Application of machine learning to predicting cancer drug response

Cilj rada je kreirati alat koji će potencijalno pomoći liječniku odabir najbolje terapije, te uštedjeti velike svote novaca svima uključenima u proces izrade lijeka. Kao polazne točke u kreiranju ovoga rada koristile su se dvije baze podataka. Spojili smo genetske karakteristike staničnih linija raka i kemijske strukture lijekova. Nad tim podacima kreirali smo model algoritma slučajne šume, te algoritme hrbatne regresije i regresije najmanjeg kuta. Pokazalo se da se algoritam slučajne šume ponaša najbolje, i naš model ima bolje rezultate ($ R^2 $ metriku) od konkurentnog rada. Pokazalo se, također, da se model ne ponaša dovoljno dobro u slučaju predviđanja rezultata za novonastali, dosad neviđeni, lijek. Zbog toga, daljnji fokus istraživanja će biti na kreiranju modela koji će se moći snalaziti u danim okolnostima.The main goal was to create tool that would, potentially, be able to help doctors in therapy decision making, as well as saving a lot of money during the development of treatment methods. As starting points in this paper two databases were used. We merged genetic characteristics of cell lines and chemical structures of medicine. We applied random forest algorithm on these data. We also trained the ridge regression and the least angle regression. It was shown that the random forest algorithm showed best results, and that my results were better than the previous work. I showed that, also, my model does not manage to work well on predicting results on previously unknown medication. The future research will be focused on creating a model that will work well in such use case