A comparison of experimental permeability for a series of captopril carboxyl ester prodrugs with the QSAR models of skin permeability proposed by Moss and Cronin (2002), Potts and Guy (1992), Potts and Guy (modified by Wilschut et al. 1995) and Robinson et al. (modified by Wilschut et al. 1995). The first series of data (left-hand side, from front to back) is the experimental data (from Moss et al. 2006) for the permeability of captopril and its C1–C6 carboxyl ester prodrugs, and each subsequent pair of columns shows the predicted permeabilities for a particular QSAR model using, in the first case, a predicted value for log P and, secondly, an experimentally determined value for log P
It is also striking that despite approximately 15 % of its data having been shown to be erroneous, the Potts and Guy (1992) algorithm—and others based on this study—is still used by researchers and other models, which incorporate the corrected data, find little utility. Whether this is through a sense of the real strength of Potts and Guy’s model—its simplicity and ease of use—in the context of an inherently variable data environment, or a sense that the points raised by Johnson et al. (1995) do not significantly alter its key message (that permeability is dependent upon lipophilicity and molecular weight) is unclear. In this latter context, it should also be commented that when these erroneous data were corrected and reanalysed, the outcome was very similar to that originally presented by Potts and Guy (Moss and Cronin 2002).
Outcomes similar to those reported by Moss and Cronin (2002) were also found by Buchwald and Bodor (2001). While their study cannot be directly compared to the work of Moss and Cronin (2002), it draws similar conclusions. For example, they not only comment on the strengths of the Potts and Guy’s approach in terms of transparency and simplicity, but also then comment on the strong interrelatedness of the significant parameters, log P and MW, as they had described previously how the former is strongly size-related (Buchwald and Bodor 1998). Thus, they commented that such a strategy is effectively a “top-down” approach which sheds no specific mechanistic information on the diffusion process. They commented that a solvatochromic approach would resolve such issues; as reported in Chap. 4, these studies highlighted the role of hydrogen bonding in the skin absorption process. This includes, for example, the early use in this field of the ΔlogP o/h descriptor (where ΔlogP o/h = ΔlogP o/w−ΔlogP heptane/w) by El Tayar et al. (1991) and its role in characterising hydrogen bond donor acidity. They commented that studies where data are collated together and not analysed separately in subgroups appear to be contradictory to El Tayar’s findings. Buchwald and Bodor described the application of their “QLogP” method to couple mechanistic insight with the simplicity of the Potts and Guy’s approach.
Sitting above the attempts of Chilcott et al. (2005) to standardise methodology in the measurement of in vitro percutaneous absorption is a diverse range of experimental methods from which such data can be derived—all of which may contribute to data sets used to develop algorithms of skin absorption. Thus, it is important to remember that the experimental protocols may be significantly different and that they reflect primarily the needs of experimenters in particular studies and not the needs of modellers who will later abstract such data from those studies for their own purposes. In a particular in vitro study, there are a large number of parameters that can be altered, by design or otherwise, during the conduct of the experiment. Thus, in the collation and use of such data for the development of structure–permeability relationships, there is the possibility of data being compiled from experiments carried out by a wide range of methods. This may affect the relevance of the models subsequently developed and the validity of their output.
The relevance and advantages of conducting in vitro studies, particularly in the earlier stages of a formulation or drug development programme, have been discussed in detail in Chap. 2. The main benefits of such tests are in cost and convenience, but they also allow specific technical modifications to be made to the experiment. For example, if required, an experiment can consider different types of diffusion cell (i.e. static for flow-through cells), different states of occlusion, the condition of skin (where, e.g., a researcher may have an interest in permeation following chemical or physical depilation), the use of different solvents in either the donor or receptor compartments of the diffusion cell and a range of experimental temperatures. Such variations in experimental technique can allow significant control of the laboratory environment and therefore specific phenomena to be studied, such as the excellent thermodynamic studies conducted by Hadgraft and colleagues, discussed in Chap. 5. They also allow important protocols for the understanding of toxicity and permeation to be developed, where the laboratory advantages of radiochemical labelling techniques can be exploited in a manner that is not appropriate in in vivo studies. Clearly, they do not consider issues such as blood flow and metabolism, and the constraints and limitations of any in vitro method, when compared to in vivo testing in particular, should be understood and any results treated in the appropriate context. It thus follows that the same caveats should apply to mathematical algorithms derived from such data, as well as being mechanistically understood in the context discussed by Buchwald and Bodor (2001).
