Exploring Predictive Approaches in Computational Drug Repositioning: A Pathway to Safer and Faster Drug Development

Abstract

The development and approval process for new drugs is time-intensive, costly and carries only a 10% chance of success. Drug repositioning - the strategy of repurposing drugs that have previously completed some or all clinical trial stages for new indications – can save substantial time and resources, while also increasing the success rate to around 25%. Historically, repositioning successes, such as the sildenafil’s approval for erectile dysfunction under the brand name Viagra, have mostly been accidental or identified retrospectively. However, the growing availability of vast biomedical datasets and advancements in computational tools have spurred in more systematic approaches to drug repositioning. In parallel, alternative compound sources that may have a higher likelihood of passing clinical trials have gained attention. Natural products, which have demonstrated safety and efficacy through centuries of medical use, are particularly popular. Therefore, this thesis investigates various computational strategies for drug repositioning, with a focus on predicting human protein targets. Uncovering and understanding drug interactions with off-targets, in addition to their primary pharmacological targets, is crucial in repositioning efforts, as these interactions can sometimes cause harmful side effects or present opportunities to affect targets involved in other diseases. To support these efforts, data on repositioning candidates from sources such as approved or withdrawn drugs and natural products were carefully curated, emphasizing data reliability and integrity to enhance prediction accuracy. This data enabled the exploration of repositioning methods ranging from similarity-based screening and human disease pathway analysis to machine learning models for predicting targets, indications or therapeutic areas. Using the developed models, well-known repositioning candidates such as thalidomide could be validated, confirming their potential in new therapeutic contexts. The resulting datasets and prediction tools were made publicly accessible with user-friendly interfaces, providing the opportunity for other researchers to benefit from and build upon this work.