ADME Profiling in Drug Discovery and a New Path Paved on Silica

3.1 Aqueous solubility and lipophilicity

As already mentioned, ionizability is one of the most important properties in PK, thus making pKa the physicochemical property with the highest impact.

Early attempts to increase the efficiency of pKa evaluation were reported by Morgan and colleagues by scaling down the classical titration and spectrophotometric methods introducing microscale versions [43].

These alterations, however, could not overcome the principle demands of each technique, which are moderate precision and frequent calibration (potentiometric), and the need for a chromophore within the analyte (spectrophotometric) [44]. Starting in 1998, capillary electrophoresis (CE) was effectively used to determine pKa of many compounds and was further upgraded from Pfizer by implementing pressure-assisted capillary electrophoresis (PACE) as a standard method, which is nowadays readily applied in industry settings showing superior features compared to the aforementioned methods [44, 45]. Other variants such as vacuum-assisted multiplexed capillary electrophoresis also exist (VAMCE) [46]. A different approach better suited for HTS is called pH gradient titration offered from Sirius Analytical Instruments but is still limited due to the UV spectroscopy technology [30].

It is well established that solubility in aqueous media is one of the most important physicochemical properties to be evaluated in oral administration. It is not only necessary for absorption in the GI tract but also a requirement for almost all in vitroand in vivoassays, which depend on a solved compound. Poor solubility can affect the reproducibility of assay results by introducing high variability and further increase development costs of leads with low solubility [26, 47]. Traditionally, solubility measurements were conducted via labor-intensive potentiometric techniques [48] or equilibrium solubility (thermodynamic; e.g., shake flask) [26]. HTS alternatives comprise laser nephelometric scans (kinetic) [47] and LC-MS/HPLC techniques, which can also be performed with DMSO solutions of the compound—the standard for HTS applications [47, 49]. It should be noted, though, that aqueous solubility, as described above, is not an optimal model for GI solubility since it does not consider the composition of the GI fluids [49].

On the other hand, generally speaking, lipophilicity is the ability of a compound to dissolve in lipids and/or organic solvents thus being able to pass biological membranes. Descriptors for lipophilicity are the logarithm of the partition coefficient (logP) or distribution coefficient (logD). Classically, logPwas determined using the shake flask method applying n-octanol/water phases. Later, UV spectroscopy became the standard, which unfortunately is not applicable for compounds without absorption in the UV range [50]. Today, RP-HPLC methods are frequently in use due to superior properties [25, 51]. As with many methods, comparison of results obtained under different conditions and in different laboratories proves to be difficult with RP-HPLC. A solution offers the implementation of a standardized lipophilicity value, for example, the chromatographic hydrophobicity index (CHI).

In recent years, a great effort has been made to improve the ability of in silicomodels to accurately predict aqueous solubility. One of the most developed model is Yalkowsky and Jain’s [52] general solubility equation (GSE), which is based on the melting point (m.p. °C − 25) and logP(the octanol-water partition coefficient of the un-ionized molecule) of a chemical substance (Eq. (1)), with a relevant prediction power as represented by the coefficient of determination (R²) = 0.96 and root-mean-square error (RMSE) = 0.53 in a dataset of 1026 organic compounds [53].

General solubility equation as proposed by Yalkowsky and Jain’s [52]:

Modifications in terms of the GSE have been proposed, for instance with the SCRATCH model, which replaces the melting point by molar aqueous activity coefficient, with comparable accuracy (R² = 0.956, RMSE = 0.859 in a dataset of 883 compounds) [54]. Ali and collaborators suggested replacing the melting point descriptor of the GSE with TPSA, aiming to overcome the issues with compounds with high melting points and also to explicitly take into account the effect of polar and polarizable atoms on the aqueous solubility [55].