Further issues that may introduce possibly significant variance into experimental protocols include the selection and treatment of skin. Standardised measures of skin integrity (such as measurement of transepidermal water loss, the pretreatment with 3H or measurement of electrical conductance) will allow the skin barrier to be characterised and for particular samples to be discarded if they fall outside acceptable ranges of use. In addition, individual researchers may prepare skin samples differently depending on specific needs or to adhere, for example, to protocols used in previous similar studies. This will include preparing the skin by removing all or part of the tissue underlying the epidermis (the dermis and/or the subcutaneous fat), which is normally achieved by the application of heat, with the use of enzymes or mechanically, for example, with a dermatome. Whether the skin used for a permeation study is used fresh or used following freezing and storage may also impact on the permeability measured.
Chapter 2 introduced the concept of diffusion cells for in vitro measurement of permeation across biological membranes, including skin. Most cells are, essentially, variations on a theme, with the majority of diffusion cells currently in use being mostly based on designs by Franz (1975). Thus, they will have two compartments, and the permeant’s ability to pass from one compartment (the donor) to the other (the receptor, which is normally stirred to ensure uniformity) via a membrane (such as mammalian skin or an artificial membrane), is determined. Cells are normally arranged vertically or side by side (where both compartments are usually stirred). Stirring is essential in order to avoid or minimise the occurrence of static diffusion layers within the diffusion cell and to avoid high local concentrations of permeant which may influence the diffusive process, affecting the results of a study (Stehle and Higuchi 1972; Lovering and Black 1974). The compartments will usually be maintained at a specific temperature which is predominately, but not exclusively, 37 °C (+1 °C) in the receptor compartment which equates to 32 °C (+1 °C) at the skin surface (Barry 1983). Inefficiencies in temperature control, even to a very small degree, can significantly influence the permeability measured in an experiment (Friend 1992; Chilcott et al. 2005), leading to a substantial source of error if such data are subsequently collated and used for model development.
The nature of the quantitative analytical method may also influence, albeit minimally, accuracy of measurements in skin permeability studies. Methods commonly used include chromatography, spectroscopy, scintillation counting and specific assays which may be biological in nature (Nugent and Wood 1980). For non-radiolabelled techniques applied to, for example, cosmetic systems, a lack of specificity in penetrants (such as triglyceride-containing natural oils or proteins which are found in some topically applied cosmetic products) may be an issue, particularly when they are derived from natural sources which may contain a range of materials which may not necessarily all be at the same concentration in all samples and which may also be similar to material found naturally within the skin. While more recent analytical techniques, particularly those incorporating mass spectrometry, may reduce or eliminate such considerations, the context of collating data from the literature may include doing so from experiments where such analytical issues were not, or could not be, addressed. Thus, care in collation of the data set for development of a mathematical model of skin absorption should extend to an understanding of the analytical processes used in the original studies from which the data were collated.
Where extraction of a material from a biological matrix forms the part of an analytical procedure, this may result in incomplete or variable recovery. One generalisation often made in the collation of skin permeability data appears to be the potential for experimental variation, particularly when conducting bioanalytical studies; factors such as stability, storage, inter- and intra-day variation are seldom accounted for, and the general assumption has been that samples can be measured properly and consistently, when in fact such analyses are prone to variance. Ideally, it would be for the benefit of models produced for the data they use to be taken from guidelines, such as the FDA guidelines for bioanalytical method development. This may extend to consideration of the nature of the solvents used in both donor and receptor compartments, solute solubility in those compartments and the effects they may exert on the skin barrier function (i.e. in the case of solvents which may, for example, delipidise the skin) or on solubility of potential permeants in the receptor compartment.
Composition of the receptor phase can thus influence solute solubility and therefore may affect the nature of the permeation process being investigated (Barry 1983; Bronaugh and Maibach 1999; Williams 2003). It is common, in measurements of in vitro percutaneous absorption, to modify the receptor composition in order to facilitate diffusion of a permeant into this compartment. It should be noted that the general models from Flynn onward are based on infinite-dose experiments also consider diffusion from saturated aqueous solutions only, as Flynn commented that it is from such vehicles that the bulk of environmental exposures would occur. Thus, in vitro systems where the receptor compartment contains a solvent other than water (or a simple buffer) are not normally considered for inclusion in data sets from which mathematical algorithms are produced.
