Showing posts with label process characterization. Show all posts
Showing posts with label process characterization. Show all posts

Monday, June 18, 2012

Viral clearance studies …. are they needed for proteins produced using bacterial or yeast fermentation processes?

In E. coli or Pichia pastoris-based bioproduction of recombinant proteins, there are no suitable host cells for amplification of viruses that are infectious for humans. Bacteria such as E. coli and yeast such as P. pastoris can only support the growth of certain bacteriophage or yeast viruses, respectively. These types of viruses are not infectious for humans or animals.
The potential for viral infection of a bacterial cell process or a yeast cell process used for production of a recombinant protein is therefore a business risk, not a patient safety risk. This is the reason why case studies of bacteriophage infection of bacterial fermentation processes or of viral contamination of yeast cell processes have not appeared in the literature (or in the news). Unlike virus contamination of animal cell-based production processes, the events involving bacterial or yeast do not have patient safety implications. When a bacteriophage or yeast virus contamination occurs, manufacturers quietly go about the business of restarting the fermentation. Remediation and prevention of future occurrences is driven primarily by business concerns, as opposed to regulatory concerns.
Manufacturers of recombinant proteins produced using bacterial or yeast cell substrates are not required to conduct cell line viral testing (though phage induction studies may be performed as part of bacterial cell line characterization – to mitigate business risk!), nor are they required to conduct lot-by-lot viral testing of bulk harvest samples. Such requirements are mandated for production processes using animal or human cells. As the title of the document “Quality of Biotechnological Products: Viral Safety Evaluation of Biotechnology Products Derived from Cell Lines of Human or Animal Origin” indicates, the ICH Q5A (R1) guidance mandating such testing is clearly directed toward manufacturing processes utilizing the types of cells which may support growth of viruses infectious for human or animal cells. This ICH document also provides guidance on the conduct of viral clearance studies.

Viral clearance studies are intended to provide evidence of the ability of downstream purification steps of a manufacturing process to remove or inactivate viruses that potentially may contaminate the process, make their way into final product, and represent a health risk to patients. As such, the evaluation typically involves both relevant viruses (i.e., viruses that are known to contaminate these kinds of processes) and model viruses (i.e., viruses representative of the types of viruses that could contaminate these processes). In the case of bacteria- and yeast-based manufacturing processes, the relevant viruses (bacteriophage and yeast viruses, respectively) are not infectious for humans or animals. In addition, there are no model viruses that are capable of infecting these kinds of cells that are of concern for humans or animals. Therefore, viral clearance studies are typically not required or conducted.

Are there any circumstances where a bacterial or yeast fermentation process could be expected to harbor a virus capable of infecting humans? The only that I can imagine is a process utilizing primary or secondary animal-derived raw materials in rather large quantities. The worry would not be that viral amplification might occur, but that some carryover of surviving animal viruses to the final product might be possible. The use of such animal-derived materials by a manufacturer of a therapeutic protein must be justified based on risk analysis. If substantial risk is introduced by the use of such raw materials, then perhaps a raw material treatment approach would need to be validated.

In the absence of the use of animal derived materials or plant materials not subject to processing steps that would inactivate contaminating animal viruses, incorporation and validation of viral clearance steps into protein production processes using bacteria or yeast cell substrates is not expected, nor would this be of practical value in assuring patient safety.

Thursday, May 24, 2012

Update: New USP General Chapter 1050.1


There is a new general chapter being prepared for inclusion in the United States Pharmacopeia (USP). It will be entitled “Design, Evaluation, and Characterization of Viral Clearance Procedures” and will be numbered 1050.1 to associate it with the current General Chapter <1050>.

A little history is called for to make this association more clear. Chapter <1050> first appeared in supplement 10 of USP23-NF-18 in 1999. It was, and still is, a verbatim copy of the International Conference on Harmonisation (ICH) document Q5A R1. In fact, it has an identical title: “Viral Safety Evaluation of Biotechnology Products Derived from Cell Lines of Human or Animal Origin”.