The argument that real drugs are actually more soluble than drug-like molecules, filtered by Lipinski’s rule of five [56], pointed out the studies in the direction of more complex models. Indeed, nowadays, the quantitative structure-property relationship (QSPR) models correlating the aqueous solubility with various molecular descriptors are often employed. As an example, Chevillard et al. reported the use of a random forest protocol to select the most accurate model among several available, both in commercial or free software packages, for each compound [57]. They report that the multimodel approach can enlarge the applicability domain given that more accurate results for solubility prediction were obtained in comparison to using individual models. This approach agrees with other reports that consensus of local QSAR models can generate predictive workflows, especially for datasets with large structural diversity [58, 59]. It is worth noting that Lipinski himself recently revisited his own rules [60], in vision of new potential classes of drugs, such as natural products, peptide-like, and fragments, which, despite the validated effect, would defy the original Ro5 limits.

3.2 Ionization state and pKa prediction

Early pKa measurement proves beneficial in lipophilicity assessment since logDvalues at any pH can be calculated from the pKa and logPvalues [25, 50]. Although octanol/water logPis similar to most components in the body, not all biological partition processes (i.e., blood-brain barrier and gastrointestinal absorption) can be easily modeled by it [25].

The prediction of ionization state of compounds, which is indicated by the pKa value, is relevant to derive several other physicochemical and ADME properties of drugs, including solubility, lipophilicity, and pharmacokinetic profile. The use of pKa prediction can be placed in two different stages along the DDD, in the beginning with fast models for larger libraries, intending to generate all possible state populations of particular compounds, and/or later on with more refined semiempirical and, computationally expensive, the density functional theory (DFT), in which more accurate ionization states can be accessed. Examples of fast prediction methods for ionization states, which are available as computer programs, are SPARC [61], MoKa [62], and Epik, which use the Hammett and Taft approaches for the pKa prediction [63]. On the other hand, once smaller subsets of molecules are being addressed, the use of semiempirical or density functional theory (DFT) with more computationally expensive models was reported to accurately incorporate the structural features and diversity into the pKa prediction [64, 65].

3.3 Permeability and the use of cellular and noncellular models

As already seen, lipophilicity (logP, logD) is highly involved in membrane permeability. Apart from the already described in vitromethods for logPand logDdetermination, systems for ex vivo/in situbut also in vivoassessment exist as “direct empirical” determination of permeability. When talking about permeability, the difference between passive diffusion and transporter-mediated active transport needs to be considered.

Cell culture methods have been applied to study intestinal absorption for several decades already [66]. Finding the correct model or cell line is crucial to assess the desired parameters such as passive or active transport. In general, it cannot be distinguished between the different transport mechanisms when using cell culture approaches, but several models exist to shift the focus on one of the parameters.

Two main cell lines are in use as models for intestinal absorption: Caco-2 and MDCK cells. Caco-2 cells are derived from a human colorectal carcinoma and possess many of the typical properties of the small intestine, therefore representing a well-established and validated assay system for absorption, permeability, and secretion studies [21, 67]. This assay is mainly used for rank ordering of compounds in terms of oral absorption and permeability in early phases of drug design. Unfortunately, results obtained in different batches and laboratories vary heavily due to several reasons, which make control compound usage necessary and represent a drawback of the technique [26]. Additional disadvantages include long preparation times (about 3 weeks) and no specific permeation mechanism evaluation. Caco-2 assays are usually used as a primary assay followed and complemented by other in vitroand ex vivomethods [68]. Recently, a 3D version of the Caco-2 assay, “Caco-2 3D spheroid permeability assay” was reported, increasing the overall performance and correlation to a human in vivodata [69].

As already stated above, transcellular permeation either occurs passively via diffusion of lipophilic molecules or is driven by membrane transporters. Important transporter includes ATP-dependent efflux transporter such as MRP2, BCRP, and P-gp and the organic solute transporter and the multidrug resistance protein 3 (MRP3) on the luminal and basolateral membranes, respectively [26].

