A comprehensive review on non-clinical methods to study transfer of medication into breast milk – A contribution from the ConcePTION project Biomedicine Pharmacotherapy

variable milk lipid content) are not accounted for by these methods. Physiologically-based pharmacokinetic (PBPK) modelling offers a mechanism-oriented strategy with bio-relevance. Recently, lactation PBPK models have been reported for some medications, showing at least the feasibility and value of PBPK modelling to predict transfer of medication into the human milk. However, reliable data as input for PBPK models is often missing. The iterative development of in vitro , animal in vivo and PBPK modelling methods seems to be a promising approach. Human in vitro models will deliver essential data on the transepithelial transport of medication, whereas the combination of animal in vitro and in vivo methods will deliver information to establish accurate in vitro/in vivo extrapolation (IVIVE) algorithms and mechanistic insights. Such a non-clinical platform will be developed and thoroughly evaluated by the Innovative Medicines Initiative ConcePTION.

variable milk lipid content) are not accounted for by these methods. Physiologically-based pharmacokinetic (PBPK) modelling offers a mechanism-oriented strategy with bio-relevance. Recently, lactation PBPK models have been reported for some medications, showing at least the feasibility and value of PBPK modelling to predict transfer of medication into the human milk. However, reliable data as input for PBPK models is often missing.
The iterative development of in vitro, animal in vivo and PBPK modelling methods seems to be a promising approach. Human in vitro models will deliver essential data on the transepithelial transport of medication, whereas the combination of animal in vitro and in vivo methods will deliver information to establish accurate in vitro/in vivo extrapolation (IVIVE) algorithms and mechanistic insights. Such a non-clinical platform will be developed and thoroughly evaluated by the Innovative Medicines Initiative ConcePTION.

Introduction
About 50 % of postpartum women need medication, and the evidence on the beneficial effects of breastfeeding is growing [1]. Yet, data on the safe use of medication during breastfeeding is lacking for a large share of medicines [2]. This knowledge gap often forces women either to postpone a much-needed treatment or to discontinue breastfeeding at the advice of their healthcare professionals. Strong scientific evidence could prevent this unnecessary choice. In fact, many medications are likely to be safe as the transfer into the breast milk and/or neonatal gastrointestinal absorption are expected to be low for a majority of medications. However, clinical lactation studies are costly, time-consuming and encounter many practical and ethical limitations. Therefore, validated non-clinical research methods could play a crucial role in filling the knowledge gap in this field.
This comprehensive review aims to provide a state-of-the art of nonclinical (in vitro, in vivo and in silico) methods to determine transfer of medication during lactation.

Methods
A comprehensive ("non-systematic") review was performed. Five distinct literature searches (see Supplemental Data S1: Search Terms) for in vitro models, in vivo animal models, empirical and semi-mechanistic models, PBPK models and the effect of maternal conditions on the macro-nutrient composition of breast milk were performed searching PubMED between September 2019 and December 2019. In addition, Embase was searched for the in vitro and in vivo models. Articles were excluded if no full text was available or if they were not written in English. The selection of the articles for the in vitro models and PBPK models was performed using Rayyan [3].

In vitro animal models
Articles regarding in vitro models for the blood milk epithelial barrier, using cell lines or primary cells, were included.

In vivo animal models
Articles on animal models to predict human breast milk exposure or human neonatal systemic exposure to maternal medications via breastfeeding were included.

Empirical and semi-mechanistic models (human)
Articles on empirical or semi-mechanistic models to predict human breast milk exposure were included.

Physiologically-based pharmacokinetic (PBPK) models
Articles about PBPK models to predict human breast milk exposure or neonatal systemic exposure to maternal medication via breastfeeding were included. A review on the use of PBPK modelling regarding lactation was conducted previously [4]. However, at that time (in 2003) PBPK models were only available for chemical substances or for the dairy industry.

The effect of maternal conditions on macro-nutrient composition of breast milk
Articles about the effect of maternal conditions on the macronutrient composition of breast milk were included.

Available in vitro models for the mammary epithelium
In vitro cell culture models for the mammary epithelium have been established to predict medication partitioning into the breast milk, based on the assumption that the mammary epithelium ( Fig. 1) is the main barrier between the systemic (maternal) circulation and the milk. Many cell culture models have been established based on human or animal mammary epithelial cell cultures, including both primary cells and cell lines (Table 1). These cell culture models can be used to study not only passive, but also active transport of medication across the blood milk barrier. In 2006, the first human model to predict medication transfer into the human breast milk was developed by Kimura et al. [6]. They used the method from Schmidhauser et al. [7] to obtain trypsin-resistant cells, which have the ability to differentiate into the lactating state. More recently, Andersson et al. [8] developed a model to evaluate the transfer of the neurotoxic amino acid beta-N-methylamino-alanine (BMAA) into the breast milk based on its uptake in the human mammary MCF-7 cell line. Other mammary cell lines have been used to investigate the mammary gland (e.g., PMC42-LA [9]) and in the field of breast cancer (e.g., R5, MCF-7, MDA-MB-231-LUC, MCF10A and primary epithelial cells [10]). In addition, MDCK-II cells transfected with human Breast Cancer Resistance Protein (BCRP) have also been used to optimize predictions of milk-to-plasma (M/P) ratios for medications [11]. Many in vitro models relying on mammary epithelial cells have been established based on cells obtained from animal tissue.