In many respects, U5P Chapter <1050> (and ICH Q5A R1) is similar in content and philosophy to the 1993 US FDA Points to Consider (PTC) document entitled "Characterization of Cell Lines used to Produce Biologicals".  Like the 1993 PTC, USP Chapter <1050> expresses an overall safety paradigm composed of three orthogonal approaches. These approaches may be summed up as consisting of testing of raw materials and cell banks, testing of unprocessed bulk harvest, and evaluation of viral clearance steps used during purification. Also like the 1993 PTC, Chapter <1050> describes potential sources of viral contamination, including raw materials and the cell substrate. The scopes of the documents are similar, applying to products derived from cell lines. Transmissible spongiform encephalopathy agents and live or inactivated intact viral vaccines and gene therapy vectors are out of scope and are covered by other guidance documents.

Due to the potential for introducing a viral contaminant through use of an infected cell substrate (e.g., SV40 in polio vaccine, or more recently PCV-1 in rotavirus vaccine), USP Chapter <1050> and the 1993 PTC address the concept of cell banking and the extensive viral testing that is required for cell banks at the Master Bank, Working Bank, and Limit of Cell Growth levels. The second approach of the safety paradigm is the requirement for lot-to-lot testing of the unprocessed bulk harvest. Testing is done at this level as opposed to purified product since purification is designed to eliminate viruses and therefore might preclude the manufacturer from discovering a viral infection. It is the intent of the regulatory agencies that potential opportunities for introduction of a viral contaminant be investigated and remediated, so testing is done prior to purification. The viral testing that is required for unprocessed bulk harvest is a subset of the tests done on the cell banks.

Both USP Chapter <1050> and the 1993 PTC discuss the limitations of viral testing as a sole means to assure viral safety, and this provides a segue for discussion of the need for viral clearance steps and the validation of the ability of such steps to clear viruses from the product. In fact, USP Chapter <1050> does a good job of providing the fundamentals of viral clearance evaluation, however there was a feeling expressed by some in the industry (and especially Mike Rubino, a member of the original ad hoc advisory panel for this chapter) that more detailed guidance on experimental design was needed. This, by the way, is true in general about Chapter <1050> (and IQH Q5A Rl upon which USP <1050> was based) that it is strong on philosophy but weak on specific methodological detail.

The original USP ad hoc advisory panel for Chapter <1050> revision was assembled with the purpose of updating the chapter to include this missing experimental detail. The panel went through each section of the existing chapter and added the detail that was felt to be lacking. This was done with the overall mandate to avoid introducing any new language that might conflict with the original language (and ICH Q5A Rl). The revised Chapter <1050> was published in the Pharmacopeial Forum in 2010. Responses obtained indicated that there was reluctance to modify in any way the language of this chapter without also modifying ICH Q5A. The situation as of February 2011 was described in an earlier posting.

As a result, the USP decided to keep Chapter <1050> intact and identical to ICH Q5A Rl.  A new ad hoc advisory panel was assembled with the purpose of producing a new chapter in the general information series that would be a companion to the existing Chapter <1050>. The new chapter received the number 1050.1 and provided the vehicle for adding the desired methodological detail.

The proposed new General Chapter <1050.1>, entitled "Design, Evaluation, and Characterization of Viral Clearance Procedures" consists of some background information on process evaluation and process characterization, then launches into experimental design for evaluating both inactivation and removal steps. The latter section includes some specific experimental design flow charts addressing virus removal by filtration and chromatography and virus inactivation. Along with the flow charts describing the design of the studies are description of the methods used to assess potential cytotoxicity or interference caused by the process materials themselves.

The subheadings of the background information are shown in the text box below.



The information presented under these subheadings reflects the panel's understanding of current regulatory expectations. Regulatory input obtained during the public comment period will assure that the panel's perceptions were correct. Discrepancies will be corrected as required.

Finally, a goal of the new chapter was to provide an updated list of the types of viruses that have been and may be used in viral clearance evaluations. For studies enabling clinical trials, it is common for evaluations to use a parvovirus such as MMV and a retrovirus such as X-MuLV as models. This provides a small non-enveloped virus to challenge filtration steps, as well as an enveloped virus to assess inactivation steps designed for lipid enveloped viruses. For BLA enabling studies, a few additional viruses may be selected from this updated list, keeping in mind that the viruses should represent a diversity of characteristics (envelope status, genome type, size, etc.).

The intention of the new chapter <1050.1> is that manufacturer’s may use the methods as appropriate to their own processes and will be able to cite the guidance in their descriptions of the study design. At the present time, no guidance having sufficient experimental design descriptions is available to cite in this respect. The planned date for publication of the proposed chapter in the Pharmacopeial Forum is January 2013.

