Evaluating the Effectiveness of Active Learning Methods in Drug Repurposing

Abstract: Creating new drugs is an expensive process both in time and resources. Drug repurposing aims to find other uses for already approved medicines or clinical candidates to speed up the development of treatments; it can be done experimentally or computationally, with in silico approaches demonstrating many advantages. One of the most popular trends in computationally drug repurposing is high throughput screen in combination with target-based drug repurposing, which unlocks the possibility of using the results of the previously mentioned procedures in machine learning tasks. Machine learning techniques can be used to automate the prediction process for new compounds so that only the most valuable ones are tested in the next round of laboratory experiments. Further, active machine learning methods assist the selection process by learning from compounds that are richer in information, making the prediction process even faster. This thesis evaluates the application of active learning techniques in classification and regression tasks. This technique was assessed in two data sets: a highly unbalanced set consisting of fingerprints and labels for the classification task and the output from a high throughput screen for the regression part. The results show that active learning can solve the regression task but fails to solve the classification one, given its inability to handle the imbalance in the distribution of the classes.