The composition of the receptor phase will therefore be highly significant in influencing percutaneous absorption, not just by its volume and mixing but by its chemical composition. For example, Kasting et al. (1987) used a receptor phase solution which contained 50 % v/v ethanol as well as a small amount of sodium azide (0.02 % w/w) as an antimicrobial agent. This receptor phase was chosen as it would ensure solubility of their penetrants in this compartment. Bovine serum albumin (BSA) has been added to receptor compartments, in concentrations as high as 4 % w/w, by many researchers, including Dal Pozzo et al. (1991) and Sartorelli et al. (1998) as it assists in the solubilisation of penetrants with a wide range of lipophilicities. Lin et al. (1996) used a receptor phase buffered at pH 6 to ensure the solubility of their permeant, which is pH dependant, throughout their experiment, and Mistry et al. (2013) employed a citrate buffer to ensure that their permeant, aluminium, was solubilised throughout their experiment as this chemical forms an insoluble oxide at neutral pH.
Such modifications to experimental protocols clearly move them further from the in vivo ideal, making in vitro–in vivo correlations more difficult to quantify, but also clearly signposts that these experiments are being carried out in vitro and should be understood in this context. A range of other factors should be considered in the context of developing in vitro models of skin absorption. These include the effect of the systemic circulation and biological processes such as active transport or metabolism. Somewhat idealistically and impracticably, it may be proposed that the most representative receptor fluid may be either human or animal blood (or plasma), although it may have to be modified to include anticoagulants and its use may raise particular handling concerns (Moss et al. 2012).
Therefore, despite commonality in generic experimental design, and attempts such as those by Chilcott et al. (2005) to develop common approaches, little standardised methodology has been adopted for the measurement of percutaneous absorption. Given the main reason for conducting such studies is not often related to developing mathematical models of permeation, this limits significantly the amount of data available in the literature from which models can be developed. It also limits the usefulness and scope of models so developed as they may be of little relevance to other systems.
Considerations relating to finite-dose experiments have been discussed previously in Chap. 8.
Analysis of the Experiments from Which Data Have Been Taken to Develop Models of Skin Absorption
The preceding section outlines the range of experimental factors that can be used in a seemingly simple Franz-type diffusion cell-based experiment. The excellent review by Friend (1992) describes in detail the range of experimental techniques available for such seemingly straightforward measurements. This experimental variance is something which underpins the data used to develop quantitative models of skin permeation and highlights the need to examine the source data used. This is again underpinned by the Scheuplein data added to the Flynn data set, which varied considerably from findings from a number of other researchers (Michaels et al. 1975; Hadgraft and Ridout 1987; Goodman and Barry 1988; Hou and Flynn 1989; Liu et al. 1991; Williams et al. 1992; Knutson et al. 1993; Yum et al. 1994; Mitragotri et al. 1995). Of the fourteen steroids examined by Scheuplein, six (aldosterone, corticosterone, estradiol, hydrocortisone, progesterone and testosterone) have measured permeabilities lower than those reported by other researchers by factors of between 5 and 77. For eight steroids, the published permeability data are all in good agreement with each other but vary from the Scheuplein data by factors of between 11 and 20. It is also interesting to note that in Johnson’s evaluation of these data, they considered potential experimental factors, such as those outlined in the previous section, but concluded that such factors as experimental temperature, establishment of steady-state conditions, methods of skin preparation and the use of radiolabelled permeants did not influence the studies and that the data are incorrect.
One point often not highlighted in the commentary of the Scheuplein data is that while other researcher’s permeability data are in agreement, this essentially means that it is within one order of magnitude. This is still a significant margin of error and indicates that even controlled studies which offer a broad consensus do so within the wider context of methodological and membrane variability which is difficult to control or minimise, despite substantial efforts in trying to harmonise methodology, including skin barrier preparation. It is also reflected in the statistical approach taken in preparing data for analysis. In general, it might be considered that all data—unless reasonably determined to be unsuitable—should be included for analysis. Whether this is then examined as a mean or median value of “repeated” data (i.e. permeability data for the same chemical from different experiments) or whether each individual data point is used in constructing a data set is an important point to be considered, as it is often a point that is irregularly presented with sufficient transparency—if at all—in some studies.
It is also insufficient to simply collate different values and then use the resultant mean or median. For example, in the steroid data summarised by Johnson et al. (1995), they showed that the different experiments summarised in Table 4.1 were conducted at different temperatures of 25, 26, 30, 32 and 37 °C. It was argued by Johnson et al. that these data could be temperature-corrected using the Arrhenius equation, resulting in only a twofold change in skin permeability. While such corrections clearly often improve the consistency of data, they still demonstrate a significant variance in the data and, it might be argued, a lack of clarity in presenting this point in their study so as to avoid the use of data obtained at different experimental temperatures without further adjustment. More broadly, it is often unclear from published studies whether the temperature quoted relates to the membrane temperature (i.e. the skin surface temperature) or the temperature in the receptor compartment of the diffusion cell.