Friday, August 26, 2011

Our take on process-specific vs. generic host cell protein assays

By Drs. Ray Nims and Lori Nixon 

The residual host cell protein (HCP assay) is used to determine the concentration, in process streams, of protein originating from the production cell (Chinese hamster, NS0, E. coli, etc.) used in the manufacture of a biologic. Host cell protein is considered to be a process-related impurity/contaminant. The most typical approach to quantitation of residual host cell protein in a test sample is the enzyme-linked immunosorbant assay (ELISA) platform.

Generic vs. process-specific assays.  A question that often arises is whether a process-specific HCP assay is required, or whether a generic assay can be used. Although there is a spectrum of “process specificity” that can be described, in this discussion, we are considering a generic assay to be one that is developed from a broad set of antigens from the host strain or closely-related strains, but not necessarily under specific process conditions mimicking the actual production/purification process of the protein of interest. That is, the reagents for a generic assay could be developed in advance of a defined production process. For example, it is often possible to purchase a “generic” HCP ELISA kit from a commercial provider, such as Cygnus Technologies, selecting a kit that matches the production cell line. Practically speaking, most firms start with generic HCP methods in early development for an obvious reason: you can’t have a process-specific assay until you have a defined process. Even after the process is defined, the lead time for developing custom (process-specific) HCP reagents and an assay using these can be as long as 1-2 years.
It should be appreciated from the outset that there is no perfect HCP assay. Production cells can express thousands of proteins, with expression patterns differing under different growth conditions. At each successive stage of the recovery and purification process, the population of HCPs is altered. At the final drug substance stage, there may be only a few HCPs present that have co-purified with the product. If you attempt to design an assay that only selects those few remaining HCPs, then some will argue that you might miss HCPs that might make it through under other circumstances (such as a process deviation).  If you attempt to design antibody reagents using the entire unpurified mixture of proteins from the production cell, the proteins of most interest may not elicit a good antibody response and you may have inadequate sensitivity to quantify HCP in your final drug substance. In fact, it has been reported that process-specific antibodies lead to HCP assays that are more sensitive for certain HCP species (and sometimes more broadly reactive) than the antibodies provided in a generic HCP assay. 
In recent years, there has been some regulatory pressure setting an expectation that a process-specific HCP assay must be developed for product registration. In our opinion, this is unjustified as a blanket prescription.  There are a multitude of approaches for creating antigen and antibody reagents, but the resulting assay should be judged primarily on its own merits.
Suitability of any HCP assay (whether generic or process-specific) must be evaluated on a case-by-case basis, and the assay validated for the specific product in question. One of the first characteristics to aim for is an assay that is able to readily quantitate HCP in the final drug substance. Without this sensitivity, it is difficult to get results that are meaningful to guide process optimization or ensure control of your product. If the available generic method(s) have limited sensitivity for the sample of interest (results are below the required range of quantitation), then it is a good idea to start planning the development of more process-specific reagents—which are likely to afford better sensitivity. Since the numerical values reported from these assays are semi-quantitative at best, sensitivity should be judged on the performance in actual samples rather than on the product specification. Of course there are other assay characteristics (non-reactivity to product, precision, dilutional linearity, spike recovery, etc) that are required to meet validation acceptance criteria. The antigen coverage should be determined, typically by 2D electrophoresis or 2D HPLC-ELISA. It is unrealistic to expect 100% coverage of all of the proteins present in the production cell; on the other hand, it is relatively more important to ensure response against bands that persist in the purification process.  All other things being equal, of course greater coverage is more desirable.
Example of 2D electrophoresis of E. coli proteins, from Kendrick Labs

Whether a generic or process-specific method is implemented, it is vital to ensure a consistent reagent supply that will last throughout the commercial life of the product. If using vendor-supplied antigen and antibody, be aware that by nature these are “single-source” reagents with associated supply risks. If creating a custom reagent set, make enough supplies to last for years to decades.
As mentioned above, most firms start with a generic assay and may move to a process-specific assay in later stages of development. A close evaluation of the generic assay during its characterization and validation may lead to the conclusion that the generic assay is suitable for your product.  In this case, be prepared to make a strong case in defense of your generic assay—based on empirical data with your product, rather than theoretical speculation.

