated for Nanoemulsion Day 2 Diameter, 90-day Diameter Increase, and Drug Loading as a Function of Each Parameter Studied in the Mixture Process Variable Design of Experiments. Statistically Significant Correlations (p Value < 0.05) Are Italicized

		Parameter
Response	Statistics parameter	Miglyol	Transcutol	PCE	Microfluidization pulses	Sonication time
Day 2 diameter (nm)	R 2	0.0275	0.2948	0.0018	0.6360	0.0116
	p value	0.5395	0.0298	0.8751	0.0002	0.6914
90-day diameter Increase (%)	R 2	0.1462	0.4975	0.4357	0.0376	0.0212
	p value	0.1438	0.0023	0.0054	0.4714	0.5904
Drug loading (%)	R 2	0.0048	0.3547	0.0258	0.0273	0.2668
	p value	0.7984	0.0149	0.5524	0.5406	0.0405

To understand factors and interactions that may cause nanoemulsions to fail to meet the CQA specification for 90-day diameter increase, multiple linear regression (MLR) models were developed to predict (1) nanoemulsion diameter on day 2 (48 h) after production (Fig. 1a) and (2) percent diameter increase over 90-day storage at ambient temperature (Fig. 1b). The mixture regression terms of these MLR models are expressed in terms of l-pseudocomponents. This linear transformation normalizes the model mixture terms and allows the input variables to be comparable in size at the onset of model fitting. The transformation is

$${x^{\prime }}_i=\frac{x_i-L_i}{\mathit{\text{Total}}-L}$$

where x′_i is the i′th pseudocomponent, x_i is the original component value, L_i is the lower constraint for the i′th component, L is the sum of lower constraints for all components, and total is the mixture total. Regression terms for developed MLR models are presented in Table V. Nanoemulsion composition and the number of microfluidization pulses were found to significantly impact nanoemulsion diameter, while percent diameter increase was found to be solely dependent upon nanoemulsion composition (Table V). In comparison with the correlation analyses, inconsistencies were observed for day 2 diameter, where transcutol was the only compositional variable that affected the diameter (Fig. 1a;Table V), and for percent diameter increase, where miglyol was observed to have no effect on the response (Fig. 1b; Table V). This is not unexpected because a correlation trend for a single parameter of interest may be difficult to detect using single linear regression, since there are multiple variations in other parameters that are unaccounted for. An accurate MLR model could not be successfully developed for drug loading. Additional statistical analyses and regression modeling methods were therefore investigated to (1) identify a predictive model for drug loading and (2) develop more accurate predictive models for nanoemulsion diameter and percent diameter increase.

Fig. 1.

Multiple linear regression models predict a day 2 nanoemulsion diameter and b percent nanoemulsion diameter increase over a 90-day period. R² and root average square error (RASE) are shown for both training and validation sets in tables beneath each model

Table V.

Model Terms, Estimates, Standard Errors, and p Values for Multiple Linear Regression (MLR) Models Used to Predict Day 2 Nanoemulsion Droplet Diameter and 90 Day % Diameter Change. Day 2 Droplet Diameter Was Found to be a Function of Both Nanoemulsion Composition and the number of Microfluidization Pulses, While 90-Day Diameter Change Was Found to Be a Function of Nanoemulsion Composition Alone

Response	Term	Estimate	Std error	p value
Day 2 droplet diameter (nm)	(Miglyol-1.9)/1.9	128.05	1.01	< 0.0001
	(Transcutol-0.2)/1.9	99.81	3.43	< 0.0001
	PCE/1.9	127.64	1.09	< 0.0001
	Microfluidization pulses (15, 45)	−6.06	0.63	< 0.0001
90-day diameter increase (%)	(Miglyol-1.9)/1.9	12.29	1.22	< 0.0001
	(Transcutol-0.2)/1.9	27.10	4.15	0.0001
	PCE/1.9	6.16	1.31	0.0011

