Brain storm!

As usual, I can't let this problem go without digging down to bedrock.  I was inspired by the discussion on linkedin, especially an idea that Ina put into my head about the magnitude of the parameter estimates.  I couldn't shake the thought and then I came across this article from the folks at Genentech.  If you get the chance, read it!  Here's my attempt to apply my understanding of their approach.

In my previous discussion about the aggregate present in the CaptoAdhere eluate was driven by the statistically significant pairwise interaction of load*load conductivity.  Now, the magnitude of the effect was only 0.02% (rounding).  Now consider this: the root mean square error (RMSE) is 0.09%!  So, even though we have a statistically significant model the expected variation is lost within the experimental variation.  Based upon this, the practical significance to the quality attribute variation is lost within the expected variation due to the process and the analytics.  As a result, we could argue the load and load conductivity are non-critical process parameters.  This can be demonstrated graphically within JMP using the prediction profiler.  What do you think?

Summary of Fit to Pairwise Interaction of Load*Load Conductivity

RSquare	0.73365
RSquare Adj	0.700356
Root Mean Square Error	0.000893
Mean of Response	0.00765
Observations (or Sum Wgts)	10

Predicted Variation to Aggregate 

Acceptable Operating Range for the CPPs

I came across this white paper from the Product Quality Research Institute (PQRI) awhile back but only just revisited it to work on a resolution of how to justify whether a process parameter was critical or non-critical.  In their paper, the authors state that a "Critical Process Parameter  (CPP) is a process input that, when varied beyond a limited range, has a direct and significant influence on a Critical Quality Attribute (CQA).  Failure to stay within the defined range of the CPP leads to a high likelihood of failing to conform to a CQA."

The challenge is the interpretation of a "limited range".  Limited relative to what, infinity?  The normal operating range (NOR) "reflects the range over which a parameter can vary without impacting critical quality attributes".  From the previous discussions, we saw that the aggregate profile doesn't change significantly within the NOR that I specified in the model AND the variation within the host cell protein data was going to be difficult to identify given the precision of the assay.  Within the context of the discussion, the data are consistent with current thinking about definitions to process parameters.  However, I haven't defined what is meant with "limited range".

Taking a lead from the PQRI white paper, the goal of the development team is to define the limits of that "limited range"in relation to the NOR.  These ranges for the process parameters become the proven acceptable ranges (PAR), or acceptable operating range (AOR).  The development team has to define these AORs/PARs using laboratory-scale models that reflect the process at manufacturing scale.  In the worked example, there is limited manufacturing scale production data that may be used to link back to the laboratory-scale results.  In this paper, the author proposes using a one-sided specification limit to define the AOR.  Let's say, for the sake of argument, that the release specifications for the drug substance are as follows:
Aggregate: <0.95%
[HCP]: < 20 ppm

Within JMP, the prediction profiles can be output and used to generate an operating space.  In this case, the non-colored regions are allowed and the colored regions are the points at which the CQA(s) would be at risk.  Within the region bounded as less than 200 g/L of load and a load conductivity of 10-30 mS/cm, the process is expected to yield a product with CQAs that are within the release limits.

Relationship of Release Specifications to the Studied Ranges

The resulting limits would attempt to be filed:

Load Conductivity NOR:  19.5 - 20.5 mS/cm

Load Conductivity AOR:  10 - 30 mS/cm

Load amount NOR: 190-210 g/L

Load amount AOR: <220 g/L

Now for the pressing question: which limits do you file for your process description?  Keep in mind that the limits you file in the process description are the ones to which you, as the sponsor, are committed to sustaining during manufacturing.  The implication is that if you exceed the limits within the process description, the expectation is that the batch would be rejected.  

Non-critical versus Critical, continued

In the last post, the aggregate percentage was found to be dependent on the process parameters; however, when these were tightly maintained the product quality variation was negligible.  As it happens, the ELISA that was developed specifically for the cell line shows a dependence on the process parameters as well.  The latter of these are at the edge of statistical significance; however, a conservative approach was taken when building the model and they were included in the final analysis.
Parameter Estimates for CHO Host Cell Protein (HCP)

Term		Estimate	Std Error	t Ratio	Prob>|t|
Intercept		16.175041	0.876267	18.46	<.0001*
Load(93,312)		8.1260841	1.155147	7.03	0.0004*
Load cond(10,30)		-2.534245	1.18887	-2.13	0.0770
Load*Load cond		-3.252868	1.312414	-2.48	0.0479*

