Predicting Drugs that Inhibit Growth in Cancerous Immortal Cell Lines

Abstract

In this work, we explore the possibilities of predicting cancers using only information from compounds, and their target growth inhibition value for each cell line. We focused on representing the compounds in different ways such as with fingerprints and skeletal formulas. These representations were then fed into various neural nets and tree-based models to try to predict the growth inhibition values. We observed that tree-based methods on fingerprints performed the best over the various combinations of models and representations, and this performance rivals that of other previous works that combine a multitude of information about cell state and expression with compound data. These results indicate that compound information alone can be used to predict drugs that inhibit the growth of cancer, although more specialized models can push the predictive power even higher. With such specialized models, this form of prediction could be used even more to guide rational drug discovery in the fight against cancer.