Advances in Predictions of Oral Bioavailability of Candidate Drugs in Man with New Machine Learning Methodology.

Oral bioavailability (F) is an essential determinant for the systemic exposure and dosing regimens of drug candidates. F is determined by numerous processes, and computational predictions of human estimates have so far shown limited results. We describe a new methodology where F in humans is predicted directly from chemical structure using an integrated strategy combining 9 machine learning models, 3 sets of structural alerts, and 2 physiologically-based pharmacokinetic models. We evaluate the model on a benchmark dataset consisting of 184 compounds, obtaining a predictive accuracy (Q2) of 0.50, which is successful according to a pharmaceutical industry proposal. Twenty-seven compounds were found (beforehand) to be outside the main applicability domain for the model. We compare our results with interspecies correlations (rat, mouse and dog vs. human) using the same dataset, where animal vs. human-correlations (R2) were found to be 0.21 to 0.40 and maximum prediction errors were smaller than maximum interspecies differences. We conclude that our method has sufficient predictive accuracy to be practically useful with applications in human exposure and dose predictions, compound optimization and decision making, with potential to rationalize drug discovery and development and decrease failures and overexposures in early clinical trials with candidate drugs.

Prediction of human pharmacokinetics - renal metabolic and excretion clearance.

The kidneys have the capability to both excrete and metabolise drugs. An understanding of mechanisms that determine these processes is required for the prediction of pharmacokinetics, exposures, doses and interactions of candidate drugs. This is particularly important for compounds predicted to have low or negligible non-renal clearance (CL). Clinically significant interactions in drug transport occur mostly in the kidneys. The main objective was to evaluate methods for prediction of excretion and metabolic renal CL (CL(R)) in humans. CL(R) is difficult to predict because of the involvement of bi-directional passive and active tubular transport, differences in uptake capacity, pH and residence time on luminal and blood sides of tubular cells, and limited knowledge about regional tubular residence time, permeability (P(e)) and metabolic capacity. Allometry provides poor predictions of excretion CL(R) because of species differences in unbound fraction, urine pH and active transport. The correlation between fraction excreted unchanged in urine (f(e)) in humans and animals is also poor, except for compounds with high passive P(e) (extensive/complete tubular reabsorption; zero/negligible f(e)) and/or high non-renal CL. Physiologically based in-vitro/in-vivo methods could potentially be useful for predicting CL(R). Filtration could easily be predicted. Prediction of tubular secretion CL requires an in-vitro transport model and establishment of an in-vitro/in-vivo relationship, and does not appear to have been attempted. The relationship between passive P(e) and tubular fraction reabsorbed (f(reabs)) for compounds with and without apparent secretion has recently been established and useful equations and limits for prediction were developed. The suggestion that reabsorption has a lipophilicity cut-off does not seem to hold. Instead, compounds with passive P(e) that is less than or equal to that of atenolol are expected to have negligible passive f(reabs). Compounds with passive P(e) that is equal to or higher than that of carbamazepine are expected to have complete f(reabs). For compounds with intermediate P(e) the relationship is irregular and f(reabs) is difficult to predict. Tubular cells are comparably impermeable (for passive diffusion), and show regional differences in enzymatic and transporter activities. This limits the usefulness of microsome data and makes microsome-based predictions of metabolic CL(R) questionable. Renal concentrations and activities of CYP450s are comparably low, suggesting that CYP450 substrates have negligible metabolic CL(R). The metabolic CL(R) of high-P(e) UDP-glucuronyltransferase substrates could contribute to the total CL.

Prediction of human pharmacokinetics--improving microsome-based predictions of hepatic metabolic clearance.