The simulator was then invoked using the same variances for the load and load conductivity that were applied in the previous post.  The ELISA assay is robust; however, the variance is a bit less.  Based upon the model fit, the root mean square error is about 2.75 ppm.  This is a pretty reasonable estimate of the method's error, so this was included in the assay's simulation.  A short time later and the model predicts quite a large variation in the ELISA when the process is operating under normal operating ranges.  As a result, variations in the ELISA results as a function of process variation are likely to be lost in the overall variation of the analytical result.
Model Simulation of CHO HCP as a Function of Load Amount and Load Conductivity

Based upon this information, and in conjunction with the data from the aggregate analysis, are the process parameters critical or not?

Critical or Non-Critical

My previous post sparked me to pose a question within LinkedIn and the resulting conversation has been really helpful.  What I want to do next is analyze what the guidelines mean in relation to a possible example of data that would be used in a filing.  In a previous post, I used data from an evaluation of CaptoAdhere to do some DOE analyses and illustrations.  I'm going to use that data to work through a thought experiment.  Let's start with some assumptions:

1. The data are for the last step in the fine purification prior to final filtration and formulation
2. The preceeding steps were a protein A capture, low pH viral inactivation, and an anion exchange column
3. The CHO HCP assay represents a cell-line specific assay
4. A risk analysis has identified both the aggregate and CHO HCP as critical quality attributes (CQA) for the product.
5. Characterization experiments were completed with a qualified laboratory-scale model of the manufacturing scale process and the results are to be used in supporting the process characterization section, the validation section and the manufacturing description of the regulatory filing
6. Finally, there exists limited manufacturing scale data, both analytical and process related, for the CaptoAdhere process.  Let's say...10 lots were produced by the time the commercial filing was submitted.

The model for the aggregate has found that the main effects of load amount and load conductivity are statistically insignificant; however, their pairwise interaction is statistically significant.

Parameter Estimates for Aggregate in the Eluate

Term		Estimate	Std Error	t Ratio	Prob>|t|
Intercept		0.0075547	0.000268	28.23	<.0001*
Load(93,312)		0.0005093	0.000353	1.44	0.1989
Load cond(10,30)		-0.000359	0.000363	-0.99	0.3614
Load*Load cond		0.0019552	0.000401	4.88	0.0028*

In contrast, the CHO HCP is dependent upon the load amount, the load conductivity and their pairwise interaction.

Parameter Estimates for CHO HCP in the Eluate

Term		Estimate	Std Error	t Ratio	Prob>|t|
Intercept		16.175041	0.876267	18.46	<.0001*
Load(93,312)		8.1260841	1.155147	7.03	0.0004*
Load cond(10,30)		-2.534245	1.18887	-2.13	0.0770
Load*Load cond		-3.252868	1.312414	-2.48	0.0479*

The prediction profiler was then used to simulate 5000 points using the following input:

1. Load: imagine the load standard deviation is 10 g/L and reflects the lot-to-lot variation in the process
2. Load conductivity: the conductivity is controlled fairly precisely at the manufacturing scale through precise, and accurate, measurements of the buffer constituents.  The result is a standard deviation of 0.5 mS/cm
3. SEC results:  The assay precision could be demonstrated in the validation of the method.  Based upon my experience, I'd be comfortable with a value of 0.04%

Simulating Manufacturing Scale Data

Based upon the model, as long as the process parameters are controlled to the precision indicated the aggregate percentage is essentially unchanging.  Based upon this data (I'll get to the CHO HCP data in the next post), would the load and load conductivity be critical or non-critical process parameters?   

Relationship of Upper Tolerance Limit to Prediction Profile

Argument against CPPs:  if the process parameter data from the manufacturing runs match, or are better, than the estimates used in the prediction profiler then the process parameters could be considered to be non-critical because within the normal manufacturing process variation, there is no statistically significant change in the critical quality attribute. 

Argument for CPPs: guidelines clearly state that a CPP is "a process parameter whose variability has an impact on a critical quality attribute and therefore should be monitored or controlled to ensure the process produces the desired quality." 