Madin-Darby canine kidney (MDCK) cells are an alternative to Caco-2 cell-based assays and the next most common cell line for passive permeability assessment as well as drug-receptor interaction [70]. MDCK cells also are ideal for transfection and overexpression experiments with human transporters and receptors due to the lack of P-glycoprotein [68, 71]. For instance, the MDCK-MDR1 cell line overexpresses the multidrug resistance protein 1 (MDR1, P-glycoprotein) and can be used in concert with other cell-based assays to specifically address the influence of MDR1 in drug efflux [72].

Immobilized artificial membranes (IAMs) were already used very early on for lipophilicity determination and are gaining interest again in recent years for direct permeability measures [25]. IAMs are also intensively used in the measurement of the volume of distribution to mimic in vivobinding to phospholipids and phospholipid bilayers (membranes). Therefore, IAMs are discussed in more detail in the following section.

The parallel artificial membrane permeability assay (PAMPA) [73] is a cheap and fast in vitroalternative to cellular-based assay systems. A very comprehensive review of recent PAMPA methodologies and applications is available [74]. In principle, PAMPA was developed to overcome cellular-based systems (Caco-2, MDCK, etc.) for passive permeability evaluation, which are error-prone, more difficult, and labor and time intensive and tend to report false negatives. Another advantage of PAMPA over conventional cell-based assays is the ability to selectively measure passive permeability, while in cell-based systems influence of membrane transporters cannot be left out. PAMPA assays can be readily applied in high throughput processes or scales and different variants exist to address ionic and H-bonding with membranes that influence permeability and complement the use of Caco-2 and other cell assays [73]. Bermejo and colleagues also showed a significant correlation between Caco-2, in siturat perfusion, and PAMPA assay data underlining applicability of the method for ADME assessment [75].

Although high-throughput applications of newly developed and standardized techniques allow gathering of an exorbitant amount of data, it is crucial to also (cor)relate the physicochemical and biomimetic properties to structural features of the compound. This will facilitate the development of QSPRs and allows the construction of in silicomodels ultimately guiding the medicinal chemistry efforts [25].

When dealing with oral administration, it is important to note that the drug is not only confronted with the hurdles of solubility and permeability in the absorption process but is also facing metabolizing mechanisms (i.e., enzymes) in the gastrointestinal tract, which are referred to as first-pass metabolism [24, 76]. These include but are not limited to P-glycoproteins, uridine diphosphate glucuronosyltransferase, and mainly cytochrome P450s (CYP450) [24]. This will be discussed more deeply in the Metabolism section.

Permeability has a direct influence on the drug absorption rate and, as discussed, despite the several in vitrocellular models available (e.g., Caco-2, PAMPA, and MDCK), the high costs justify the use of in silicoprediction. Further, QSPR study developed using a large compound dataset of Caco-2 permeability data (1272 compounds) presented good apparent permeability prediction accuracy (R² = 0.81 for the test set) using the polar volume, number hydrogen bond donors, and the surface area as main descriptors [77].

However, we are far from a model that can predict overall permeability and, the current status, rather focuses on individual compartments and tissues, such as the gastrointestinal (GI) tract, skin, buccal membrane, and the blood-brain barrier (BBB). Since the first BBB permeability correlations with logPin 1977 [78], models to predict BBB permeability, particularly logBB(Eq. (2)), have greatly advanced. Current models using an array of machine-learning methods such as multilinear regression, support vector machine (SVM), and artificial neural network (ANN) against a dataset of 320 unique compounds had good predictive power (R² = 0.89) [79]. The work of Shen et al. developed SVM models using 1593 compounds (1283 BBB⁺ and 310 BBB⁻) by using different pattern selection methods and obtained the overall accuracy of 98.2% [80]. Both methods have the limitation of unbalanced datasets (where the number of BBB+ is higher than the BBB− within the training set), which was addressed on the work of Wang et al. by using resampling methods coupled with the machine-learning techniques, to achieve accuracy rates of 0.919 in external test data [81]. Wang and collaborators compiled a dataset of 439 unique molecules, which were employed to generate a diverse set of QSAR models and consensus (R² = 0.504 for external dataset prediction). They also reported the use of transporter profiles as additional biological descriptors to develop hybrid QSAR BBB models, with an improved correlation coefficient R² = 0.526 for external validation [82].

