Feature Discovery in Small-Sized Experiments in Early Drug Development

Abstract

The abstract consists of the thesis' description from the book of abstracts because it does such an excellent job of summarising the research done in this thesis.

"Nowadays we hear a lot about artificial intelligence (AI) and machine learning (ML) as methods for predicting outcomes. In medical and pharmaceutical sectors, these methods have also become very popular. We will focus on prediction problems in preclinical discovery research in pharmaceutical companies. In this stage of the drug development, no large data sets are available. That is: only a small number of observations on a very large number of features (e.g. data on thousands of gene expression, metabolomics, .... on only 10 to 50 subjects). Scientists use such datasets aiming at finding predictors (e.g. genes) to predict disease status (diagnostics) or to identify responders to a treatment for a disease (personalised medicine). Many of these scientists believe in the power of AI and ML, whereas, however, the success stories of AI and ML come from big data applications (i.e. data from thousands of subjects in the training data).

The goal of this thesis is to study the behavior of AI and ML methods on small sample-sized studies in early drug development. It involves the use of many ML prediction methods in small sample-sized simulation studies. The ultimate goal is to formulate guidelines."