Characterizing a Chromatography Step

I'm back!  It was an epic vacation that required almost as much time to recover from!  That's a good vacation, if you ask me.  :-)

Previously, there were seven factors identified for characterization for a cation exchange step of a biosimilar.  Now, a full-factorial approach would require 128 runs for all 7 factors.  What a mess in terms of time and resources!  To reduce the design complexity, the authors take a cue from the literature where the buffer concentration and pH are wrapped up into one parameter for both the equilibration and elution buffers.  By combining these two parameters, the authors study the process in worst/best case situations for the equilibration and elution.  With 5 factors, a full factorial approach would take 32 experiments - still quite a bit of time and sample commitment.  In lieu of that, the authors apply a fractional factorial approach to track all the pairwise interactions that results in a reduction to 16 experiments.  Not too shabby!  I liked this strategy for its economy while maximizing the amount of knowledge.

The downside of the approach is the authors will miss the impact of the individual factors of buffer concentration and pH.  A Placket-Burman study with all 7 factors would take only 12 experiments.  Based upon the chemistry of the cation exchange resin, it's not a tremendous leap to imagine that load pH, elution pH, and aggregate amount in the load would be the dominant factors affecting the process performance.  Augmenting the original 12 experiments to evaluate these would require an additional 8 experiments (20 total).  For an additional four experiments, the authors could have obtained greater definition of the process space and reduce the number of process parameters necessary for including as part of their control strategy.

Sadly, the authors don't provide the data in their paper; however, they do provide the parameter estimates from their results that could be used to generate a simulation of the data (I may include that next time).  They found that the amount of aggregate in the load and the equilibration buffer were the driving process parameters for this antibody.  The downside of the result: the equil buffer is characterized by both the concentration of the buffer and the pH.  When the time comes to file these results, I'd bet the agencies are going to want to see a clear scientific picture of that the sponsor understands how the chemistry is driving the aggregate clearance and that the sponsor has the appropriate controls in place to ensure the product remains safe for patients.

I've grappled with how to define critical and non-critical process parameters - especially when there's limited raw material lots used in the early stages of a campaign.  The authors use the statistically significant approach to define critical and non-critical process parameters.  I like their approach; however, what happens when there's a statistically significant model but it is practically meaningless?  I'm going try to tackle that question in my next entry.

Vacation!

After a very long two weeks, I'm off for some R&R.  I'll pick up the posts next week.  In the meantime, fire up the grill, grab your favorite libation, and enjoy yourselves.

Preparing for characterization of the process space

With the FMEA completed and the process parameters identified, the next step is to qualify the scale down model.  One question that I've been considering is, "does the qualification lend itself to phase appropriate strategies?"

During efforts that carry a program from a research molecule to IND, there will most likely be a limited number of lots manufactured that lead to a limited process experience.  Further, for biologics the process can be expected to evolve under the pressures of cell culture optimization.  With such a dynamic process in place, a classic strategy for the purification team would be to leverage the knowledge from 1.5 cm and 10 cm column performance for establishing the scale-down model.  At these scales, the drug substance being generated is typically for pre-clinical research and is completed within the development lab.  As a result, the team has a clear picture of the scale comparability.  For example, the tubing type, plumbing length, pumps, on-line meters, etc., can all be evaluated along with the change in column scale (a 44x scale-up in this case).   When combined with the chromatographic performance and process performance (elution volume, recovery, etc) at each scale, the team can document the starting point of the process development history.  This also presents an opportunity to leverage process knowledge from the literature into the scale-down documentation and create a foundational document for future process development activities.

As the program moves into pivotal phase 2 status, the push is to deliver material to the clinic in sufficient quantities to support the clinical program.  With a bit of luck and great planning, the cell culture process has been completed to the point that they can be scaled-up.  For companies working with their CMOs, the path to scale-up becomes a bit more challenging as the purification equipment to support an 80 cm column is much different than that of the earlier models: the scale difference is 2844 between an 80 cm and 1.5 cm column.  The plumbing becomes stainless steel, the hold up volumes increase, water quality, and the differences continue and all will need to be documented.  Looping back to a previous post, the qualification of the laboratory-scale model presents a good opportunity to initiate (or revisit) the FMEA to prioritize the scale differences with the greatest potential to impact the process and product quality attributes.  To bridge the scale performance back to the early phase results, data from the 80 cm columns has to be compared with data from the laboratory-scale models.  To maximize the opportunity for success, the ideal situation would be to use process eluate retains from the clinical manufacturing scale in the laboratory-scale models whose process parameters are run as close as possible to those at clinical manufacturing scale.  An important point to consider is that there are relatively few raw material lots being used in the cell culture process during this stage of the program.  Consequently, the anticipated variation in the product quality will be relatively modest and this will need to be included in the discussion of the scale-dependent, or independent, process performance or product quality differences.