LogBBcan be calculated by the log of the ratio between the concentration within the brain (C_brain) by the bloodstream concentration (C_blood) of the determined chemical entitiy

Finally, beyond the usual ADME parameters of interest in DDD, there are several other unusual ones that also can be predicted; as examples, we here point the permeability of the models for skin permeability, which evolved from simple diffusion models based on molecular weight and n-octanol/water partition coefficient [83, 84], until more sophisticated models, such as (non)-linear QSPR models and even molecular dynamics simulation (as extensively reviewed by [85, 86]).

4.1 Absorption

Oral bioavailability is defined as the amount of drug that reaches the site of action after oral administration and is influenced by factors like drug solubility and dissolution, chemical and enzymatic stability in the gastric and intestinal lumen, interacting luminal contents (food), gastrointestinal transit time, enterocyte permeability, and intestinal and hepatic metabolism [24]. Recently, bioavailability has been also described as the rate and speed of the drug to reach systemic bloodstream, considering the initial formulation as the starting point.

Oral administration includes a pharmaceutical phase—prior to PK and PD phases—that comprises disintegration and dissolution of the dosage form. When using oral dosage forms, the shape and chemical composition (e.g., tablets) play an important role since they contribute to the time needed for disintegration and dissolution.

Following the pharmaceutical phase, absorption is the first step in the pharmacokinetic phase and is defined as the movement of the drug from the site of administration to the bloodstream. The main properties determining the rate of oral absorption for small molecules are permeability and solubility [87].

As such, the rate of dissolution and ionization, which are described by the Noyes-Whitney and Henderson-Hasselbalch equation, respectively, is the key factors in lead optimization for oral administration and is complemented by lipophilicity as an additional factor influencing membrane permeation and solubility of the compound [31].

Dissolution can be expressed by a function of the aqueous solubility of a compound, the surface area of the administered tablet (or the particles in other solid formulation), and a specific dissolution rate constant. Altering any of these parameters directly affects the dissolution profile [26]. While solubility is an endpoint value indicating the amount of a compound that is soluble in a solvent, dissolution describes the kinetic process of a compound being solved in a solvent [88].

On the other hand, ionization reflects if a compound is present in the charged or uncharged state and is at least influenced by two major parameters. The physicochemical property responsible for ionization is the pKa and describes the ionization state of that entity at a given pH. It is also referred to as aqueous ionization constant [30]. Thereby, it is directly influenced by the pH of the environment, the second parameter, which drastically changes on the way through the GI tract, from about pH 1 to 8 in the stomach and ileum, respectively.

The determination of the ionization state of a compound in the gastrointestinal system (stomach, jejunum, ileum, and colon) is crucial for absorption since it not only influences the solubility of a compound but also the lipophilicity and permeability [26, 30, 89]. About 60–70% of all drugs (effective 1999) are ionizable, which underlines the role that ionization plays in ADME assessment [30, 90]. While charged molecules easily dissolve in aqueous systems (GI tract), they do not permeate membranes via passive diffusion and are reliant on active transport. The contrary is true for uncharged molecules, which pass biological membranes passively but show low solubility in aqueous solutions. Mechanisms of drug absorption include passive diffusion, active transport, and receptor-mediated endocytosis, which are influenced by different factors and can themselves influence the bioavailability.

Similar to model and prediction, the absorption of a drug is a complex process, which is influenced not only by the physicochemical properties of drugs themselves but also by the physiological state of the tissue in question. As such, there are a large number of prediction models available, which were generated based on the physicochemical properties involved in the absorption process, such as membrane permeability and drug solubility. These models can help formulation scientists to optimize drugs with poor absorption due to low aqueous solubility.