Culture conditions
Mammary epithelial cells are either obtained via isolation from normal breast tissue [25], tumor breast tissue [25] or breast milk [26]. Currently, some cell lines can also be obtained from commercial cell suppliers. Mammary epithelial cell basal medium (MEBM) (Table S1), with addition of several supplements (Table S2), is recommended for the growth of HMEC [6]. RPMI 1640, DMEM and DMEM:F12 (50:50) have been used for human mammary epithelial cell lines and cells from animal origin [8]. Basal RPMI 1640 and DMEM:F12 (50:50) are not specific for epithelial cells. However, the addition of particular supplements makes them suitable for the growth of mammary epithelial cells. Different supplements concentrations (Table S2) are recommended by the suppliers and in previously mentioned in vitro culture models. The choice of supplements and in particular, prolactin, is strategic when working with a model of secreting cells. For instance, Freestone et al. [9] were able to make a model for the resting, lactating and suckled mammary epithelium, by adding either no, 200 ng/ml, or 800 ng/ml prolactin, respectively, to the PMC42-LA cell line.

Characterization
Characterization of the in vitro models, especially of the transporters (Fig. 2), is important to ensure that the model is applicable to the in vivo situation for a specific lactation period and a given compound. An overview of the available information regarding transporters has recently been given by Ventrella et al. [38].

