Ronald D. Snee, PhD, is president of Snee Associates, 10 Creek Crossing, Newark, DE, 19711 tel: 610.213.5595, fax 302.369.9971.

Ronald D. Snee

To help pharmaceutical and biotech companies improve their operations, FDA has been promoting the use of quality by design (QbD) (1, 2). Much has been written about the QbD concept of a “design space” from a manufacturing process perspective. Little has been published, however, about using QbD in formulation development, including the development of formulation design spaces. 

	Formulation studies typically involve the optimization of multiple ingredients including the API, lubricants, binders, and disintegrants. Such optimization is a difficult challenge when a large number of components are involved, which is typically the case. 

	An additional challenge is that, in many formulations, the amounts of components in the formulation must add up to 100% or some fixed amount. A pharmaceutical formulation, for example, might consist of: 15% API, 35% lactose, 45% microcrystalline cellulose (MCC), 4% Starch, and 1% magnesium stearate. The percentages of the ingredients add up to 100%; thus, when levels of one or more components are changed (i.e., increased or decreased), the percentages of one or more of the remaining components must be changed (i.e., decreased or increased by a corresponding amount). 

	The statistical design-of-experiments approach, DoE, which is at the core of QbD, has the flexibility to deal with multiple components that involve constraints of many different types. In the statistical approach, a series of formulations is created and tested in a planned sequence. The levels of the candidate components are varied in a blending design and the performance of the formulations is measured. A model is fit to the data and critical components, and then blending characteristics are identified. 

	Response surface contours are examined graphically and analytically to determine the design space, the region of the best values of the response that meet specifications. Additional confirmatory formulations are typically tested to verify the model predictions. The statistical approach has many important benefits, as discussed by Montgomery (3), Snee and Hoerl (4), and many others. As a result, formulation understanding is greatly increased as well as the probability of successful formulation development (4-7). This article discusses some of the issues involved, and outlines a strategy for dealing with them. To clarify abbreviations in the text, MCC refers to Avicel PH105; CP is Hiviswako 104; and PVP is Providone K90.

	Use of DoE speeds up formulation development, but more is needed and more is possible if one thinks strategically about formulation development. 

	Over the years, it has been recognized that experimentation is more effective when it is approached with a strategy in mind. To be effective, any strategy must recognize that the design, or sequence of designs, should match the experimental environment; that experimentation is sequential; and that the DoE tools must be embedded in the strategy and linked and sequenced to guide the experimenter. 

	This experience leads to the following principles, which can enhance experimental strategies:

		Plan ahead; define the series of experiments needed to satisfy the objective of the program.

		At the beginning, include (or at least consider) all factors (Xs) that may possibly be important. Recall the Pareto effect, and the fact that most of the variation will be caused by a small subset of the factors so that, as one moves through the experimentation, the important factors will be discovered and tested further in later experiments.

		Don’t spend all resources on a single experiment: an issue is rarely resolved in a single experiment.

	A strategy that uses these principles was developed at DuPont in the 1960s, and offered in public workshops beginning in the 1970s. In the case of formulation, this strategy identifies two experimental environments: screening and optimization. The objective of each of the phases and the designs used is summarized in 

	Briefly stated, the screening phase explores the effects of a large number of ingredients (components) with the objective of identifying a smaller number of components to study further in optimization experiments. Additional screening experiments involving additional factors may be needed when the results of the initial screening experiments are not promising. On several occasions, the screening experiment solves the problem.

	In the optimization phase, a predictive model is developed for the system that can be used to find useful operating conditions using response surface contour plots and perhaps mathematical optimization. The result is the design space.

	The end result of each of these sequences is a completed project. There is no guarantee of success in a given instance, only knowledge that the DoE strategy will “raise your batting average” (8). The strategy used depends on the experimental environment, which includes the objectives of the experimental program, the nature of the components (Xs) and responses (Ys), resources available, quality of the information to be developed, and the theory available to guide the experiment design and analysis. A careful diagnosis of the experimental environment along these lines can have a major effect on the success of the experimental program.