Initial absorption models can be separated into dispersion and compartmental models [91]. While dispersion models treat the GI as a continuous system, with variable pH and surface area, compartmental models take into account physiological factors such as transporters. The compartmental absorption transit (CAT) was one of the first models to regard distinct physiological properties, such as the minimal absorption in the stomach and colon, while assuming some mathematical simplifications, such as the instant dissolution of the drug and linear kinetics [92]. CAT was further modified as advanced CAT (ACAT), by including nonlinear absorption kinetics and the effects of the first-pass metabolism. ACAT also considers the gastrointestinal tract as nine subsections, each with unique physicochemical properties, such as pH, allowed solubility, particle size, and permeability [93]. Novel developments have included other absorption routes other than the GI, which have been recently included in commercially available software, such as oral absorption for the development of sublingual zolpidem tablets [94]. The absorption constant (K_a, expressed in terms of h⁻¹ min⁻¹), or also called first-order absorption rate constant (to not be confounded with pKa), is employed in most of the aforementioned models and is determined as a result from the changes in mass of absorbable drug over time at the site of administration. K_a can be derived from the decrease in the drug amount of absorbable present at the site of administration over time; however, it is often indirectly determined by the drug amounts measured in the blood and/or urine.

Along physicochemical models, which have a global application, machine-learning techniques were extensively employed to model absorption (as comprehensively reviewed by Kumar et al. [95]) and are inclined to be local models, since they are mostly based on a small, homogeneous dataset that influences their applicability domain.

4.2 Distribution and the role of plasma-binding proteins

After being absorbed and entering the circulatory system, the drug moves reversibly between different compartments within the body, which is described as distribution and influenced by several physicochemical properties of the drug and biological factors of the body. One of the most important properties is lipophilicity, and as such logP/logD, since it reflects the ability of the compound to pass biological membranes to reach other sites, tissues, and organs within the body [25]. Additional factors include phospholipid and (plasma) protein binding, which reduces the free drug concentration within the body, can prevent the migration to the receptor side/side of action, and causes drug-drug interactions [25, 96]. Interestingly, binding to plasma proteins can also prolong the drug action by releasing the drug over a longer period of time. It is also important to note that the influence of lipophilicity on plasma protein binding is hypothesized to be higher for acidic compounds than for bases, meaning that negative charges contribute highly to plasma protein binding and prevent tissue binding, which leads to diminished volumes of distribution (V_d, Eq. (3)). The V_d is the amount of drug that is freely available in the blood, thus not bound to plasma proteins or other components [25, 97, 98]. V_d is an apparent volume that increases proportionally to the extravascular drug binding and not an anatomically defined volume. Consequently, extensive drug binding outside the bloodstream leads to increasing values of volume distribution.

Volume distribution (V_d) is defined by the ratio between the amount of drug in the body (A) and the drug concentration in plasma (C, comprising both free drug and protein-bound drug):

The parameter describing protein binding is the plasma protein affinity constant K_i. Many efforts to determine distribution led to chromatography-based methods, such as (RP-)HPLC to mimic n-octanol/water logPor lipophilicity to measure distribution. In general, chromatographic methods are believed to resemble biological partition processes more than octanol/water partition [25]. In the beginning, stationary phases in (RP-)HPLC were either silica-based or polymer-based but both had difficulties to reproduce logPand logDvalues despite several additives in the mobile phases [99]. The introduction of biomimetic (stationary) phases coated with human serum albumin (HAS), α₁ acid glycoprotein (AGP), or immobilized artificial membranes (IAM) revolutionized the methodology since they allowed a better approximation of the biological system [25, 100].

A method to address plasma protein binding is the use of HSA and other plasma proteins (e.g., α₁ acid glycoprotein) coupled with RP-HPLC [25, 101]. On the other hand, HPLC combined with IAMs is a popularly accepted technique for phospholipid interaction and partition and several IAM columns are commercially available for DDD projects. Both techniques represent good assay systems to model in vivoV_d in high-throughput scale [98]. Problems with HPLC techniques, which are also true for biomimetic phases, include the lack of a gold standard that is needed to calibrate and later standardize results to make a comparison possible [25].