Human in vitro models to predict medication transfer into the breast milk
Two human models have currently been reported for the prediction of medication transfer into the breast milk. Kimura et al. [6] used HMEC: they were able to obtain a monolayer with a transepithelial electrical resistance of 227 +/-11 Ohm. cm 2 after three trypsin treatments, indicating a tight monolayer. They detected beta-casein mRNA, which demonstrates that the monolayer is differentiated into the lactating state. Shipman et al. [39] showed that beta-casein is only expressed in the lactating state. Besides beta-casein, Kimura et al. [6] also detected mRNA of organic cation transporter (OCT)1 and OCT3. Interestingly, they found that OCT1 mRNA increased, whereas OCT3 mRNA decreased with increasing number of trypsin treatments. This observation was consistent with the observations of Alcorn et al. [29], who observed that OCT1 mRNA levels are higher, while OCT3 mRNA levels are lower in vivo in the lactating state compared to the non-lactating state. Kimura et al. [6] also investigated the function of OCT and organic anion transporter (OAT) with the substrates tetraethylammonium (TEA) and p-aminohippurate (PAH), respectively. A clear directionality was observed for TEA, indicating that functional OCT is present. However, no directionality was observed for PAH. Other transporters that have not been investigated in this study, for instance BCRP, play an important role in vivo. Therefore, further characterization of this in vitro model is required in order to conclude whether this is a good model to evaluate medication transfer into the human breast milk.
Andersson et al. [8] developed a human in vitro model to investigate the transfer of D-BMAA and L-BMAA into the human breast milk based on their uptake in MCF-7 cells. MCF-7 (Fig. 3) is a cell line derived from a metastatic site of an adenocarcinoma, via pleural effusion [40]. MCF7 expresses estrogen and progesterone receptors and is known to have some characteristics of the differentiated mammary epithelium [40]. Andersson et al. [8] found that uptake was higher in the differentiated model. Furthermore, via inhibition studies with natural amino acids, they concluded that several amino acid transporters might be involved in the uptake. Additionally, they found some differences in mRNA levels of orthologous transporters in the MCF-7 cell line compared to the mouse HC11 cell line. Finally, they compared mRNA expression in undifferentiated and differentiated HC11 cells and found that mRNA increased for some transporters, whereas mRNA decreased for others.
Besides MCF-7, many other human cell lines have been used to investigate the mammary gland. For example, MCF-10A has been used commonly as a model to investigate normal breast cells. MCF-10A is an immortalized, non-tumorigenic cell line obtained from benign proliferative breast tissue [41]. MCF-10A does not express estrogen or progesterone receptors. Furthermore, no beta-casein or alfa-lactalbumin were detected [41]. In addition, Ying Qu et al. [41] questioned whether MCF-10A cells are a good model for normal breast cells. They conclude that further investigations are required.
Another frequently used cell line is the MDA-MB-231, which was obtained via pleural effusion of a patient with metastatic mammary adenocarcinoma [42]. The MDA-MB-231 cell line does not express an estrogen or progesterone receptor. Moreover, MDA-MB-231 might not be suitable as a model for normal lactating mammary epithelial cells, as it is a highly aggressive, invasive and poorly differentiated cell line that is mainly used to investigate triple-negative breast cancer.
Finally, PMC42-LA is an epithelial cell line derived from PMC42, a mesenchymal breast carcinoma cell line that has been obtained from a pleural effusion. Although PMC42-LA has not been used as frequently as the previously mentioned cell lines, it might be a good model for the lactating mammary gland. In fact, Freestone et al. [9] were able to develop a resting, lactating and suckling in vitro model with this cell line by using different concentrations of prolactin. They indicated that the capacity to differentiate into a lactating state is a major advantage of this cell line compared to many other cell lines. It has also been shown that PMC42-LA cells express beta-casein after stimulation with lactation hormones [9]. The mammary epithelium consists of lobes, containing alveoli and milk ducts. The luminal epithelial cells or acini cells secrete milk into the lumen of the alveolus. The surrounding myoepithelial cells contract under the stimulation of oxytocin and cause excretion of the milk into the ducts. Transport of endogenous compounds as well as medication is possible via passive or active transport mechanisms. Transcellular transport implies crossing of cell membranes, whereas paracellular transport occurrs via the confined spaces between cells. 3.1.5. Animal in vitro models to predict medication transfer into the breast milk Prediction of medication transfer into the human breast milk based on animal in vitro models might be difficult due to species-specific differences (e.g. differences in enzyme and transporter expression and activity). However, these in vitro models, in combination with animal in vivo models, can play a pivotal role in the in vitro to in vivo extrapolation of human in vitro data, as they can generate fundamental and mechanistic knowledge.
Two main categories can be distinguished: rodent and non-rodent models. Among the rodent cell lines: Rat Mammary Epithelial (RME) cells are derived from normal mammary glands of 50− 60-day-old virgin female Lewis rats [15], while both mouse cell lines (HC11 and CIT3 cells) are derived from COMMA-1D cells, obtained from mammary tissue of BALB/c mice in the middle of pregnancy. CIT3 are selected for resistance to triple trypsinization, while HC11 have been immortalized. Both cell lines have been used for active transport studies of medications such as nitrofurantoin.
Among non-rodent cell lines, Bovine Mammary Epithelial cell line BME-UV is a clonal cell line established from primary epithelial cells from a lactating Holstein cow. This cell line expresses functional markers such as microvilli and desmosomes and secreting properties [43], and expresses functional organic anion and cation transporters [21,22]. Primary goat mammary epithelial cells (pgMECs) derived both from mammary tissue of lactating or non-lactating juvenile goats grow in vitro for several passages and remain hormone and immune responsive with a secreting phenotype (55). Porcine mammary epithelial cells (pMECs) can be derived from non-pregnant and non-lactating gilt [17]. These primary cells can be maintained in culture for at least 15 passages and could represent a good model for studying molecular regulation and synthesis of milk (excretion). Alternatively, porcine mammary epithelial cells can be obtained from mammary gland after parturition [18].
Rodents are usually considered an excellent model in many fields of research, but their metabolic and digestive patterns and milk composition are quite different from humans, with significant differences in medication levels reached in blood and the potential lactational transfer [56][57][58]. Indeed, in such species, the general aspects of reproduction (age of sexual maturation, hormone sensitivity, reproductive lifespan, litter size) are very different when compared to humans [49]. Nevertheless, most studies clarifying the development of mammary cancer have been performed in rodents due to their relatively easy manipulation and housing requirements [59,60]. An additional issue with working with rodents relates to the body dimension. More specifically, rodents allow only for small volume milk sampling limiting the possibility of end point analysis.
Medication transfer from blood to milk has been extensively studied in ruminants [46,61], mainly for human safety reasons, due to their role as food producing animals (in particular milk and dairy products). Nevertheless, these animals differ significantly from human from an anatomical point of view and mainly, in absorption and metabolic capacity due to their peculiar gastrointestinal physiology.
Swine offer a generally accepted model in translational medicine based on their anatomical and physiological similarities with humans. Its use as a model for nutritional physiology, medication testing and metabolism has been generally acknowledged [45,52,[62][63][64][65]. Mammary gland anatomy shows macroscopic differences but, at a molecular Transporters detected either in primary human mammary epithelial cells from commercial suppliers and/or in mammary epithelial cells isolated from milk or breast tissue. The arrows indicate the direction of the transport. If no evidence was found for the location of the transporter in the mammary gland, the transporter was placed on the expected location based on locations in other tissues (indicated with asterisk). Furthermore, transport protein expression was also found for ABCA7, ABCA13, ABCF2, ABCF3, SLC1A4, SLC1A5, SLC3A1, SLC4A2, SLC4A7, SLC5A3, SLC7A1, SLC7A2, SLC9A1, SLC9A3R1, SLC9A3R2, SLC12A1, SLC12A4, SLC12A9, SLC15A4, SLC16A2, SLC16A3, SLC20A2, SLC26A2, SLC27A1, SLC27A2, SLC30A1, SLC31A1, SLC35E1, SLC39A14, SLC44A2 & SLC52A2. However, these transporters were not included in the figure since they are currently not known to play an important role in the transport of medication.
level, the presence of medication transporters similar to human has been reported (See Table 2 from Ventrella et al. [38]). However, the lactation duration is shorter, compared to humans. These points have to be considered when studying the lactational transfer in the different phases of lactation, especially for colostrum composition [55]. The litter size allows for an easy milk collection without interfering with the lactation process and piglet's growth; piglets can be analyzed individually and have been already utilized as animal model for pediatric PK/PD [66]. In recent years, the minipig, and in particular Göttingen Minipigs, has been proposed as a more relevant animal model in translational medicine in particular for toxicological studies [67,68]. They show, together with all the positive characteristics described for domestic swine breeds, a reduced growth rate that makes them easier to use. Moreover, since some minipig breeds, such as the Göttingen one, are specifically bred and produced for experimental purposes, they allow for a much more precise genetic and microbiologic standardization, as well as a standardization of the interactions between animals and the housing context, including operators, leading to a significant reduction in the number of animals needed for experimental trials [69,70]. The fact that Göttingen Minipigs are available in a constant and uniform quality worldwide forms the basis of their recognition by regulatory bodies as a suitable non-rodent species for toxicology studies. However, data regarding qualitative and quantitative composition of the milk, as for the domestic pig, are lacking.
The most relevant papers on animal models of lactational transfer are listed in Table 2with indication of the relevant applications and the species studied.