(9) presented a pharmaceutical tablet study that investigated the formulation involving the compound paracetamol, which was known to have poor flowability (a measure of  how well the materials flow through the tableting equipment) and compressibility properties. The study involved seven ingredients, varied over the ranges that are shown in 

	Nine responses were measured. The focus of the following discussion will be for illustrative purposes, the repose angle response (

). A 19-blend extreme vertices design shown in 

 was used to design the formulations to be tested. This design was selected using the D-Optimality criterion. As in any screening experiment, it is necessary to know what components are most important as measured by their effect (positive or negative) on the response. This information will enable one to answer the following questions:

		Can we select a formulation based on the results of this particular experiment?

		Is additional experimentation needed? If so, which components should be the focus of future experimentation?

	It is not unusual for a screening experiment to provide the information needed to choose a desirable formulation. Such was the case in this particular study. If additional experimentation is needed, the screening experiment results provide a firm basis for designing the optimization experiment. In this case, a seven-term linear blending model of the following form was developed (

	Y = b1X1 + b2X2 + b3X3 + b4x4 + b5x5 + b6X6 + b7X7  

	where Y represents the response of interest, in this case Y=Repose Angle; X1, X2 …. X7 represent the levels of the components, and the bs are regression coefficients associated with the Xs. 

 for each of the responses and used the equations to develop an optimal formulation, which, when tested, produced measured responses that were very close to those predicted by the linear blending model. Thus, in this case, additional experimentation was not needed.

	The principal output of a screening experiment is the component effects. The component effects shown in 

 for the repose angle response were calculated using methods proposed by Cox (10). The linear blending model defined by 

 fit the data with an adjusted R-Square value of 82%, a respectable value for a screening experiment. The overall model gave a statistically significant fit to the data (p=0.000).

, note that the Microcel and Aerosil have a significant effect on the tablet repose angle. The Aerosil effect is negative and much larger than the Microcel effect, which is positive. These effects are seen in the component prediction profile plots in 

 the effect lines for all the components except Microcel and Aerosil have a very small slope, indicating no effect.

	When evaluating component effects to better understand formulation systems, the following are recommended:

		Evaluate the component effects using the Cox effect directions to assess the nature and magnitude of the component effects as shown in 

		Study the component effects to identify components that may have equal values indicating similar blending behavior.

		Assess whether similar blending of these components is supported by subject-matter science and business knowledge. Components with effects that are not statistically significant identify components that have no effect. The response to finding components that have no effect will depend on the objectives of the experiment. Options include setting the component with no effect at any desirable level within the range studied. If zero is at or near the lower end of the range, one can consider removing the component from the formulation. In all cases, the selected action should take into account subject-matter science and business knowledge (8).

 (11) present an optimization case study focused on the development of a three-component sustained release tablet of chlorpheniramine maleate. The objective of the study was to find a formulation with a release rate >30 units. The ranges of the three components shown in 

 suggests the use of an extreme vertices response surface design. 

	These component ranges produce an experimental region that has six vertices (

). The region is irregular, and there are two pairs of vertices that are close together (points 2 and 3 and 4 and 5, Table IV). 

 shows the six vertices, the six-edge centroids, and the overall centroid, as well as the response surface contours that will be discussed later.

	In a formulations experiment design made up of the six vertices, the six-edge centroids (including the two short edges) and the overall centroid, which was tested three times for a total of 15 blends, were evaluated. The design and release rate data are shown in 

. In this table, five of the design blends (5, 6, 10, 11, 12) have release rates > 30. This provides the scientist’s assurance that the objectives of the study will be met, even before any statistical modeling and analysis are done.