The points raised above have resulted in some researchers challenging the very concept of developing mathematical models of skin absorption from experimental data. Not only are such models limited by the formulation type applied to the donor compartment, they are beset also by variable data which are difficult to validate (e.g. Scheuplein’s original, disputed data have a variance of approximately 25 %, which generally meant that skin researchers felt that it was reasonable given the inherently variable nature of the barrier being used), they may not properly address the variability in their output, and they are prone to misinterpretation and misuse.
For example, Walters and Brain (2000) commented on the lack of relevance of QSAR-type models of skin absorption, which followed the comments outlined above. In particular, they highlighted where such models were not relevant by considering a range of permeant/vehicle interactions and choices of solvents used in experiments from which models are developed. However, rather than commenting on the lack of relevance of such models, it may be argued that they addressed the issue of model perception and misuse, highlighting the limitations of a range of models rather than their redundancy. As an example, they commented that as most Potts and Guy-type models for skin absorption relate to permeation from aqueous solutions, any predictions or mechanistic insight would be relevant for such systems only. This is indeed the case, and both the reasons for the use of such a system and its potential limitations were highlighted in Flynn’s original work. What Walters and Brain do not offer is a solution to these limitations, but this is an issue addressed by others, which will be discussed below. So, while they cited earlier work of theirs (Walters et al. 1997) which highlighted the differences between formulations and how this has a significant effect on permeation and flux, this essentially highlights that modelling of percutaneous absorption is still a relatively new field of research and that it is one which has led to significant insights to the wider issue of skin permeation—this is seen from the outcomes of many studies discussed in Chaps. 4–6 in particular and also in the formulation-focused studies discussed in Chap. 8. However, the misuse or misinterpretation of models is the real underlying issue in such studies. The mathematical algorithms developed for skin permeation are limited to their input data and while significant criticism is valid—for example, that which relates to data quality, variance in experimental results and membrane variance—other criticisms of models seem to be informed from an incomplete knowledge of model construction and validation, as well as a lack of understanding of their domains of applicability. Nevertheless, the issues raised by Walters and Brain are significant in the context of formulation and solvent effects.
Systematic studies examining models of skin absorption were conducted previously (Moss et al. 2005, 2006). They synthesised and characterised a series of new prodrugs of captopril, an ACE inhibitor whose potential as a transdermal therapeutic is based on the maintenance of a steady blood plasma concentration over a fixed period of time. They modified the parent drug at the thiol and carboxyl groups, which had a wide range of lipophilicities and molecular weights—all of which were still within the “molecular space” of the Flynn’s data set. Predictions of permeability were obtained from a number of QSAR-type models and were compared to in vitro Franz-type diffusion cell studies using a variety of membranes, including fresh and frozen porcine skin and a polydimethylsiloxane membrane.
The findings of both studies are summarised in Fig. 9.1. This compares experimental permeabilities for a series of captopril carboxyl ester prodrugs with predictions from a number of QSAR-type algorithms. While the model is limited in the nature of the comparison—given the species used experimentally—it demonstrates several fundamental issues with the use of QSAR-type models of skin permeation. The first clear issue is that despite any limitations due to the membrane used, the trends for measured permeability in skin are, in most cases, significantly different from the predictions in a number of ways. They initially show reasonable agreement between experimental and predicted permeabilities for the less lipophilic prodrugs, generally up to a point where the log P of the permeant is at, or below, 2.0. However, above this point, the models begin to differ significantly from the experimental findings, which are a set of results that empirically agree with our understanding of skin permeability outside the context of algorithm development. Specifically, the experimental results show an optimum permeability (generally around the C2–C4 prodrugs), followed thereafter by a decrease; similar conclusions have been drawn in the earlier chapters in the context of nonlinear modelling. So, while the general trends between experimental and predicted permeability data do not compare across the stated range of validity for the prodrugs examined, another significant issue which is clearly apparent in Fig. 9.1 is the overestimation of permeability by most models compared to experimental findings. This may be due to a number of factors—for example, issues with skin thickness may influence such direct comparisons—but the scale of overestimation is significantly larger than any such experimental factors, which in themselves are not considered by the models as they are based on data from a range of diverse experiments. Overestimation is at its worst for the prodrugs with the highest lipophilicity.