Boosted decision tree regression models were developed for nanoemulsion diameter, percent diameter increase, and drug loading (Table VI, Supplemental Fig. 2). In this iterative modeling approach, the data are divided into training and validation sets. The training data are used to develop a single decision tree, or layer. The calculated residuals are used to fit an additional layer, which is added to the first layer to form a combined tree. Residuals are then calculated again for the combined tree and used to fit a third layer. This process is repeated until the user-defined maximum number of layers is reached or an additional layer fails to improve the validation statistic. Using boosted decision tree regression, the total sum of squares and the sum of squares attributable to each parameter are calculated. Thus, the percentage contribution of each parameter to the final model is reported. Parameters that significantly contribute to the studied output will have a larger percent contribution. In the presented work, column (parameter) contributions, R², and root average square error (RASE) terms were calculated for each boosted decision tree model (Table VI). Boosted decision tree models predicted nanoemulsion diameter and percent diameter increase more accurately than the MLR models. Further, an accurate boosted decision tree prediction model was developed for drug loading. Transcutol content and the number of microfluidization pulses made up 93.43% of the contributions for the boosted decision tree nanoemulsion diameter model. Transcutol content and PCE content made up 93.16% of the contribution for the percent diameter increase model, and transcutol content and sonication time made up 73.43% of the contribution for the drug loading model. An additional 22.61% of the drug loading model was determined by PCE content. Therefore, 96.04% of the drug loading boosted decision tree model was defined by sonication time, transcutol content, and PCE content.

Table VI.

Boosted Tree and PLS Model Statistics. Column Contributions for Boosted Tree Models, Including the Number of Splits Defined by Each Parameter, the Sum of Squares Attributed to Each Parameter, and the Proportion of the Sum of Squares Attributed to Each Parameter. Rows Highlighted in Gray were Found to have a Significant (p < 0.05) Correlation (see Table IV). Parameters with Significant Correlations Contributed the Most to the Boosted Tree Equations. R² and Root Average Square Error (RASE) Values are Also Shown for the Training and Validation Sets. Criteria Used to Select PLS Models for Nanoemulsion Day 2 Diameter, 90-Day Percent Diameter Increase, and Drug Loading, Including the Number of Factors in the Model, Root Mean Prediction Error Sum Of Squares (PRESS), and the R² and Root Square Average Error (RASE) Values for the Training and Validation Sets. PLS Models were Developed that Incorporated All Parameters, Including Main Effects, Mixture-Mixture Interactions, and Process-Process Interactions. Variable Importance for Projection (VIP) vs. Coefficient Plots were Used to Identify the Parameters that Significantly Impact the PLS Model, and These VIP Parameters Were Used to Generate a Second PLS Model for Each Studied Response. Selected PLS Models Are Italicized

	Response	Term	Number o f splits	Sum of squares	Portion	Training R2 (RASE)	Validation R2 (RASE)
Boosted tree	Day 2 diameter (nm)	Microfluidization pulses	41	795.04	0.6069	0.9823 (0.8415)	0.9804 (0.9446)
		Transcutol	21	428.84	0.3274
		PCE	23	85.08	0.0649
		Sonication Time	6	0.86	0.0007
		Miglyol	5	0.085	0.0001
	90-day diameter increase (%)	PCE	24	221.48	0.5565	0.9168 (1.0409)	0.9621 (0.8412)
		Transcutol	17	149.30	0.3751
		Sonication Time	57	11.34	0.0285
		Miglyol	1	9.68	0.0243
		Microfluidization Pulses	89	6.20	0.0156
	Drug loading (%)	Transcutol	11	78.77	0.4094	0.8298 (1.4505)	0.8361 (2.2110)
		Sonication Time	16	62.50	0.3249
		PCE	15	43.51	0.2261
		Microfluidization Pulses	24	7.62	0.0396
		Miglyol	0	0	0
PLS	Response	Parameter	No. of Factors	PRESS	Training R2	Validation R2	Training RASE	Validation RASE
	Day 2 diameter	All (9)	3	0.3714	0.9371	0.8679	1.5867	2.4539
		VIP (4)	4	0.2206	0.9577	0.9534	1.3004	1.4574
	90-day diameter increase (%)	All (9)	2	0.3500	0.8171	0.9067	1.5433	1.3192
		VIP (5)	2	0.3558	0.7804	0.9036	1.6912	1.3412
	Drug loading (%)	All (9)	2	1.0126	0.5336	0.5361	2.4014	3.7191
		VIP (6)	6	0.8970	0.6804	0.6360	1.9881	3.2945

The boosted decision tree models confirmed the findings observed using correlation plots. The two parameters with the highest percentage contribution corresponded to the statistically significant correlation trends (see gray highlights, Table VI). Further, the boosted decision tree models were more accurate than MLR models in predicting diameter, percent diameter increase, and drug loading. It appeared that MLR may not be the most accurate approach for predicting these three CQAs or for understanding the relationships between nanoemulsion composition, manufacturing processes, and the CQAs of interest. However, boosted decision tree modeling does not allow for the study of interaction variables. To study possible interactions and to further confirm the accuracy of the correlation plots and boosted decision tree models, partial least squares (PLS) models were developed for the CQAs of nanoemulsion diameter, percent diameter increase, and drug loading (Fig. 2; Table VI).