Thursday, August 26, 2010

Do We Have Clearance, Clarence?

By Dr. Scott Rudge

As in take offs and landings in civil aviation, the ability of a pharmaceutical manufacturing process to give clearance of impurities is vital to customer safety. It’s also important that clearance mechanism be clear, and not confused, as the conversation in the classic movie “Airplane!” surely was (and don’t call me Shirley).

There are two ways to demonstrate clearance of impurities.

The first is to track the actual impurity loads. That is, if an impurity comes into a purification step at 10%, and is reduced through that step to 1%, then the clearance would typically be called 1 log, or 10 fold.

The second is to spike impurities. This is typically done when an impurity is not detectable in the feed to the purification step, or when, even though detectable, it is thought desirable to demonstrate that even more of the impurity could be eliminated if need be.

The first method is very usable, but suffers from uneven loads. That is, batch to batch, the quantity and concentration of an impurity can vary considerably. And the capacity of most purification steps to remove impurities is based on quantity and concentration. Results from batch to batch can vary correspondingly. Typically, these results are averaged, but it would be better to plot them in a thermodynamic sense, with unit operation impurity load on the x-axis and efflux on the y-axis. The next figures give three of many possible outcomes of such a graph.


In the first case, there is proportionality between the load and the efflux. This would be the case if the capacity of the purification step was linearly related to the load. This is typically the case for absorbents, and adsorbents at low levels of impurity. In this case (and only this case, actually) does calculating log clearance apply across the range of possible loads. The example figure shows a constant clearance of 4.5 logs.


In the second case, the impurity saturates the purification medium. In this case, a maximum amount of impurity can be cleared, and no more. The closer to loading at just this capacity, the better the log removal looks. This would be the point where no impurity is found in the purification step effluent. All concentrations higher than this show increasing inefficiency in clearance.


In the third case, the impurity has a thermodynamic or kinetic limit in the step effluent. For example, it may have some limited solubility, and reaches that solubility in nearly all cases. The more impurity that is loaded, the more proportionally is cleared. There is always a constant amount of impurity recovered.

For these reasons, simply measuring the ratio of impurity in the load and effluent to a purification step is inadequate. This reasoning applies even more so to spiking studies, where the concentration of the impurity is made artificially high. In these cases, it is even more important to vary the concentration or mass of the impurity in the load, and to determine what the mechanism of clearance is (proportional, saturation or solubility).

Understanding the mechanism of clearance would be beneficial, in that it would allow the practitioner to make more accurate predictions of the effect of an unusual load of impurity. For example, in the unlikely event that a virus contaminates an upstream step in the manufacture of a biopharmaceutical, but the titer is lower than spiking studies had anticipated, if the virus is cleared by binding to a resin, and is below the saturation limit, it’s possible to make the argument that the clearance is much larger, perhaps complete. On the other hand, claims of log removal in a solubility limit situation can be misleading. The deck can be stacked by spiking extraordinary amounts of impurity. The reality may be that the impurity is always present at a level where it is fully soluble in the effluent, and is never actually cleared from the process.

Clearance studies are good and valuable, and help us to protect our customers, but as long as they are done as single points on the load/concentration curve, their results may be misleading. When the question comes, “Do we have clearance, Clarence?” we want to be ready to answer the call with clear and accurate information. Surely varying the concentration of the impurity to understand the nature of the clearance is a proper step beyond the single point testing that is common today.

And stop calling me Shirley.

Thursday, June 17, 2010

Informing the FMEA

By Dr. Scott Rudge
Risk reduction tools are all the rage in pharmaceutical, biotech and medical device process/product development and manufacturing. The International Conference on Harmonization has enshrined some of the techniques common in risk management in their Q9 guidance, “Quality Risk Management”. The Failure Modes and Effects Analysis, or FMEA, is one of the most useful and popular tools described. The FMEA stems from a military procedure, published in 1949 as MIL-P-1629, and has been applied in many different ways. The most used method in the health care involves making a list of potential ways in which a process can “fail”, or produce out of specification results. After this list has been generated, each failure mode in the list is assessed for its proclivity to “Occur”, “Be Severe” and “Be Detected”. Typically, these are scored from 1 to 10, with 10 being the worst case for each category, and 1 being the best case. The scores are multiplied together, and the product (mathematically speaking) is called the “Risk Priority Number”, or RPN. Then, typically, development work is directed towards the failure modes with the highest RPN.