Empirical and semi-mechanistic models (human)
Purely empirical models describe the correlation between data, while mechanistic models (including PBPK models) also account for the underlying physiological processes. Semi-mechanistic models lay between the empirical models and the mechanistic models. For some aspects they rely on physiologically relevant mechanisms, whereas other aspects of the model are not physiologically relevant.
Several attempts have been made to predict the transfer of medication into human milk using different physicochemical parameters. In 1959, Rasmussen et al. [71] first assumed pH-dependent diffusion of medication, while Notarianni et al. [72] developed an equilibrium dialysis model to test the partitioning of medication between freeze dried plasma and baby formula powder over a dialysis membrane. Other diffusion models have been developed by Atkinson et al. [73] and Fleishaker et al. [74]. Meskin and Lien [75] each developed an in silico model based on the relation between physicochemical properties (the molecular weight, partition coefficient and degree of dissociation) and the transfer of medication into the milk. This model was later extended with an artificial neural network by Agatonovic-Kustrin [76]. Quantitative structure activity (or property) relationship tools have been explored as well [77]. Many others have tried to predict milk transfer using similar methods. However, the major limitation with all of these methods was that they do not take transporter-mediated processes into account.
the milk was compared to the fraction of unbound medication in the plasma for each medication to determine whether passive diffusion or transporter-mediated transfer is the most likely route for medication transfer into the human breast milk. For the medications with passive diffusion as main pathway, an equation describing the relation between the physicochemical properties and the secretion and reuptake clearance values was determined by multiple linear regression. This semi-mechanistic model was applied to determine M/P Area Under the Curve (AUC) ratios for 71.9 % of the 49 medications, be it within a 3-fold error compared to observed values [78].

Physiologically-based pharmacokinetic (PBPK) modelling
A PBPK model is defined by the European Medicines Agency as "a mathematical model that simulates the concentration of a medication over time in tissue(s) and blood, by taking into account the rate of medication Absorption into the body, Distribution in tissues, Metabolism and Excretion (ADME) on the basis of interplay between physiological, physicochemical and biochemical determinants" [81].
PBPK modelling is for instance often applied to predict drug-drug interactions and to select an initial dose for pediatrics and first-inhuman trials [81]. Furthermore, PBPK modelling can be used to predict transfer of compounds into the breast milk, and subsequent neonatal systemic exposure. Research in this field has focused on long lasting, bio-accumulative substances (e.g. trichloroethylene) in the field of toxicology, and milk transfer in animals providing milk for human consumption [4].. More recently, some PBPK models for the prediction of transfer of medication into the human breast milk and subsequent neonatal systemic exposure, further referred to as lactation PBPK models, have been reported (Table 3). Five articles and four conference abstracts about lactation PBPK models were retrieved, all within the last decade. The reported lactation PBPK models consist of a maternal PBPK model coupled to a neonatal PBPK model (Fig. 4) allowing them to predict milk transfer and neonatal systemic exposure via breastfeeding. The PBPK models for alprazolam, caffeine and tramadol only consist of a maternal PBPK model, and can thus only predict milk transfer [82,83]. The model for escitalopram is a combination of population pharmacokinetics (popPK) to analyze human breast milk data and PBPK modelling to predict infant exposure [84]. These five cases will be discussed in the next sections.
The goal of the lactation PBPK models was to predict the exposure of neonates to maternal medication via breastfeeding. In the case of codeine, the focus was on differences in exposure due to maternal and neonatal differences in CYP2D6 genotype and therefore, morphine formation [86]. The PBPK model for isoniazid also aimed to investigate the impact of the polymorphic N-acetyltransferase-2 [89]. No genotype specific simulation was performed for escitalopram, rifampicin or ethambutol [84,88]. In the case of efavirenz, CYP2B6 polymorphism is known to have an impact on the metabolism. Olagunju et al. did not perform genotype specific simulations but used reported data for  pediatric CYP2B6 protein expression in which they induced variability.