Fig. 2.

PLS models developed to predict nanoemulsion diameter 2 days (48 h) after production (panels a–c), percent diameter increase over 90 days (panels d–f), and drug loading (panels g–i). Data shown include actual vs. predicted plots for the training data (panels a, d, g) and validation data (panels b, e, h) and variable importance for projection (VIP) vs. coefficient plots (panels c, f, i). VIP variables are parameters that contribute significantly to the PLS model and have a VIP value of greater than 0.8 (parameters above the dotted line)

Two sequential PLS models were developed for each CQA of interest (Table VI). The first PLS model included nine terms (5 main effects, 3 mixture-mixture interactions, and 1 process-process interaction). The second PLS model removed terms that were below a set threshold for variable importance for projection (VIP) scores. PLS models with the smaller prediction error sum of squares (PRESS), larger R², and smaller root average square error (RASE) were selected (Table VI). Actual vs. predicted plots for training and validation sets, as well as VIP vs. coefficient plots, are shown for these selected PLS models (Fig. 2). Transcutol content and the number of microfluidization pulses were the most significant contributors to the nanoemulsion diameter PLS model (Fig. 2c). These findings are consistent with those from the correlation plots (Supplemental Fig. A–E) and boosted decision tree models (Table VI).

VIP variables for the percent diameter increase PLS model included miglyol, transcutol, PCE, miglyol*PCE, and miglyol*transcutol (Fig. 2f). As with the MLR and boosted tree models, percent diameter increase is dependent solely upon nanoemulsion composition. There was a discrepancy between the boosted decision tree and PLS models for percent diameter increase, as miglyol was a VIP variable in the PLS model, but only contributed to 2.43% of the boosted decision tree model. It is worth noting that transcutol and PCE contribute more significantly to the PLS model than miglyol, based upon the higher VIP scores and coefficient values for these variables (Fig. 2f).

VIP variables for the drug loading PLS model include transcutol, sonication time, and PCE content, as well as the interaction terms miglyol*transcutol and sonication time*microfluidization pulses (Fig. 2i). This is consistent with the boosted decision tree model (Table VI), in which transcutol, sonication time, and PCE make up 96.04% of the model.

Select Nanoemulsions for Further Evaluation

Three of the 16 nanoemulsion formulations (nanoemulsions 3, 4, and 16) met all CQA specifications (Table III). Of these three formulations, nanoemulsion 4 had the highest drug loading (84.68%). For this reason, nanoemulsion 4 was selected to proceed to phase 2 CQA testing. Depending upon imaging needs, it may not be necessary to include PCE in the formulation. Therefore, it is desirable to have a stable PCE-free formulation. Six of the 16 nanoemulsions developed did not contain PCE (nanoemulsions 1, 5, 6, 10, 11, and 12, Table III). Nanoemulsions 11 and 12 failed to meet CQA specifications for drug loading and 90-day percent diameter increase. The remaining four formulations only failed to meet the CQA specification of 90-day percent diameter increase. Of these four formulations, nanoemulsion 10 was the closest to meeting this CQA specification (12.3% diameter increase). Therefore, nanoemulsion 10 was also selected to proceed to phase 2 CQA testing.

The two nanoemulsions selected for phase 2 CQA testing were produced two additional times so that each selected formulation was replicated in triplicate. Nanoemulsion 4 represents an optimal formulation of PCE nanoemulsion, while nanoemulsion 10 represents an optimal formulation of PCE-free nanoemulsion. The first replicate was drug free and the second replicate contained celecoxib. Droplet diameter distribution measured 2 days (48 h) after production was consistent between nanoemulsion 4 and its two replicates (Fig. 3a). Additionally, there was no significant change in average droplet diameter or PDI of the replicated nanoemulsions over a 30-day period (Supplemental Fig. 3A–B). All select nanoemulsion replicates met phase 2 CQA specifications of filtration stability (Supplemental Fig. 3C), centrifugation stability (Supplemental Fig. 4), and serum stability (Supplemental Fig. 5). Further, all select nanoemulsion replicates met phase 2 CQA specifications for the thermal cycling study (Fig. 3b) and sterility (Fig. 3c, d).