In vitrostandard methods for unbound plasma fraction calculation include equilibrium dialysis and ultrafiltration among several others as the two most commonly used methods and are considered the gold standard for binding assessment [26].

To calculate the V_d “a priori”/nonexperimentally, plasma protein binding, experimental logDand pKa are necessary. Then again, based on the Vd, the half-life (t_1/2) of a compound can be calculated [102]. Apart from protein binding, tissue binding is also involved in the distribution of the compound. Generally, “tissue” here comprises several components of the human body such as lipids, DNA, or RNA and is also referred to as nonspecific binding [26].

In silicomodels to predict the V_d are often based on lipophilicity and solubility descriptors, which correlate with the fractions of the drugs that are either bound to plasma proteins or freely available. The work of del Amo et al. not just accurately predicts V_d and unbound drug fraction but also compares the model’s performance against the commercially available software VolSurf+ with comparable accuracy (R² = 0.70 and 0.71, respectively) [103].

Expanding these studies, the work of Lombardo and Jing generated a set of models to predict the V_d in the steady state (V_ss), using a dataset of 1096 diverse compounds [104]. They compared models generated by linear (PLS) with nonlinear (Random Forest) models, recommending the latter, with 33 descriptors, as the optimal method for V_ss prediction.

The V_d of drugs is greatly influenced by binding to plasma proteins with several machine-learning–based models generated to predict this interaction. Protein-protein interaction (PPI) information derived from molecular docking was employed to derive a PPI-QSAR model for a small dataset of antibiotics (65 unique compounds), which resulted in an accurate model (R² = 0.86 for the test set) [105]. Additionally, global quantitative models using an array of classification and regression models using physicochemical and molecular descriptors derived from a dataset of 794 compounds were shown to correctly classify the binding status of the test set compounds and could be used as a prescreening [106]. Another recent QSAR study using an extensively curated training set of 967 diverse pharmaceuticals aimed to predict plasma protein-bound fractions (fb) using models generated by six machine-learning algorithms with 26 molecular descriptors [107]. This study is particularly interesting where the applicability domain is concerned allowing to differentiate whether the classification derives from (un-)favorable regions.

del Amo et al. recently reported one of the first QSPR models to predict intravitreal volume of distribution and clearance of small molecules [108]; the model relies on the LogDand hydrogen bond capacity to understand phenomena such as intraocular pressure and guide drug discovery. Complementarily, the prediction of the drug passage through the blood-ocular barrier was described to be an important factor to evaluate volume distribution in this organ [109].

Recently, as a novel approach bridging the animal experiments with human results, it was shown that in PXB mice, a chimeric mice linage with a humanized liver, plasma concentration-time profiles could be used to infer human’s compound half-life [110].

Volume of distribution is also closely related to half-life and clearance parameters. As the V_d is a relative measurement of the free concentration of drug in the blood, this same amount could be excreted by kidneys in the glomerular filtration (clearance). Consecutively, the rate of clearance (discussed below in Excretion section) directly influences the amount of available drug. Naturally, the concentration of free drug that can bind its molecular target is related to the therapeutic dosage and the half-life of the administered drug (as seen in Eq. (4)).

Half-life definition. Half-life is calculated by a ratio between the Napierian logarithm multiplied by the volume of distribution (V_d) and renal clearance (CL):

4.3 Metabolism

Drug metabolism normally involves enzymatic modification or degradation of the compound to facilitate excretion via one of the major clearance organs: liver, kidney, spleen, or bile. While phase I enzymatic reactions include modifications such as oxidation, hydrolysis, and reduction to either introduce a functional group to the molecule or make it accessible, phase II reactions are conjugation mechanisms (e.g., methylation, acetylation, glutathione conjugation, amino acid conjugation, and others) that result in polar products that can be actively effluxed [26]. Thus, isozymes of the CYP450 family and efflux transporters such as P-glycoprotein and members of the multidrug resistance transporter MRP family are highly involved in the metabolism of drugs as well as drug-drug interactions, which are a major attrition cause. For instance, CYP3A4, CYP2C9, and CYP2D6 together catalyze the hepatic metabolism of about 50% of drugs, which underlines the importance of the superfamily. Interestingly, when CYP3A4 is expressed, usually P-glycoprotein is as well [8, 10, 14, 24, 111]. An approximation for metabolic behavior analysis is the use of either liver microsomes or S9 fractions although also recombinantly expressed proteins are partially in use [24, 26].