Physiologically-based pharmacokinetic (PBPK) model structure
Transfer of medication into the human breast milk can be modelled in several ways [89]: (i) direct transfer of the medication from the blood into the human breast milk [91]; (ii) via uptake into the breast adipose tissue [92]; or (iii) a combination of both routes [93].
All three approaches have been used for chemical substances. The currently available lactation PBPK models for medication all used the first approach as described below. Distinct software packages have been used to develop the existing models.
The lactation PBPK models used a whole-body PBPK modelling approach. They combine a maternal PBPK model including a breast compartment with a neonatal PBPK model. The model for escitalopram combines popPK on maternal monitoring data with a neonatal PBPK model [84]. Willmann et al. [86] combined four PBPK models: maternal PBPK models for codeine and its main active metabolite morphine, and neonatal PBPK models for codeine and morphine.
Different approaches are used to couple the maternal and neonatal models. Delaney et al. [84] first analyzed the escitalopram concentrations in the breast milk using popPK. In a next step, they calculated daily infant doses using a random combination of the predicted milk escitalopram concentrations, milk volumes per feed and frequencies of feeding. The calculated daily infant doses are then administered to the neonatal PBPK model as a single dose. Willmann et al. [86] used a similar approach, but they administered the medication to the neonatal PBPK models as multiple doses. They assumed breastfeeding, and thus dosing the neonate to take place each 3 h. The doses were calculated using the medication concentrations at the time of breastfeeding predicted by their maternal PBPK model and the breast milk volume. Willmann et al. assumed that the absorption of the medication in the neonatal model is fast and complete. Furthermore, Olagunju et al. [87] calculated the milk concentration by multiplying the M/P ratio with the simulated plasma concentration. Subsequently, they calculated the infant dose per breastfeeding session by multiplying the milk volume with the milk concentration.
Garessus et al. [89] used another approach, assuming that the breast is completely emptied during each feed. This means that the dose given to the neonatal PBPK model is equal to the amount of isoniazid in the breast milk at the time of breastfeeding. Dosing of the neonatal model is repeated every two hours. Partosch et al. [88] developed a PBPK model for ethambutol and a PBPK model for rifampicin. The PBPK models for both medications have a similar structure. They included the breast compartment in the maternal PBPK model as a reservoir. Excretion into the reservoir can be calculated by multiplying the milk volume with the milk concentration. The milk concentration is calculated by multiplying the plasma concentration with the M/P ratio. Every 4 h, the reservoir is opened for 30 min, allowing the medication to transfer to the neonate via a milk dose compartment. Table 4 gives an overview of the different breastfeeding parameters that have been used in the lactation PBPK models. A review aiming to  Abbreviations: SD: standard deviation.
quantify breast milk intake feeding parameters for input into PBPK models has recently been published by Yeung et al. [94]. They developed a nonlinear regression equation on the weight-normalized human milk intake. Their results show a maximum intake of 152.6 ml/kg/day and a weighted mean feeding frequency of 7.7 feeds/day for exclusively breastfed term infants. PBPK modelling aims to predict in vivo concentration-time profiles based on: (i) medication-specific parameters; and (ii) physiological parameters.
The medication-specific data required for the development a PBPK model using Simcyp™ (Certara, UK) are summarized in the manuscript of Zuhang et al. [95]. Several sources of input data can be used. Delaney et al. [84] measured in vivo escitalopram concentrations in breast milk from 18 lactating women. They used clearance values from literature, which had been determined in vitro. Age-dependent algorithms were used to scale the parameters for the neonatal PBPK model. Garessus et al. [89] used in vivo data, including an AUC based M/P ratio, obtained from literature. AUC based M/P ratios are preferred over single M/P ratios, since single M/P ratios vary over time. Some of the in vivo data were re-calculated to match their population. Partition coefficients were calculated according to an algorithm from Schmitt et al. [96], which was also used by Partosch et al. [88]. Both Garessus et al. [89] and Willmann et al. [86] used adult clearance values fitted from in vivo data, whereas the neonatal clearances were in vivo values obtained from literature. Willmann et al. [86] used a range of M/P ratios based on several in vivo values reported in literature for both morphine and codeine. Partosch et al. [88] used physiological data from literature. In vivo clearance values were obtained from literature. For ethambutol, the M/P ratio was based on two in vivo data pairs from literature. For rifampicin, an algorithm to estimate the M/P ratio was used, because it was not clear how the measurements were done for the sparse available in vivo data [88]. Olagunju et al. [87] used an in vivo M/P AUC ratio. They used anthropometric values to predict organ weight and blood flows based on a HIV positive cohort of breastfeeding women. For the maternal PBPK model, CYP450 abundances were taken from in vivo data. For the neonatal PBPK model, data from human liver microsomal samples were used.