	Plots of predicted release rate versus component levels for each component are shown in 

 (sometimes referred as the prediction profile). In this figure, CP has a negative effect while MCC and PVP have positive effects. The component linear effects as measured along the Cox axes show that the CP effect (-16.5 units) is dominant as compared to the MCC and PVP effects which are 7.9 and 9.5 units, respectively. In formulation optimization studies, the quadratic model is the model of choice, particularly at the beginning of the analysis process. The three-component quadratic blending model of the form is shown in 

	Y = b1X1 + b2X2 + b3X3 + b12X1X2 + b13X1X3 + b23X2X3

	Where Y represents the response of interest. In this study case, Y=Release Rate; X1, X2, and X3 represent the levels of the components, and bs are model regression coefficients.

) gives a good fit to the data as judged by the adjusted R-square value of 94%. As expected, there was some significant curvilinear blending involving MCC and CP and PVP and CP. 

	The centroid blend was evaluated three times. These replicate blends enable us to test the lack-of-fit of the model. This test was not significant (p=0.638), providing further indication of the adequacy of the fit of the quadratic model. 

	The next step is to construct the response surface contours to identify the formulations that will produce release rates > 30 units. In 

, over the design space, levels of CP < approximately 0.25 will produce release rates > 30. 

 shows contours for release rates of 18, 20, 24, 28, and 32; associated with decreasing CP moving left to right. The curved contour lines result from the curvilinear blending noted above. Components that blend linearly produce straight line contours. The fit of the model was checked by running two confirmation experiments (

, blends 16 and 17), which produced “observed” and “predicted” release rates of 32.7 vs 30.0 and 20.7 vs 20.0, respectively. These accurate predictions added further evidence of the adequacy of the model.

	The right data in the right quantity at the right time

	The strategy of formulation development described in this article enables formulation scientists to speed up formulation development by getting the right data in the right amount and the right time. In the process, the formulation scientist’s “batting average” is raised (10). 

	Proven techniques, and available software

	The strategy and methodology outlined here have been used successfully, and tested over time. They work, and there is software available to ease the computational burden, as well as to create graphics that aid in analyzing the data and visualizating the results. In short, the approach seems worthy of consideration by formulation scientists. 

	1. ICH, Harmonised Tripartite Guideline: Pharmaceutical Development, Q8, Current Step 4 Version, Nov. 10, 2005.
	2. R. Snee, “

Building a Framework for Quality by Design

” pharmtech.com, October 2, 2009.
	3. D.C. Montgomery, Design and Analysis of Experiments, 8th Edition (John Wiley and Sons, New York, NY, 2013.)
	4. R.D. Snee, and R. W. Hoerl, Strategies for Formulations Development: A Step-by-Step Guide using JMP (SAS Press, Cary, NC, 2016.)
	5. R.D. Snee, “Understanding Formulation Studies,” Technometrics, 37(1), pp.131-132, 1995.
	6. R.D. Snee, “Understanding Formulation Systems-A Six Sigma Approach,”Quality Engineering, 23 (3), pp. 278-286, 2011.
	7. R.D. Snee and G. F. Piepel “Assessing Component Effects in Formulation Systems,” Quality Engineering, 25(1), pp.46-53, 2013.
	8. R.D. Snee, “Raising Your Batting Average: Remember the Importance of Sequence in Experimentation,”Quality Progress, December 2009, pp. 64-68. 
	9. T. Martinello, et al., Int. J. Pharmaceutics, 3(22), 87-95, 2006.
	10. D.R. Cox, “A note on polynomial response functions for mixtures,” Biometrika, 58(1), pp. 155-159 . 1971 .
	11. M Hirata, et al., “Formulation Optimization of Sustained Release Tablet of Chlorpheniramine Maleate by Means of Extreme Vertices Design and Simultaneous Optimization Techniques,” Chemical Pharmaceutical Bulletin, 40(3), pp.741-746, 1992.


	Vol. 41, No. 9
	September 2017
	Pages: s4-s11

	When referring to this article, please cite it as R. Snee, "Speeding Up Formulation Development," 

Pharmaceutical Technology APIs, Excipients, & Manufacturing 2017 

	Ronald Snee is principal of Snee Associates, LLC, based in Newark, Delaware. He worked at DuPont for 24 years, and then as a consultant for companies that include Tunnell Consulting. He is an adjunct professor in the pharmaceutical programs at Temple and Rutgers Universities. He received his BA from Washington and Jefferson College and his MS and PhD degrees from Rutgers University, and can be reached at Ron@SneeAssociates.com.