When available, the 3D structure of those proteins could be employed in molecular docking and molecular dynamics simulations aiming to predict the binding affinity of drugs or drug candidates aiming the estimation of a PK profile [112]. The metabolism prediction combines mathematical models to predict whether the target compound could be a substrate of a specific enzyme in combination with metabolism site predictions. Usually, those initial predictions are followed by molecular docking simulations and quantum mechanics simulations due to the dependency of electronics structure from both substrate and enzyme in catalyzed reaction [113, 114].

Nowadays, several attempts have been made to develop in silicomodels for predicting drug metabolism, specifically site-of-metabolism (SOM), and quite often are also converted into online server prediction tools for general use, for instance, the FAst MEtabolizer (FAME) model, which was generated from a diverse chemical datasets of more than 20,000 molecules and their respective experimentally determined metabolism sites. FAME prediction rates were comparable to other metabolism site predictors focused on specific enzyme families, despite using only seven chemical descriptors [115]. Similarly, SMARTCyp server, which initially relied on the 2D structure of the molecule, without considering electronic properties or generating 3D structures, to predict CYP2D6 [116], was later expanded for other CYP isoforms. A more refined version was later updated to include the atomic solvent accessible surface area, which is independent of 3D coordinates, slightingly improving the overall prediction accuracy for different CYP isoforms [117]. The newest SMARTCyp version (3.0) uses the activation energies calculated by the density functional theory (DFT), meaning the difference between the energy of the transition state and the reactant complex, to predict SOMs of molecular fragments of the query in an unsupervised fashion. SMARTCyp 3.0 also calculates the similarity between the query molecule and the model fragment [118].

IDSite approach aims to overcome the ligand-based bias of SOM prediction by using it as a part of a large framework, more precisely by combining it with molecular docking, where an atom can be considered a significant SOM by a P450 enzyme when accessible to the reactive heme iron center, and/or quantum calculations, where the candidate atom must have some degree of reactivity in the absence of the enzyme [119]. Similarly, the work of Kingsley et al. combined different approaches into a framework to predict CYP2C9 substrates. They validated the predictions from SMARTCyp in an ensemble docking, followed by a QSAR model to account for influences of both the inherent reactivity of each atom and the physical structure of the CYP2C9 binding site [120]. This combined approach resulted in 88% of true SOMs accurately predicted among the top ranked sites.

4.4 Excretion

Excretion is guided by one of the major clearance organs, and the assessment of clearance behavior sometimes involves isolated organs or tissues [24]. Humans rely on the kidney clearance as a major route for xenobiotic excretion, despite other available routes such as feces, bile, sweat, and breath. The excretion pathways directly impact the concentration of available drugs and are often measured in terms of half-life and the initial administered dose.

The renal clearance of a drug is another important parameter, which is usually employed to predict drug excretion. Experimentally, clearance is defined by the drug concentration drug along a defined time of renal excretion by a linear equation (Eq. (5)).

Equation for renal clearance. mis the substance’s mass generation rate, Kis the clearance and Cis the concentration at the time, and Vis the volume where the drug is distributed, or for systemic approaches the total body water.

Gombar et al. developed SVM- and MLR-based QSAR models to predict both systemic clearance and apparent volume of distribution from intravenous data [121] using as input structural fingerprints and electro-topological states (so-called E-states), respectively. The model performed with high accuracy, despite the highly diverse initial dataset employed for its generation, which points the importance of those models in early steps of the drug-discovery pipeline.