Evaluation of the physiologically-based pharmacokinetic (PBPK) models
PBPK models should be evaluated for their ability to predict in vivo pharmacokinetic data. All lactation PBPK models exist as a maternal PBPK model coupled to a neonatal PBPK model [84,[86][87][88][89]. The lactation PBPK models, except for codeine [86] and escitalopram [84], used a separate evaluation for the maternal PBPK models and the neonatal PBPK models. The evaluation for the maternal PBPK models was done by comparing in vivo plasma concentration profiles from literature with predicted plasma concentration profiles. Garessus et al. [89] used matched dosing regimens and also compared breast milk concentrations. For escitalopram, medication milk concentration data was coupled to a neonatal PBPK model. They first evaluated an adult PBPK model against in vivo data and then extrapolated it to neonates and again verified this with in vivo data [84]. A bootstrapping technique was used by Delaney et al. to evaluate the adult PBPK model, for which they pre-specified that the PBPK model would be accepted if the mean plasma AUC ∞ of the observed data fell within a 95 % confidence interval of the mean of the predicted data. Olagunju et al. [87] mentioned an acceptance criterion of a 2-fold difference against observed data.
Different approaches were taken to evaluate the neonatal PBPK models. Delaney et al. [84] and Olagunju et al. [87] compared predicted neonatal plasma concentrations to in vivo plasma concentrations obtained after exposure via breastfeeding, while Garessus et al. [89] and Partosch et al. [88] compared their predicted neonatal plasma concentrations with in vivo concentrations after direct oral or intravenous dosing of the neonates. Partosch et al. [88] only evaluated the neonatal model for rifampicin. Delaney et al. [84] also compared four age groups within the first year of life and concluded that the variation was limited for escitalopram. Olangunju et al. [87] also predicted the infant exposure for four age groups. Garessus et al. [89] also simulated a worst-case scenario by implementing breastfeeding at the time of maximal breast milk concentration and using the highest reported individual M/P ratio.
Willmann et al. [86] simulated a situation comparable to a reported fatal case of codeine use during breastfeeding and compared the simulated breast milk and plasma concentrations with the values observed in this case. Willmann et al. (132), Garessus et al. [89], and Partosch et al. [88] all performed a sensitivity analysis. Willmann et al. (132) investigated the effect of different values for maternal and neonatal morphine clearances in the sensitivity analysis. They found that the morphine concentration in the neonate was mainly dependent on morphine (as codein metabolite) clearance by the neonate and the maternal daily dose of codeine. Garessus et al. [89] found that the maternal PBPK model for isoniazid for the fast metabolizers was most sensitive to the isoniazid clearance, the partition coefficient (Kp) of isoniazid between liver tissue and plasma, the dose, the liver organ blood flow and the breast milk volume. Partosch et al. [88] found that the maternal model was most sensitive to the dose, the clearance and the partition coefficient of the liver, whereas the neonatal model was most sensitive to the M/P ratio and the bioavailability in the infant.
One of the assumptions made by the PBPK models is that a general, 'mean' milk composition exists for macronutrients. However, milk composition changes, especially during the first days after delivery, but also during the time course of each feed, differences related to either preterm or term delivery, and the duration of lactation. Milk contains carbohydrates, fat, proteins, vitamins and minerals. Delaney et al. [84] investigated the intra-feed alteration by comparing escitalopram concentrations in foremilk (refers to the milk at the start of a feed, containing less fat) with concentrations in hindmilk (refers to the milk at the end of a feed, containing more fat). They concluded that there was a significantly higher escitalopram concentration in hindmilk, but the impact of the sampling phase for monitoring of the concentration will be negligible. In addition, maternal characteristics might have an impact on the milk composition. An overview on the effect of several maternal conditions on milk composition is presented in Table 5. Overall, the differences in macro-nutrient composition are rather limited but can be considered during modelling for condition specific settings, e.g., diabetes, or when antibiotic concentrations in human milk are collected in women with mastitis as an indication for these antibiotics.

Animal PBPK models
Lactation PBPK models have also been developed for animals. Generally, the animal PBPK models are based on the same concepts as the human PBPK models. One of the advantages of animal PBPK models, in comparison to human PBPK models, is that the systematic collection of in vivo data is easier. However, a limitation is that the knowledge of the animal physiology is limited compared to humans. The animal lactation PBPK models address different types of research questions. Animal PBPK lactation models have been used to address risk assessment questions in food producing animals about residual medication in edible tissues and milk for human consumption (e.g., [138][139][140]). Other animal lactation PBPK models, typically for rodents, aim to get insight into (human) toxicology (e.g. [141][142][143],). Animal PBPK lactation models have also been used as the basis for the development of human lactation PBPK models [91]. In this case, the animal PBPK lactation model is first developed and validated for animals and thereafter extrapolated to humans by interspecies scaling of the physiological factors.

Table 5
Effect of specific maternal conditions (pre-pregnancy existing, pregnancy related or related to logistic settings, environment and lifestyle) on human milk composition for macro-nutrients.

