Leveraging Chemical Background Knowledge for the Prediction of Growth Inhibition

We show how chemical background knowledge can he used to improve the prediction performance in structure-activitity relationships (SARs) for non-congeneric compounds. The goal of the study is to build a model of growth-inhibition for the NCI DTP human tumor cell line screening data. The SAR model is based on frequent molecular fragments generated from the structure data. Background knowledge in the form of standard anti-cancer agents (ACAs) grouped by known mechanisms of action is available and used twice: First, the standard agents are treated separately in the fragment generation process. Second, we represent each molecule in terms of the similarities with structures known to be associated with certain mechanisms of action. In experiments, we show that using chemical background knowledge in this way reduces the mean absolute error (MAE) by about 5% compared to initial experiments, and by 9% compared to a previous publication. We conjecture that specific instances and groups of instances are a commonly occurring type of background knowledge that is particularly easy to use and effective in practice