Prediction and uncertainty quantification of small molecule bioactivity from chemical structure and phenotypic data

Due to its diversity and increasing availability, in-vitro assay data can be a valuable source for predicting bioactivities, monitoring side effects, and repurposing drugs. Machine learning methods have been proven to be exceptionally useful tools for exploiting this information in order to find small molecular ligands with a desirable activity profile. Macau is a Bayesian matrix factorization model that was previously developed at KU Leuven and learns a latent representation of complex interactions from highly incomplete data. The aim of this work is to improve, extend and combine the benefits of the existing Macau and its deep neural network based successor, the SparseChem model. Furthermore, auxiliary information beyond molecular fingerprints will be incorporated, including, amongst others, imaging data and expression data for compounds and targets. These new model architectures will allow uncertainty estimation outside the limitations of point estimates while keeping the representation power of deep learning models, therefore enabling accurate and reliable predictions of bioactivities. Ultimately, we aim to develop reliable tools that enable the identification of ligands with a desired bioactivity profile at an  early stage in the drug discovery pipeline.