Fig. 3.

a Size distribution overlays of the selected PCE nanoemulsion (NE-4, NE-4 replicate 1, and NE-4 replicate 2). b Thermal cycling study on selected PCE and PCE-free nanoemulsions. Panels c, d Sterility test. Select nanoemulsions were diluted 1:80 v/v and incubated for 14 days in c trypcase soy broth at 25°C or d fluid thioglycolate medium at 35°C. pH measurements were taken before and after the incubation period

DISCUSSION

CQAs were divided into phase 1 and phase 2 to enable rapid and efficient identification of the best formulations. Phase 2 CQAs were included to provide the extensive evaluation necessary to confirm that nanoemulsions would maintain colloidal stability under in vitro pharmacological testing (filtration, centrifugation, incubation in serum, sterility) and transportation to collaborators for in vivo testing (thermal cycling study). Specifically, filtration is a standard processing step used to sterilize nanoemulsions prior to all in vitro biological testing and is necessary before using nanoemulsions in animal studies. Centrifugation and serum studies are designed to mimic stresses that nanoemulsions will experience during cell culture studies while thermal cycling is designed to ensure nanoemulsions can withstand temperature changes during transportation.

FMECA is rarely used to study nanoemulsions and has never been used to evaluate perfluorocarbon nanoemulsions. In the presented work, FMECA allowed us to systematically rank failure methods and identify the process parameters that were most likely to impact nanoemulsion CQAs. This provided a rational approach that directed us to construct an experimental design which allowed us to understand the impact of nanoemulsion composition and process parameters on nanoemulsion stability (90-day diameter increase) and drug loading in an efficient manner. With the production of only 16 nanoemulsions, we were able to save significant time and resources. The FMECA provided accurate predictions as to which failure modes presented the highest risk. For example, the number of microfluidization pulses received the highest RPN score for the CQA specification of nanoemulsion diameter. Correlation plot, MLR, boosted decision tree, and PLS models all indicate that the number of microfluidization pulses significantly impacts nanoemulsion diameter.

Other high RPN scores were associated with the 90-day diameter increase and drug loading specifications. Accordingly, more nanoemulsions failed to meet these two CQA specifications than any other specification. Specifically, high solubilizer content impacting the 90-day diameter increase specification received an RPN score of 60 (Table II). This was verified, as a significant positive correlation trend (R² = 0.4970, p value 0.0023) was observed between transcutol content and the 90-day percent diameter increase (Supplemental Fig. 1G), and transcutol content was a significant variable in all three developed percent diameter change regression models. Further, the interaction between transcutol and miglyol was found to have a high VIP score in the percent diameter increase PLS model. Long sonication time leading to drug precipitation also received an RPN score of 60 (Table II). This was also verified, as a significant negative correlation trend (R² = 0.2668, p value 0.0405) was observed between drug loading and sonication time (Supplemental Fig. 1O), and sonication time was a significant variable in the drug loading-boosted decision tree and PLS models. Additionally, an interaction between the sonication time and the number of microfluidization pulses was found to have a high VIP score in the drug loading PLS model.

It is worth noting that the highest drug loading was 90.84% (nanoemulsion 2, Table III). This formulation had the highest studied transcutol content, the lowest studied number of microfluidization pulses, and did not undergo sonication. One of the limitations of this work is that the impact of cold temperature on drug precipitation during microfluidization and sonication was not studied, though these failure methods received RPN scores of 60 (Table II). It is possible that cold temperature, high shear force during microfluidization, or high frequency during sonication may cause a portion of the drug to precipitate during nanoemulsion production. This will be investigated in greater detail in future studies. In the presented work, stable nanoemulsion candidates were identified that satisfied the CQA drug loading specification, and these nanoemulsions also passed all phase 2 CQA specifications (Fig. 3, Supplemental Figs. 3–5). However, it is advisable to consider the impact of processing parameters on drug loading, particularly when using sonication during nanoemulsification.

Increasing nanoemulsion solubilizer (transcutol) content will increase drug loading, but this comes at the expense of decreased nanoemulsion stability in the form of droplet diameter growth over time. This instability may be mitigated by inclusion of PCE, as there is a significant negative correlation trend (R² = 0.4357, p value 0.0054) between the amount of PCE and the percent diameter increase (Supplemental Fig. 1H). PCE content was found to be a significant variable in all three regression models and was found to be the most significant contributor to the boosted decision tree regression model. Therefore, inclusion of PCE into the nanoemulsion may improve its stability, thus extending its shelf life.

