"In silico" prediction of drug transport across physiological barriers

Physiological barriers maintain and safeguard homeostasis of certain body compartments by an increased resistance against free diffusion. Distribution and pharmacokinetics of drugs can be altered as well, if they have to cross these barriers in order to reach their target. Knowledge of the physicochemical and structural requirements for drug permeation is a key topic in drug design, development, and clinical application.
To assess processes on cellular barriers, in vitro methods are usually applied to elucidate single transport mechanisms or to study isolated transport. As the pharmacokinetics of a living system are often more complex and composed by a concatenation of several barriers, in vivo methods are required. However, this time consuming and expensive testing is not suited to answer the need for high-throughput screening of thousands of compounds in chemical databases. For these purpose in silico methods are ideally suited, which produce computational models to predict pharmacokinetics, drug distribution, or transport across single barriers.
The first project of this thesis concerned the modeling of human intestinal absorption. After oral administration and intestinal dissolution, a drug has to cross the gut wall in order to become available for the body. The process is mostly determined by passive diffusion and active transport. Active export and import of molecules on the enterocyte is regulated by a multitude of transport proteins and metabolic enzymes. A dataset of small drug-like compounds, on which information on their human intestinal absorption was available, was collected. Models trained on these data predicted human intestinal absorption with high accuracy. Several machine learning methods were compared as well as different feature sets. The features used to predict intestinal absorption resembled those known from modeling passive diffusion, which are measures of charge and lipophilicity. The models revealed also less commonly used descriptors to model human intestinal absorption, such as gravitational indices and moments of inertia.
The aim of the second project was to develop computational models to predict blood brain barrier (BBB) permeation. Development of new central nervous system (CNS) active drugs is hampered by limited brain permeation. As invasive methods have proven themselves to be ineffective and risky for patients, systemic application is the preferred route for drug administration into the brain. Hence, BBB permeability is a feature absolutely mandatory for any drug, which targets the CNS. Limited passive diffusion and active efflux and influx systems account for the complexity of this highly regulated barrier. To establish our models, a database of 163 compounds with information on the in vivo surface permeability product (LogPS) in rats was collected. Decision trees performed with high accuracy (CCR of 90.9 - 93.9%.) and revealed descriptors of lipophilicity and charge, which were yet described in models of passive BBB permeation. However, other descriptors as measures for molecular geometry and connectivity could be related to an active drug transport component. Moreover, a fragment-based approach indicated the involvement of stereochemistry to predict LogPS values.
The third project explores the physicochemical and structural requirements for drugs to pass from maternal blood into breast milk. While experimental assessment in humans is limited, computational methods are appropriate to model drug permeation into breast milk. Data preparation for these models was a challenging endeavor. Endpoints were reported in imprecise ways, which asked for a careful selection and binning of the instances. Despite these facts, the 10-fold cross-validated decision trees predicted the endpoint with high accuracy (CCR: 85.3 - 95.3%). Prominent descriptors were measures of molecular size, branching, charge and geometry. Importance of polar fragments was revealed by a fragment-based analysis.
The efflux transporter MRP2, a member of the ABC transporter family, was subject of the fourth study. Efflux transporters contribute substantially to barrier function by extruding potentially toxic substances. Three datasets were assembled from literature for MRP2 substrates, inducers, and inhibitors. For inducers and inhibitors, decision trees with high accuracy were grown. However, the substrate dataset did not qualify for decision tree induction, due to an underrepresentation of negative instances.
The fifth project deals with an ant colony optimization (ACO) algorithm, which was adapted for fragment based feature selection. The paradigm was tested to predict antimalarial activity of molecules. ACO was able to reveal chemical substructures characterizing antimalarial drug activity, which comprised passive diffusion through the erythrocyte membrane and parasite toxicity. The paradigm outperformed other algorithms such as decision trees or artificial neural networks on the same dataset.