Physiologically based methods generally perform poorly in predicting in-vivo hepatic CL (CL(H)) from intrinsic clearance (CL(int)) in microsomes in-vitro and unbound fraction in blood (f(u,bl)). Various strategies to improve the predictability have been developed, and inclusion of an empirical scaling factor (SF) seems to give the best results. This investigation was undertaken to evaluate this methodology and to find ways to improve it further. The work was based on a diverse data set taken from Ito and Houston (2005). Another objective was to evaluate whether rationalization of CL(H) predictions can be made by replacing blood/plasma-concentration ratio (C(bl)/C(pl)) measurements with SFs. There were apparently no or weak correlations between prediction errors and lipophilicity, permeability (compounds with low permeability missing in the data set) and main metabolizing CYP450s. The use of CL(int) class (high/low) and drug class (acid/base/neutral) SFs (the CD-SF method) gives improved and reasonable predictions: 1.3-fold median error (an accurate prediction has a 1-fold error), 76% within 2-fold-error, and a median absolute rank ordering error of 2 for CL(H) (n = 29). This approach is better than the method with a single SF. Mean (P < 0.05) and median errors, fraction within certain error ranges, higher percentage with most accurate predictions, and ranking were all better, and 76% of predictions were more accurate with this new method. Results are particularly good for bases, which generally have higher CL(H) and the potential to be incorrectly selected/rejected as candidate drugs. Reasonable predictions of f(u,bl) can be made from plasma f(u) (f(u,pl)) and empirical blood cell binding SFs (B-SFs; 1 for low f(u,pl) acids; 0.62 for other substances). Mean and median f(u,bl) prediction errors are negligible. The use of the CD-SF method with predicted f(u,bl) (the BCD-SF method) also gives improved and reasonable results (1.4-fold median error; 66% within 2-fold-error; median absolute rank ordering error = 1). This new empirical approach seems sufficiently good for use during the early screening; it gives reasonable estimates of CL(H) and good ranking, which allows replacement of C(bl)/C(pl) measurements by a simple equation.

Prediction of human pharmacokinetics--evaluation of methods for prediction of hepatic metabolic clearance.

Methods for prediction of hepatic clearance (CL(H)) in man have been evaluated. A physiologically-based in-vitro to in-vivo (PB-IVIV) method with human unbound fraction in blood (f(u, bl)) and hepatocyte intrinsic clearance (CL(int))-data has a good rationale and appears to give the best predictions (maximum approximately 2-fold errors; < 25% errors for half of CL-predictions; appropriate ranking). Inclusion of an empirical scaling factor is, however, needed, and reasons include the use of cryo-preserved hepatocytes with low activity, and inappropriate CL(int)- and f(u, bl)-estimation methods. Thus, an improvement of this methodology is possible and required. Neglect of f(u, bl) or incorporation of incubation binding does not seem appropriate. When microsome CL(int)-data are used with this approach, the CL(H) is underpredicted by 5- to 9-fold on average, and a 106-fold underprediction (attrition potential) has been observed. The poor performance could probably be related to permeation, binding and low metabolic activity. Inclusion of scaling factors and neglect of f(u, bl) for basic and neutral compounds improve microsome predictions. The performance is, however, still not satisfactory. Allometry incorrectly assumes that the determinants for CL(H) relate to body weight and overpredicts human liver blood flow rate. Consequently, allometric methods have poor predictability. Simple allometry has an average overprediction potential, > 2-fold errors for approximately 1/3 of predictions, and 140-fold underprediction to 5800-fold overprediction (potential safety risk) range. In-silico methodologies are available, but these need further development. Acceptable prediction errors for compounds with low and high CL(H) should be approximately 50 and approximately 10%, respectively. In conclusion, it is recommended that PB-IVIV with human hepatocyte CL(int) and f(u, bl) is applied and improved, limits for acceptable errors are decreased, and that animal CL(H)-studies and allometry are avoided.

Prediction of human pharmacokinetics--gastrointestinal absorption.

Permeability (P(e)) and solubility/dissolution are two major determinants of gastrointestinal (GI) drug absorption. Good prediction of these is crucial for predicting doses, exposures and potential interactions, and for selecting appropriate candidate drugs. The main objective was to evaluate screening methods for prediction of GI P(e), solubility/dissolution and fraction absorbed (f(a)) in humans. The most accurate P(e) models for prediction of f(a) of passively transported and highly soluble compounds appear to be the 2/4/A1 rat small intestinal cell model (in-vitro and in-silico), a newly developed artificial-membrane method, and a semi-empirical approach based on in-vitro membrane affinity to immobilized lipid bilayers, effective molecular weight and physiological GI variables. The predictability of in-vitro Caco-2, in-situ perfusion and other artificial membrane methods seems comparably low. The P(e) and f(a) in humans for compounds that undergo mainly active transport were predicted poorly by all models investigated. However, the rat in-situ perfusion model appears useful for prediction of active uptake potential (complete active uptake is generally well predicted), and Caco-2 cells are useful for studying bidirectional active transport, respectively. Human intestinal in-vitro P(e), which correlates well with f(a) for passively transported compounds, could possibly also have potential to improve/enable predictions of f(a) for actively transported substances. Molecular descriptor data could give an indication of the passive absorption potential. The 'maximum absorbable dose' and 'dose number' approaches, and solubility/dissolution data obtained in aqueous media, appear to underestimate in-vivo dissolution to a considerable extent. Predictions of in-vivo dissolution should preferably be done from in-vitro dissolution data obtained using either real or validated simulated GI fluids.