Active Learning for drug discovery

Student thesis: Doctoral ThesisDoctor of Philosophy


This thesis describes work conducted to enable Robot Scientist Eve to autonomously evaluate drug-like chemicals during high throughput experiments. Eve tests libraries of chemical compounds against yeast-based targets expressing parasite and host (human) proteins (i.e. DHFR, NMT & PGK); the parasites included in this study are responsible for an array of neglected tropical diseases. The raw data for yeast growth curves from an initial screen were evaluated, and decision tree rules were constructed to describe the relative activity and toxicity of compounds. These rules were verified, and versions were subsequently developed for application to routine mass and confirmation screens. Consequently, many potential lead drug-like candidates have been identified in the Maybridge Hitfinder library; several compounds from an approved drug library (the Johns Hopkins Clinical Compound Library) have also been confirmed as exhibiting activity against these yeast-based targets. Further in vivo study of some JHCCL compounds is in progress using extracted parasite proteins; preliminary results indicate the potential for repositioning Triclosan and Tnp-470 as having anti-malarial behaviour based on their interaction with Plasmodium sp. DHFR proteins. In the second phase of the programme, a prototype Active Learning strategy was applied (active k-optimisation) to partial mass screen data as a seed; this allowed Eve to select compounds by assessing and predicting quantitative structure activity relationships (QSAR) between seed and unknown compounds. Simulations of learning and testing QSAR cycles showed that Eve would be able to select active compounds more efficiently under such a regime. Other strategies have been developed that further improve selection efficiency for active compounds, and also promote the ability to find rare category compounds. An econometric model has been developed to demonstrate the potential beneficial impact of Active Learning strategies on the execution costs for such screens
Date of Award20 May 2014
Original languageEnglish
Awarding Institution
  • Aberystwyth University
SupervisorChuan Lu (Supervisor) & Nigel Hardy (Supervisor)

Cite this