MLR, boosted decision tree, and PLS regression models were developed for nanoemulsion diameter and percent diameter increase. An MLR model could not be developed for drug loading, but boosted decision tree and PLS regression models were developed for drug loading. These models are all summarized in Table VII. Boosted decision tree models were the most accurate predictors of nanoemulsion diameter, percent diameter increase, and drug loading, while MLR was the least accurate predictor of these three CQA specifications. MLR assumes that all input parameters are independent of one another. This assumption is not the case in the presented work, as the mixture variables are not independent. PLS, and on the other hand, is capable of dealing with co-linearities. This is a potential explanation for why the PLS models have better predictive capabilities than the MLR models for all three studied CQA specifications. Finally, an advantage of boosted decision tree models is that this type of model is recursive. The residuals of existing trees are used to construct subsequent trees. This process enables the identification of the parameters that are most significant. This is a potential explanation as to why the boosted decision tree regression models were the most accurate predictors of the studied CQA specifications in the presented work. The modeling approaches presented here can be applied to other nanomedicine product development.

Table VII.

R² and Root Average Square Error (RASE) Values for MLR, Boosted Tree, and PLS Models Used to Predict the Responses of Nanoemulsion Day 2 Diameter, 90-Day Percent Diameter Increase, and Percent Drug Loading. Boosted Tree Modeling Resulted in the Most Accurate Prediction Equations for All Three Studied Responses

Response	Training/validation	Prediction equation	R 2	RASE	Frequency
Day 2 diameter (nm)	Training	MLR	0.9464	1.4647	12
		Boosted tree	0.9823	0.8415	12
		PLS	0.9577	1.3004	12
	Validation	MLR	0.9181	1.9319	4
		Boosted tree	0.9804	0.9446	4
		PLS	0.9534	1.4574	4
90-day diameter increase (%)	Training	MLR	0.7288	1.8792	12
		Boosted tree	0.9168	1.0409	12
		PLS	0.8171	1.5433	12
	Validation	MLR	0.9030	1.3453	4
		Boosted tree	0.9621	0.8412	4
		PLS	0.9067	1.3192	4
Drug loading (%)	Training	Boosted tree	0.8298	1.4505	12
		PLS	0.6804	1.9881	12
	Validation	Boosted tree	0.8361	2.2110	4
		PLS	0.6360	3.2945	4

Droplet diameter distribution measured 2 days (48 h) after production was consistent between nanoemulsion 4 and its two replicates (Fig. 3a). Additionally, there was no significant change in average droplet diameter or PDI of the replicated nanoemulsions over a 30-day period (Supplemental Fig. 3A–B). All select nanoemulsion replicates met phase 2 CQA specifications of filtration stability (Supplemental Fig. 3C), centrifugation stability (Supplemental Fig. 4), and serum stability (Supplemental Fig. 5). Further, all select nanoemulsion replicates met phase 2 CQA specifications for the thermal cycling study (Fig. 3b) and sterility (Fig. 3c, d).

Select nanoemulsion replicates were produced by three independent researchers. Nanoemulsion droplet diameter and PDI were consistent among replicates (Fig. 3a). Further, nanoemulsion droplet diameter and PDI in response to filtration, centrifugation, serum, and thermal cycling stability tests were consistent among replicates (Fig. 3b, Supplemental Figs. 3C, 4–5). Measured pH values as part of sterility testing were also consistent among nanoemulsion replicates (Fig. 3c,d). The consistency among the replicates demonstrates the reproducibility of these formulations. The fact that all selected nanoemulsion replicates demonstrate robust colloidal stability under all CQA specifications indicates that these nanoemulsions can exhibit a long shelf life and withstand a variety of environmental changes and physical stresses during storage and long-distance transportation. Nanoemulsions can also survive the physical stresses experienced during in vitro cell culture studies, which require centrifugation and serum contained biological media exposure. Sterility protocols strictly followed guidelines from US Pharmacopeia 34 microbiological test/71 sterility test (34). This helped to make the formulations closer to the pharmaceutical industry standards, which may ease the translation from pre-clinical to clinical experiments. It is also worth noting that the select nanoemulsion formulations which met CQA phase 1 specifications also met all phase 2 CQA specifications. This indicates that phase 1 CQA specifications are critical and appropriate to select the stable formulations. This saved time and resources and enabled rapid identification of stable formulations by avoiding the performance of all CQA phase 2 quality control tests on all formulations. This presented quality by design approach provided us with a deeper understanding of the processes that impact nanoemulsion drug loading and droplet stability and allowed us to rationally identify stable nanoemulsion candidates suitable for further formulation development.

CONCLUSIONS

The presented QbD approach applied to nanoemulsion development assisted in loading ten times the amount of celecoxib into a formulation compared to previous studies. This methodology also enhanced product and process understanding. Specifically, through construction of several types of regression models, we found that the presence of perfluorocarbon oil can increase long-term colloidal stability, whereas the presence of solubilizer can increase drug loading but decrease long-term colloidal stability. We believe these methods can be useful in development of various nanomedicines to produce higher-quality products.