Understanding the primary and secondary pharmacology of drugs is imperative for delivering a drug molecule which has the necessary efficacy and safety profile in humans. The application of machine learning and data mining methods to drug discovery has the promise to accelerate the drug discovery process by learning from the large amounts of existing data available. This thesis focusses on in silico approaches to address challenges related to selectivity and toxicity in drug discovery. The first section of this thesis is concerned with the prediction of ligand selectivity profiles using proteochemometric (PCM) modelling, a technique which uses both compound similarity and protein target similarity as input into machine learning models for the prediction of ligand-target interactions. We showed that employing a multi-target PCM model outperformed other methods, including QSAR models, on the same bioactivity dataset for the bromodomain-containing proteins. Furthermore, we established the applicability domain of the model by employing conformal prediction, which was further used to aid the selection of compounds for prospective experimental testing in bromodomain assays. We achieved a 31 % hit rate for actives from our experimental selections, including the discovery of diverse novel hits. The PCM models were interpreted to reveal residues in the protein active site important for the classification of activity for each bromodomain, which were further validated by the generation of co-crystal structures for new ligands bound to bromodomains, as well as literature evidence. The PCM technique can be used in virtual screening, in silico off-target panel screening of compounds and to aid future structure-based design. The second section of this thesis is concerned with the translation of toxic effects between preclinical and clinical studies. Animal models of toxicity are used in the drug discovery process to assess the risk of toxicity in humans. However, these effects do not always translate into human studies as demonstrated by previous retrospective concordance analyses. Here, we asked “what more can animal models tell us about toxicity in humans?” by extending the previous concordance approaches to find associations between toxicities in animals and humans which were not described by the same adverse event (AE) terms. Using data from PharmaPendium, we derived 2,050 statistically significant associations using the mutual information and subsequently located preclinical AEs which, when observed for a drug, were indicative of a high risk of experiencing a clinical AE, as measured by a positive likelihood ratio of greater than 10. To find mechanistic links for associations with the highest mutual information values, we conducted an analysis to find intersecting genes between preclinical AE, clinical AE and the drugs which were reported to cause both AEs, finding genetic evidence for 25 % of associations that were analysed. We suggested that the protein targets identified by this method can be considered for inclusion in in vitro toxicity panel screening to enable the earlier prediction of toxicity. In summary, we quantified from existing data where animal models were useful in clinical toxicity prediction and where they were not, as well as generated mechanistic hypotheses for the connection between in vivo and clinical toxicities. This study can provide evidence to support the case for in vivo animal model usage for the prediction of certain clinical toxicities, as well as provide suggestions for targets to incorporate into in vitro screening panels. Both outcomes contribute towards the aim to replace, reduce and refine animal usage in drug discovery. Overall, this work contributes to the prediction and understanding of the primary and secondary pharmacology of drugs and how this relates to the concepts of selectivity and safety. The four main applications of this work included; 1. a machine learning tool for the prediction of selective bromodomain inhibitor chemotypes, 2. a guide for determining which binding site residues to interact with to optimise binding to bromodomains, derived from machine learning models, 3. a quantitative assessment of the drug-induced toxicities in animals which indicate a high risk of drug-induced toxicities in humans to guide preclinical safety decisions, 4. proposed targets for further investigation for inclusion in in vitro secondary pharmacology screening panels for early detection of potential toxic liabilities. These techniques can be applied to other target classes or toxicity datasets in the future.