Articles

QbD principles and strategic thinking can reduce the time required to optimize formulation. 

Speeding Up Formulation Development

	Process control is and has been an important aspect of manufacturing, test method operations, and other bio/pharmaceutical processes. FDA’s process validation guidance (1) calls for continued process verification (CPV) as Stage 3 of the guidance. Control charts are important tools for assessing, monitoring, and maintaining process control. Sometimes, organizations resist using control charts due to the perception that too many false out-of-control (OOC) signals are produced. False OOC signals can happen, particularly when the wrong model is used to construct the control chart. This article describes a fix that works in many situations.

	A common example of the use of control charts is monitoring the production of a batch of tablets. Depending on the tablet characteristic (weight, hardness, dissolution, etc.) samples of 3-20 tablets are taken by the process operator periodically (e.g., every 30 minutes) during the production of the batch. Using an Xbar control chart, the averages of the samples are plotted versus the order in which the samples were taken. The control limits are computed from within the sample standard deviation.

	An example of the resulting chart is shown in 

. The variation within the control limits, called common-cause variation, represents the variation inherent in the process. Values (points) outside the control limits are said to be due to special-cause variation. The control chart helps detect and reduce the special cause variation, thereby improving the performance of the process. 

	When using control charts, typically two types of non-random patterns are observed:

		Sample results outside of the control limits (typically set at the process average ± 3 standard deviations). Such events are referred to as OOC signals.

		Non-random patterns such as trends, drifts, and shifts, up and down. These events are referred to as out-of-trend (OOT) signals and are detected using the Western Electric rules (2).

	The process is said to be in a state of statistical control if there are no points outside of the control limits and no non-random patterns (i.e., cycles and shifts) are detected. The OOC and OOT signals provide a basis for human intervention, to identify and eliminate sources of variation, which may require some work (problem solving) to identify. This example demonstrates one of the principles of statistical thinking; namely, understanding and reducing variation provides an opportunity for improvement (3).

	Both OOC and OOT signals are based on statistical tests of significance and are included in many statistical software packages (4). OOC and OOT signals can trigger an investigation that can be a wasted effort and organizational anxiety if the signals are indeed false positives. The control chart limits are typically based on ± three standard deviations to provide a good balance between detecting false alarms with missing important signals as recommended by Shewhart (5).

	Example 1 shows a process in which 27 samples of 10 tablets each were sampled at hourly intervals across the production of the product and the weight measured on each tablet. 

is the resulting Xbar chart with the limits based on the within sample tablet-to-tablet standard deviation.

	The chart shows no major trends but six OOC signals and one OOT signal are observed. In this chart and others in this article, “1” denotes a point outside the control limits. The numbers 2 through 8 denote different types of OOT points as identified by the Western Electric rules (2).

, three of the OOC signals are on the low side and three are on the high side. The one OOT signal is associated with a potential process shift (2 out of 3 points more than 2 standard deviations from center line) on one side of the control limits. Ignoring the control limits for a moment, the process appears stable with no major shifts and trends. Is this process really out of statistical control? Is an investigation required? Or are these false positive signals? To answer these questions, it is helpful to review how the control limits should be set.

	Variation in the product measurements has three principle sources: manufacturing operations, sampling, and testing. The control limits should be based on short-term manufacturing, sampling, and testing variation as it is the long-term stability of the manufacturing process that is being assessed by the control chart.

	Unfortunately, many times the control limits are based on testing variation alone, such as the tablet-to-tablet variation within the sample in 

. When the short-term manufacturing and sampling variation are real--as they almost always are--and the limits are based on only testing variation, the limits will be too narrow and false OOC and OOT signals will result. The control limits should be based on all the variation inherent in the averages being plotted.