In vitro human and animal models
Several in vitro models using mammary epithelial cells are available to investigate the transfer of medications into breast milk, while other epithelial cell lines derived from different tissue (intestine, kidney) can be used as reference models for the epithelial cell barrier ( Table 6). The in vitro model needs to be 'biorelevant', i.e., representative for the in vivo human physiology. Therefore, HMECs are expected to be a good model for the blood milk epithelial barrier. HMECs have previously been used by Kimura et al. [6], but further characterization is required before concluding whether this is an adequate model for medication crossing over the mammary epithelium.
Human cell lines are an alternative for HMEC. The advantage is that cell lines are generally easier to culture and have a longer life span. The cell lines should be a surrogate for the in vivo mammary barrier physiology. Therefore, MCF-7, MCF-10A or PMC42-LA could be good options. In addition, animal cells and/or cell lines can be explored. This might especially be useful to obtain mechanistic insights and scaling information for in vitro to in vivo extrapolation (IVIVE) and the development of PBPK models. Even if rodent mammary epithelial cells are the most widely studied and provided many biological insights, it is evident that the rodent mammary gland is not fully representative of the human setting. Therefore, other in vitro animal models have been explored, including bovine, goat and porcine. From a physiological, anatomical and metabolic point of view, ruminants provide a model very far from humans, whereas the porcine species is recognized as an excellent model for translational purposes [155][156][157][158]. Among the different in vitro models of mammary epithelial cells available, primary cell cultures offer the opportunity to study the factors that regulate physiologically relevant development of normal mammary epithelial cells under defined conditions.
Overall, it can be concluded that several in vitro models are available which allow for determination of the partitioning of medications over the epithelial blood milk barrier. However, characterization of the cell culture models for this application remains rather limited.

In vivo animal models
Medication excretion in milk during lactation can be successfully investigated utilizing in vivo studies in lactating animals [38,44]. The principal benefits of in vivo animal studies in this field are: (i) the possibility to clarify also the mechanistic aspect of milk/blood barrier; (ii) the possibility to evaluate the influence of various parameters on the rate of medication excretion in milk (milk composition, timing of milking, drug-drug interaction, and different models of excretion even at molecular level); and (iii) the possibility to evaluate the effects of excreted medication or metabolites on the offspring.
The combination of the animal model with an in vitro-based preliminary screening phase may reduce the number of animals needed and Across 9 different countries, protein content in mature human milk Total protein: steady decline from 30 to 151 days of lactation, significantly higher in the second month of lactation compared with the following 4 months (y = 23.251x-0.1554 (g/l), x = lactation days); True protein: steady decline from 30 to 151 days of lactation, significantly higher in the second month of lactation compared with the following 4 months (y = 18.86x-0.1705 (g/l), x = lactation days); Individual amino acid: steady decline from 30 to 151 days of lactation, significantly higher in the second month of lactation compared with the following 4 months. There is a high level of consistency in the protein content and amino acid composition of human milk across geographic locations, with Chile as an outlier. Stage of lactation explained 22.9 and 16.9 % of the variation in total protein and total amino acid concentration [128] High-altitude adapted population (Tibet) Fat: averaged 5.2 ± 2.0 g/100 ml

Empirical and semi-mechanistic models (human)
Koshimichi et al. [78] showed that semi-mechanistic models can be used to predict the transfer of compounds into human breast milk. The main advantage of this semi-mechanistic model over other reported methods to predict medication transfer into the human milk is that the model of Koshimichi did consider that milk and plasma concentration-time profiles do not change in parallel. Koshimichi et al. found that secretion and reuptake values are similar for most compounds, suggesting mainly passive diffusion. However, for some compounds, transporter-mediated secretion or reuptake plays an important role. The model developed by Koshimichi et al. thus allows to distinguish between compounds that undergo passive diffusion into the milk and compounds that undergo transporter-mediated partitioning. However, with a 3-fold error tolerance, there is still room for improvement.
The main advantage of PBPK models over empirical and semimechanistic models is that PBPK models are based on the underlying in vivo physiological mechanisms. This will allow inclusion of transporter-mediated milk secretion or reuptake in medication-specific PBPK models. Therefore, it can be expected that more reliable predictions can be obtained with PBPK modelling.

