Machine Learning for Modelling Tissue Distribution of Drugs and the Impact of Transporters

The ability to predict human pharmacokinetics in early stages of drug development is of paramount importance to prevent late stage attrition as well as in managing toxicity. This thesis explores the machine learning modelling of one of the main pharmacokinetics parameters that determines the therapeutic success of a drug - volume of distribution. In order to do so, a variety of physiological phenomena with known mechanisms of impact on drug distribution were considered as input features during the modelling of volume of distribution namely, Solute Carriers-mediated uptake and ATP-binding Cassette-mediated efflux, drug-induced phospholipidosis and plasma protein binding. These were paired with molecular descriptors to provide both chemical and biological information to the building of the predictive models.

Since biological data used as input is limited, prior to modelling volume of distribution, the various types of physiological descriptors were also modelled. Here, a focus was placed on harnessing the information contained in correlations within the two transporter families, which was done by using multi-label classification. The application of such approach to transporter data is very recent and its use to model Solute Carriers data, for example, is reported here for the first time. On both transporter families, there was evidence that accounting for correlations between transporters offers useful information that is not portrayed by molecular descriptors. This effort also allowed uncovering new potential links between members of the Solute Carriers family, which are not obvious from a purely physiological standpoint.

The models created for the different physiological parameters were then used to predict these parameters and fill in the gaps in the available experimental data, and the resulting merging of experimental and predicted data was used to model volume of distribution. This exercise improved the accuracy of volume of distribution models, and the generated models incorporated a wide variety of the different physiological descriptors supplied along with molecular features. The use of most of these physiological descriptors in the modelling of distribution is unprecedented, which is one of the main novelty points of this thesis.

Additionally, as a parallel complementary work, a new method to characterize the predictive reliability of machine learning classification model was proposed, and an in depth analysis of mispredictions, their trends and causes was carried out, using one of the transporter models as example. This is an important complement to the main body of work in this thesis, as predictive performance is necessarily tied to prediction reliability.