Logo: to the web site of Uppsala University

This thesis explores the implementation of Bayesian Neural Networks (BNNs) for predicting drug synergy within the RECOVER framework. To improve model robustness and accuracy, we implemented Sparse Variational Dropout on BNNs. In addition, we examined different prior distributions for BNNs, including Single Gaussian, Scale Mixture of two Gaussian, and Non-Gaussian (Laplace).

The goal is to get a better understanding of future works and possible improvement on the RECOVER pipeline by evaluating the performance of different BNN architectures in comparison to the standard RECOVER model-Neural Network. BNNs achieve higher accuracy than RECOVER in predicting synergy scores and performance. Regarding prediction, the BNN with a Laplace prior comes out as the strongest performer by showing higher accuracy and better uncertainty quantification.

Furthermore, in the integration of BNN to the active learning pipeline, efficiency in finding promising drug combinations is improved considerably compared to the RECOVER model. This efficiency translates to faster repurposing processes in laboratories and hopefully changes treatment discovery. Overall, these findings suggest that BNN can be a promising foundation for advanced drug repurposing.