	Returning to Example 1, the easy fix is to compute the averages of the different samples and plot the averages on an individuals-moving range (I-MR) chart. The variation of the individual averages is due to short-term manufacturing variation, sampling variation, and test variation. The I-MR chart bases the control limits on the moving range of the averages, which reflects the short-term manufacturing, sampling, and test variation. In essence, the I-MR chart defines average-to-average variation as the variation inherent in the averages and any variation beyond this as special cause variation to be identified and assessed. 

 shows the I-MR chart for the data in Example 1. In 

, there are no points out of control limits and there is no OOT signals.

	Short-term manufacturing and sampling variation have not been taken into account when it is observed that the process appears fairly stable and there are approximately the same number of OOC signals on the high side and the low side. Such a pattern commonly occurs when the control limits are based on an under estimate of the common-cause variation of the averages plotted in the chart.

	It is also important to keep in mind that control charts can produce OOC signals when the process is stable and the correct sources of variation have been used to construct the control chart. Setting control limits at ± three standard deviations from the process average and assuming the distribution is normal will, in the long run, produce a 0.3% false positive rate.

Real OOC and OOT events are still detected

	The approach of assessing process control by constructing an I-MR chart of the sample averages raises the question: Does this approach limit the ability to detect significant (real) non-random patters such as shifts and trends? The following examples show that this is not the case. 

) is for tablet weight of 36 samples of 3 tablets per sample with control limits based on within-sample standard deviation. Figure 2A shows nine OOC signals; some high and some low. Some shifts up and down and a small negative trend (OOT signals) can be seen. 

 shows the I-MR chart of the sample averages. Here, there is now only one sample outside of controls limits. The shifts and negative trend (OOT signals) are still apparent and found to be statistically significant. 

	In Example 3, tablet thickness (53 samples of 10 tablets per sample), the Xbar control chart (

, control limits based on the within-sample tablet-to-tablet standard deviation) identifies five OOC and five OOT signals.  

) shows no OOC signals, but a level shift is detected. This indicates that an investigation of the level shift should be considered, but there is no need to worry about any other OOC signals.

	An important question is whether the sample size has an effect on the frequency of OOC and OOT signals. While larger samples will reduce the variation in the sample averages, it doesn’t reduce the variation due to the manufacturing process or the act of sampling. Use of the I-MR chart of the sample averages is still a useful approach.

	Don Wheeler points out that the I-MR chart is appropriate when the successive values plotted on the chart are logically comparable and the ranges between the values reflect the local, short-term, routine variation inherent in the values (6). In effect, the process short-term variation is used to create control limits to detect the presence of long-term variation, which indicates process instability at the points in time where the OOC signals appeared.

Understand the sources of variation reflected in the control chart

	Control charts that are properly constructed and maintained provide a useful method for monitoring a process. Organizations must understand the sources of variation that are contributing to the variation in the data being evaluated by using the control chart. The proper use of any control chart requires careful thought about the sources of variation that would be considered to be common cause and the sources we would consider to be special cause and want to detect. This will ensure that we only detect the signals that we are interested in.

	1. FDA, Guidance for Industry, Process Validation: General Principles and Practice, (Rockville, Md., 2011).

	2. D.C. Montgomery, Introduction to Statistical Quality Control, 7th Edition (John Wiley and Sons, New York, 2013)

	3. R. W. Hoerl and R. D. Snee, Statistical Thinking-Improving Business Performance, (John Wiley and Sons, Hoboken, NJ, 212).

	4. L. D. Torbeck. “OOS, OOT, OOC and OOSC”, Pharm Tech, 34 (10), 46-47.

	5. W. A. Shewhart, Economic Control of Quality of Manufactured Product (Van Nostrand Co., New York, 1931).

	6. D. J. Wheeler, Guide to Data Analysis (SPC Press, Knoxville, Tenn., 2005). 

Control charts that are properly constructed and maintained prevent false out-of-control signals and provide a useful method for monitoring a process.