What are the metrics establishing the scale down model?  Based upon the above discussion, the argument can be made that these metrics will change, and become more robust, with the phase of the program (see links at the start of this entry).  Documenting, and providing scientific justification, through the drug development stages will establish the basis of the QbD program because the laboratory-scale model qualification links the design space knowledge to the manufacturing scale.  The importance of this link cannot be under appreciated when the time comes to file the commercial process because the sponsor will have to provide a scientifically sound justification that the results obtained in laboratory-scale models are predictive of the process at manufacturing scale.

Handling the FMEA

In the previous posts, JMP was used to illustrate how a fishbone diagram could be used to sketch the how the different elements of a chromatography process could affect the product quality and how there needed to be sufficient quality systems in place to enable efficient conversion of this data into knowledge.

The FMEA should enable the reviewer to see a clear connection between the process parameters and the critical quality attributes (CQAs) of the product.  The next step is a bit trickier.  The FMEA should also provide a roadmap for developing the process knowledge relating the effect of these process parameters to the CQAs.  What I really liked about the paper by Xu, et al was the use of a Pareto chart to establish the number of factors that would be studied in their process characterization work.  The Pareto plot allows the visualization of the risk priority number (RPN) as well as the cumulative effect of the number of factors.  By setting the bar at 90% of the cumulative effect, the argument may be made to the regulators that the majority of the risk to the product quality has been accounted for in the process characterization work.

Pareto Plot of RPN Score by Operating Parameter (Adapted from Xu, et al)

Setting the bar at 90% means there will be seven factors that need to be studied for the process characterization work.  These can be readily handled with a screening design to identify which have the most significant effect and then augment the design for refinement of the operating space.  A Plackett-Burman would need only 12 experiments to find the main effects!  Nice.  Before tackling these experiments, a justifiable scale-down model has to be in place for comparison with the manufacturing scale.  In the next entry, I'll be talking about strategies for justification of a scale down model.

Putting the Quality in QbD

In the previous post, I laid out an example of how to begin documenting the QbD process.  Over at LinkedIn, there's a discussion about the challenges of bringing a QbD/PAT program into existence that is really worth the read.  My plan for this entry is to outline some of the pathways that need to be in place for QbD to be effective and deliver on its promises.

I believe the important, and obvious, starting point is in the Q of QbD.  The agency's have clearly articulated their expectations through the guidelines (see link to the right).  What are the practical implications of these?  Paperwork.  We have to work with the quality systems to enhance, or build, the infrastructure to support a QbD approach.  For example, there are lots of examples of risk analysis in the literature; however, the approach has to be coded into the quality system to ensure that people are appropriately trained in the methodology and the results documented, reviewed and filed within the quality system for reference.  For pharmaceuticals and biopharmaceuticals, the risk assessment has to be phase appropriate and continuous: what is done for phase 1 is not what is done for phase 3/validation and the risk analysis should evolve as a program pushes through from phase 1 to phase 3.  Oh, the QbD process also has to be flexible to handle in-licensing of a product.  All of this, and more, represent the foundation of a solid QbD program.

What does this mean on the day to day?  The lines of communication to the various quality groups have to be reinforced every day.  Conversations over a cup of coffee, formal strategy meetings, and more all serve to develop those relationships that will enable the organization to leverage its knowledge effectively and efficiently.  

At the end of a long day...

I'm flattered so many of you, from so many different countries, have been visiting.  Thanks so much and I look forward to your next visit!

Analysis of Visits to "In the Process Stream"

CEX Chromatography and QbD

I was going through my previous posts on the media selection process last night while listening the to the Bruins game.  Those last two minutes were heartbreaking. Anyways, using the augmenting process to build up a body of knowledge for the model led to a different conclusion than that proposed by the authors despite using the same data set.  Which is more correct?  What's the best way to process a CCD data set?  I'm a big proponent of starting simple and then adding complexity.  With a CCD data set, I'll start with the full-factorial model to identify the statistically significant effects.  From there, the addition of center points allows a check for curvature.  If there's none, then the analysis is complete!  If the center points are in alignment with the full-factorial model, including the axial components becomes an important step in identifying what's driving higher order behavior.  Last, tying the results back to the underlying chemistry will become a vital link when writing up the work for archiving.  Being able to have traceability and sound scientific understanding of process behavior is SO important when the time comes from authoring the IND or, if you're lucky, the BLA.