The problem is, it’s very hard to assign a ranking from 1 to 10 for each of these categories in a scientific manner. More often, a diverse group of experts from process and product development, quality, manufacturing, regulatory, analytical and other stake holding departments, convene a meeting and assign rankings based on their experience. This is done once in the product life-cycle, and never revisited as actual manufacturing data start to accumulate. And, while large companies with mature products have become more sophisticated, and can pull data from other similar or “platform” products, small companies and development companies can really only rely on the opinion of experts, either from internal or external sources. The same considerations apply to small market or orphan drugs.

Each of these categories can probably be informed by data, but by far the easiest to assign a numerical value to is the “Occurrence” ranking. A typical Occurrence ranking chart might look something like this:

These rankings come from “piece” manufacturing, where thousands to millions of widgets might be manufactured in a short period of time. This kind of manufacturing rarely applies in the health care industry. However, this evaluation fits very nicely with the Capability Index analysis.

The Capability Index is calculated by dividing the variability of a process into its allowable variable range. Or, said less obtusely, dividing the specification range by the standard deviation of the process performance. The capability index is directly related to the probability that a process will operate out of range or out of specification. This table, found on Wikipedia (my source for truth), gives an example of the correlation between the calculated capability index to the probability of failure:

As a reminder, the capability index is the upper specification limit minus the lower specification limit divided by six times the standard deviation of the process. The two tables can be combined to be approximately:
How many process data points are required to calculate a capability index? Of course, the larger the number of points, the better the estimate of average and standard deviation, but technically, two or three data points will get you started. Is it better than guessing?

Wednesday, April 7, 2010

What's Your Velocity?

By Dr. Scott Rudge

With the development of very high titer cell culture and fermentation processes, downstream processing has been identified as a new bottleneck in biotechnology. The productivity of chromatography in particular, has become a bottleneck. There are two schools of thought for scaling up chromatography: in one, linear velocity (flow rate divided by column cross sectional area) is held constant; in the other, the total (volumetric) flow rate divided by the volume of the column is held constant. In the former method, the length of the column has to be held constant. In the latter method, the geometry of the column is not important, as long as the column can be packed efficiently and flow is evenly distributed. This makes the latter method more flexible, and accommodating of commercially available off the shelf column hardware packages. But does it work?

In my experience, holding flow rate divided by column volume constant between scales works very well. There is plenty of theoretical basis for the methodology as well. Yamamoto has published extensively on the reasons that this technique works. This method is also the basis for scale up described in my textbook. Here, briefly, and using plate height theory, is the theoretical basis:

The basic goal in chromatography scale up is to maintain resolution. “Resolution” is a way to describe the power of a chromatography column to separate two components. It depends on the relative retention of the components, which is fixed by the thermodynamics of the column and remains constant as long as the chemistry (the resin type, the buffer composition) remains constant. It also depends on the peak dispersion in the column, which is a function of the transport phenomena, and is only related to the chemistry by the inherent diffusivity of the molecules involved. Otherwise, it is dependent on mass transfer, flow rate, temperature, flow distribution. Treating the thermodynamics as constant, we can say:

where Rs is Resolution, and N is the number of theoretical plates. N is the ratio of the column length L to the plate height, H, so
 
In liquids, H is approximately a linear function of linear velocity v, according to van Deemter, as discussed in a previous post. So we can say that H = Cv (where C= the van Deemter constant). Now, the linear velocity is the flow rate divided by the cross sectional area of the column, A, and the column volume, V, is the cross sectional area times the column length. The total flow rate (F) divided by the column volume is held constant between scales, we’ll call this constant “f”.
 
This bit of mathematical gymnastics says that Resolution only depends on two fundamental properties of the scale up, van Deemter’s term “C”, which considers the dispersion caused by convection relative to mass transfer to and from a resin particle, and “f”, the flow rate relative to the column volume that is chosen. There is no need to hold column length or linear velocity constant as long as the flow rate relative to column volume is held constant. You might also notice from the math that doubling the column length has the same effect on resolution as dropping the flowrate by half. However, doubling the length costs more in terms of resin, and consumes more solvent than decreasing the flow rate (that’s for you, Mike Klein!) and increases the pressure.

