Multi-Omics Visible Drug Activity Prediction with a Biologically Informed Neural Network Model

Drug discovery is a challenging task, it takes several years for a drug to be introduced on the market, with most of<br> the studied drugs not even passing the first phase. The understanding of the mechanisms influencing response to drugs<br> can reduce failures and accelerate drug development. Virtual drug screening, based on Machine Learning models, is a<br> promising field for the prediction of the outcome of a treatment. However, the complex relationships between the features<br> learned by these models are still poorly understood and not easy to interpret.<br> We have designed a Neural Network model for drug sensitivity prediction that leverages a Visible Neural Network, an<br> easily interpretable model, due to its biologically informed nature. The trained model can be inspected to study which<br> biological processes were fundamental for the prediction and to identify the drug properties that affect sensitivity. It<br> combines multi-omics data from various types of tumor tissues and drug representations based on molecular descriptors.<br> The mechanisms learned from the network can also be exploited to find candidate drugs for synergy to predict the effect<br> of combined therapies. We consider the unbalanced nature of public drug screening datasets and show that our model<br> outperforms state-of-the-art visible machine learning models.