The next article of merit is from Genor Biopharma in Shanghai.  The authors take an integrated approach to developing a cation exchange step using QbD principles.  One of their first steps is to define a fish-bone diagram of the factors affecting the process.  JMP allows you to do this quite easily as a quick sketch up of the process.

Start by making a two-column file in JMP (see below).  In this case, I've labeled them Parent and Child to make it easy to remember the order.  In every case, the child belongs to the parent and this would be repeated until you had all the children accounted for each parent in your process.

Setting up for a Fishbone diagram

In my version of JMP, I go to Graph and then diagram:

Next, put the categories into the appropriate placing within the GUI

and the result follows!

I like these diagrams throughout the design process.  They provide an easy reference for the preliminary FMEA that prioritizes the experiments and for documenting along the way which has the greatest impact on the process.

In the next posting, I'll take a deeper look at their QbD approach.

Last round of Media Analysis

I spent a fair amount of time noodling around with the problem and then did a quick visualization to confirm my thoughts.  When the points of the full-factorial are shown with the center points, notice they all fall at the origin.  As a result, there's no way to figure out which term is the leading to quadratic behavior.  I'm embarrassed to say the amount of time I spent reaching that conclusion!
Distribution of Coded Points within the Design

What to do?  We're at 35 experiments (32 for the full factorial, 3 for the center points).  What I decided to do was augment, yes, again, but this time I'll add axial points.  However, for which factors?  I decided to stick with the main effects that had the greatest statistical significance: Glutamine, NEAA, and ITS.  In the augmentation, I also included an additional 3 center point runs.  The experimental total is now at 44!

Fitting to a full factorial model that includes square terms for Glutamine, NEAA, and ITS gives some good results.  The ANOVA results aren't spectacular, but the model fits the data better than normal variations.  More importantly, the square term behavior is coming from the NEAA and ITS.
ANOVA for Augmented Design with Axial Components

Source	DF	Sum of Squares	Mean Square	F Ratio
Model	9	296.45290	32.9392	8.2079
Error	34	136.44500	4.0131	Prob > F
C. Total	43	432.89790		<.0001*

Parameter Estimates from Fit

Term		Estimate	Std Error	t Ratio	Prob>|t|
Intercept		10.829294	0.785856	13.78	<.0001*
Glutamine		-0.985914	0.320352	-3.08	0.0041*
NEAA		1.7462562	0.319869	5.46	<.0001*
ITS		0.4912445	0.320681	1.53	0.1348
Glutamine*NEAA		0.2518681	0.35905	0.70	0.4878
Glutamine*ITS		0.2831327	0.357962	0.79	0.4345
NEAA*ITS		-1.104809	0.358561	-3.08	0.0041*
Glutamine*Glutamine		-0.110353	0.392928	-0.28	0.7805
NEAA*NEAA		-1.504103	0.392928	-3.83	0.0005*
ITS*ITS		-1.241603	0.392928	-3.16	0.0033*

After simplifying the model to just the statistically significant terms, the ANOVA improves as well as nearly all the factors are highly relevant.  The outlier factor is the main effect of ITS (p>0.05).  I can rationalize an argument for including and excluding it in the final model to the data; in the end, you'll have to decide which way you'd take the model.

ANOVA from Simplified Model Fit to Augmented Design with Axial Components

Source	DF	Sum of Squares	Mean Square	F Ratio
Model	6	291.39175	48.5653	12.6985
Error	37	141.50615	3.8245	Prob > F
C. Total	43	432.89790		<.0001*

Parameter Estimates from Simplified Model Fit to Augmented Design

Term		Estimate	Std Error	t Ratio	Prob>|t|
Intercept		10.681397	0.599025	17.83	<.0001*
Glutamine		-1.002546	0.311948	-3.21	0.0027*
NEAA		1.7072543	0.309969	5.51	<.0001*
ITS		0.4520094	0.310588	1.46	0.1540
NEAA*ITS		-1.125011	0.348986	-3.22	0.0026*
NEAA*NEAA		-1.485901	0.376752	-3.94	0.0003*
ITS*ITS		-1.223401	0.376752	-3.25	0.0025*

An interesting observation is that my approach leads to a different conclusion about the model that fits the data than the conclusion presented in the paper.  I'll discuss the implications of this in the next, and final, entry about these experiments.

Prediction Profiler from Simplified Model Fit to Augmented Design