The plate count analysis is very phenomenological, but it does hold up under practice (otherwise it would be abandoned). And the more delicate mathematical models predict the same performance, so confidence in this scale up model is high.

One common mistake made by those using the constant linear velocity model is in adding extra column capacity. Since most people are unwilling to pay for a custom diameter column, but base their loadings on the total volume of resin, they add bed volume by adding length. But since they are unwilling to change the linear velocity, they end up decreasing the productivity of the column (because, for example, the resolution they achieved at a smaller scale in a 12 cm long column at 60 cm/hr is now being performed in a 15 cm long column, at 60 cm/hr, therefore taking 25% more time).

If the less well known constant F/V model is used for a process involving mammalian cells, it would be imperative to explain and demonstrate this model in the scale down validation that is a critical part of the viral clearance package.

But how can you get even more performance out of your chromatography? Treating the unit operation as an adsorption step, and scaling up using Mass Transfer Zone (MTZ) concepts will be treated in a future posting.

Thursday, March 4, 2010

QbDer, Know Thy Model!

By Dr. Scott Rudge

Resolution in chromatography is critical, from analytical applications to large scale process chromatography. While baseline resolution is the gold standard in analytical chromatography, it is seldom achieved in process chromatography, where “samples” are concentrated and “sample volumes” represent a large fraction of the column bed volume, if not multiples of the bed volume. How do you know if your resolution is changing in process chromatography if you can’t detect changes from examining the chromatogram?


Many use the Height Equivalent to a Theoretical Plate technique to test the column’s resolution. In this technique, a small pulse of a non-offensive but easily detected material is injected onto the column, and the characteristics of the resulting peak are measured. The following equation is used:


Where H is the “height” in distance, tr and tw,1/2 are the retention time and time of the width of the peak at ½ height, L is the length of the column and N is the “number” of theoretical plates in the column. The lower the H, the smaller the dispersion, the greater the resolution.

It is well established that the flowrate and temperature affect the plate height. In fact, when the plate height is plotted against flow rate, we generate what is typically called the “van Deemter plot”, after Dutch scientists (van Deemter et al., Chem. Eng. Sci.,5, 271 (1956)) who established a common relationship in all chromatography (gas and liquid) according to the following equation:


Where A, B and C are constants and v is the linear velocity (flow rate divided by the column cross sectional area) of fluid in the column. It was later proposed by Knox (Knox, J., and H. Scott, J. Chromatog., 282, 297 (1983)) that van Deemter plots could be reduced to a common line for all chromatography if the plate height was normalized to resin particle size, and linear velocity was normalize to the resin particle size divided by the chemical’s diffusivity. While this did not turn out to be generally true, it is very close to true. Chemical engineers will recognize the ratio of linear velocity to diffusivity over particle size as the Peclet number, “Pe”, a standard dimensionless number used in many mass transfer analyses.

Since diffusivity is sensitive to temperature, it is logical that Pe is also sensitive to temperature, decreasing temperature decreases the diffusivity, and increases Pe. Thus, Pe is inversely proportional to temperature. We measured plate height as a function of linear flowrate and temperature in our lab on a Q-Sepharose FF column, using a sodium chloride pulse, and found the expected result, shown in the graph below.


We can easily use this graph as a measurement of van Deemter’s parameters A and C, and find the dependence of the diffusivity of sodium chloride on temperature. Based on these two points, we find the Peclet number proportional to 0.085/T where T is in degrees Kelvin. We also find the dependence of plate height on linear velocity, and we can predict that resolution will deteriorate in the column as flow rate increases.

We can also use Design of Experiments to find the same information. Analyzing the same data set with ANOVA yields significant factors for both flow rate and temperature, as shown in the following table:

Term Coef SE Coef T P
Constant0.097560.056041.740.157
Temp-0.0079350.001858-4.270.013
v0.04970.021472.320.082




But since the statistics don’t know that the temperature dependence is inverse, or that the Peclet number is a function of both temperature and flowrate, the model yields no additional understanding of the process.

It is possible to use analysis of variance to fit a model other than linear, as the van Deemter model clearly is. But one must know that the phenomenon being measured behaves non-linearly in order to use the appropriate statistics. Using Design of Experiments blindly, without knowing the relationship of the factors and responses, leads to empirical correlations only relevant to the system studied, and should only be extended with caution.