Human lactation PBPK models
Recently, some PBPK models became available for the prediction of breast milk exposure, and neonatal systemic exposure to maternal medication via breastfeeding [84,[86][87][88][89]. The available models, despite some limitations, show the value of PBPK modelling in this research field. The models illustrate that PBPK modelling can be used to handle several research questions, including breast milk exposure and neonatal systemic exposure via breastfeeding. A major advantage of PBPK modelling is that non-clinical data can be used to predict in vivo PK behavior of medication. This is especially important, given that clinical studies in a vulnerable population like lactating women and their neonates are time-consuming, and give rise to ethical and practical issues. One of the main challenges is that there is currently no breast milk compartment available in either Simcyp™ (Certara, UK) or PK-Sim® (Open Systems Pharmacology). Another important challenge is the need for high quality input data. The knowledge regarding the physiology of lactating women and nursing neonates is growing, but further research in this field is required to optimize existing, and develop new PBPK models. Furthermore, an immense information gap exists regarding the excretion of medication into the human breast milk and subsequent neonatal gastrointestinal absorption. However, information will become available within the course of ConcePTION. ConcePTION is a project funded by the Innovative Medicines Initiative, which aims to reduce uncertainty about the use of medication during pregnancy and lactation [159]. The quality of the input data is critical for the quality of the final PBPK model. The lack of robust data is illustrated by the hurdles that some of the articles had for obtaining the M/P ratio for the respective model medication. Furthermore, AUC-based M/P ratios, which are more reliable than single M/P ratios, are not always available. In the absence of (high quality) clinical data, some of the M/P ratios were estimated Table 6 Overview of selected epithelial cell culture models that may be useful when mimicking the blood-milk barrier.. using the Schmitt et al. algorithm [96]. However, a major limitation of this algorithm is that it does not account for transporter involvement. The guideline on the reporting of PBPK modelling and simulation of European Medicines Agency [81] and the Physiologically Based Pharmacokinetic Analyses -Format and Content Guidance for Industry from Food and Drug Administration [160] should be followed. This will assure the quality of the developed PBPK models and also their suitability for registration purposes. Multiple aspects are important to consider during the development of the PBPK models, as indicated by the available models. For example, some of the lactation PBPK models performed a genotype-specific simulation, whereas others did not take genotype into account. However, the example of codeine shows the importance of genotype specific simulation for certain medications. Codeine has long been considered safe until the death of a neonate was attributed to it [161]. This event has led to caution using codeine while breastfeeding, although it has to be noted that there is some doubt in the literature about whether neonates can develop opioid toxicity from breastfeeding [162]. The PBPK model of Willmann et al. [86] suggests that a combination of worst case factors, including the genotype of mother and infant, could have led to the death of the neonate. Neglecting the genotype might lead to the conclusion that a medication is safe, while this is not the case for all genotypes. Especially since the prevalence of polymorphisms in the infant are not disconnected from the mother and the limited metabolic (glucuronidation) capacity of neonates.

In vitro
Another important aspect is how the transfer of medication into the breast milk and breastfeeding of the neonate is implemented in the PBPK models. The transfer of medication has been modelled as either direct transfer from the blood to the breast milk or as transfer via uptake into the breast tissue. Furthermore, there was some variation in the parameters that have been used to implement breastfeeding in the PBPK models (e.g., duration of breastfeeding, frequency of breastfeeding and daily milk consumption). Recently, Yeung et al. quantified breast milk intake feeding parameters in infants for input into PBPK models [94]. Even though literature shows that the milk composition is relatively constant in several maternal conditions, possible effects of the specific disease population on the milk composition should be kept in mind. Also, Delaney et al. [84] showed that there was a significant difference in medication concentrations between foremilk and hindmilk. The composition of the milk might thus play a role in the exposure of infants to maternal medication via breastfeeding. It is therefore important to understand the factors that influence the milk composition. Delaney et al. also investigated the variation between different age groups within the first year of life. Although the conclusion was that the variation is limited between the age groups for escitalopram, the age might have an important effect on the exposure to some medications, as it has been shown that clearance is dependent on the age of the neonate [163]. Olugunja et al. [87] did take this into account by doing separate predictions for different age groups. All of these factors can also be used to simulate worst-case scenarios.
PBPK models should be evaluated for their ability to predict the in vivo exposure. Delaney et al. [84] and Olagunju et al. [87] were the only ones to use pre-specified acceptance criteria. There is no consensus on which criteria should be used for acceptance of PBPK models as this depends on the purpose, but a 2-fold deviation is often used as default in literature. The guideline for PBPK models [81] indicates that a comparison of the simulated and observed individual plasma concentration-time profiles should be presented as plots and tabulations. Matched predicted and in vivo data should be used.

Animal lactation PBPK models
Besides human lactation PBPK models, several animal PBPK lactation models have been reported. Typically, these animal lactation PBPK models have been developed for dairy animals or rodents. The goal can either be to gain information for the modelled animal species or to translate the information to humans. Translation of animal PBPK information to humans is especially valuable in case in vivo human data are lacking. However, species differences, for example in transporter expression, complicate the direct translation of data from animal PBPK models to humans. Nevertheless, lessons learned during IVIVE-PBPK modelling while relying on animal in vitro data for the blood-milk barrier, followed by comparison with corresponding in vivo data will be instrumental for improving human PBPK lactation models. A similar approach was taken by Parrott et al. to predict oseltamivir disposition in neonates and infants [164]. At least initial estimates for scaling factors for the IVIVE step can be derived in this way. Simcyp™ v18 allows to build PBPK models in the rat, dog, mouse and monkey, while PK-Sim® also supports mini-pig.

Conclusion
The iterative development of a non-clinical platform should allow to predict breast milk transfer, and subsequent neonatal systemic exposure to maternal medication via breastfeeding. First, a human-relevant in vitro cell culture model, representative for the in vivo physiology should be developed and characterized. Transepithelial permeability data, generated with the in vitro model can then be subjected to IVIVE followed by PBPK modelling. Essential scaling information for IVIVE can be generated by the paired interpretation of animal in vitro and in vivo models. Furthermore, the in vivo animal models can deliver key mechanistic insights to support the physiological plausibility of the PBPK models. The non-clinical tools should be evaluated for their predictive performance using diverse model compounds for which in vivo data are available. The iterative development of several non-clinical tools will ultimately lead to robust predictions of breast milk transfer and neonatal systemic exposure to maternal medication, for which data are currently lacking, ultimately driving a paradigm shift in the domain of pharmacotherapy during lactation.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.