Also, the work of Kusama et al. established a chemoinformatic-based classification model to predict the major clearance pathways of 141 approved drugs based on four physicochemical parameters: charge, molecular weight, lipophilicity, and protein unbound fraction in plasma, resulting in a final model with an accuracy of 88% [122]. This model approach was further refined by using support vector machine and increasing the number of relevant descriptors [123]. In order to better model the biotransformation processes, often the major triggers of excretion, the work of Berellini et al. used ELASTICO (Enhanced Leave Analog-Structural, Therapeutic, Ionization Class Out) to provide an appropriate sampling during the validation process. Their partial least-square models resulted in a highly accurate model derived from 754 compounds [124].

On another topic, ABCB1, also known as P-glycoprotein (P-gp or MDR1), is a membrane protein member of the ATP-binding cassette (ABC) transporters superfamily. Together with the hERG channel and CYP3A4, P-gp is one of the most widely studied antitarget, where its inhibition could bring consequences for several processes, such as the absorption, distribution, and excretion of drugs. Classical studies used chemometric methods to describe bioavailability in terms of P-gp and CYP enzyme activities, generating QSAR models based on 805 unique drug molecules with high accuracy (R² = 0.80 for the test set) [125]. Alternatively, an approach to predict P-glycoprotein inhibition using molecular interaction fields, derived from a literature collection of more than 1200 structures, generated a pharmacophore model for competitive P-gp inhibition [126].

The most recent reported studies involving prediction of drug clearance, both from human and rat hepatic in vitrosystems, were based on microsomes, with a recent emphasis on the use of hepatocytes. Wood et al. discuss the inherent limitation of using human hepatocyte predictions, due to underprediction when compared to in vivoclearance data, and the comments on the potential causes for those divergences [127].

As the pinnacle of ADME in silicoapproaches, the holistic physiologically based pharmacokinetic (PBPK) modeling was initially conceptualized by Teorell [128], aiming to enable the prediction of drugs’ pharmacokinetic behavior in humans using preclinical data. Recent PBPK models benefit from the large amount of available ADME data not only to aid the drug discovery process and dose regiment selection but also to guide the risk assessment for regulatory reviews [129]. PBPK models are compartmentalized representations of the different organs, and each compartment can be described by a specific tissue volume and blood flow rate, which communicates with the blood (venous and arterial). Each organ/tissue has a unique volume, permeability, and eliminating anatomical constants and terms, which are determined independently from the studied drug, while other physiological drug-specific parameters are later incorporated, such as affinity toward plasma proteins, tissue-to-plasma distribution rate, and even on target activity (K_m, V_max, or binding kinetics).

One of PBPK models’ important features is the perspective for the mechanistic and prospective prediction of a drug’s pharmacokinetic profiles. The use of drug-dependent parameters includes, but is not limited to, physicochemical properties, solubility and permeability values, and also the role of individual enzymes and transporters in the metabolism. Those parameters can be determined in vitroor calculated from the compound structure with other in silicoapproaches, which allows the early use of PBPK in the DDD (the bottom-upapproach). Concurrently, it is also noteworthy that the model construction and parameter fine-tuning are a source of knowledge for the hit development, where the predictions from the ongoing model can help to understand the model’s accuracy itself along the way (called as middle-outapproach) and then prospectively be applied to simulate unstudied scenarios. Currently, there are several free-to-use and commercially available PBPK and ADME prediction options (Table 2), which are also extensively reviewed and discussed by the works of Madden et al. [130].

Early PBPK models, such as the work of Varma et al., described another layer of complexity by including drug-drug interactions (DDI). The dosing time-dependent model considering the interaction between repaglinide with rifampicin was able to predict repaglinide plasma concentrations along a day. The model also predicted the drug interaction with other CYP3A4 and OATP1B1 inhibitors, which could result in further DDIs. Reports of DDI leading to complications in patients with particular genotype stimulated studies such as the one performed by Fermier et al. [131], where the effects of polymorphic cytochromes provided the basis for a more accurate DDI prediction.