Deciding which experiments to execute in biology is an important task. Experimentation is costly, so having a good idea which hypotheses will be of value is important. Our idea is to use information contained in online semantic databases (e.g. SGD) to predict the value of automated generated hypotheses. These knowledge graphs have information from various sources, as well as ontologies describing the classes used. Specifically, we will investigate neuro-symbolic methods to learn and predict from theknowledge graphs, e.g., through learning embeddings. By utilising such rules search spaces can be reduced, as well as making the models and predictions more interpretable than other NN-based models by grounding them in semantic meaningful representations. These embeddings will be combined with graph neural networks (GNNs) for prediction of links and properties in the graph. Apart from developing neuro-symbolic approaches for knowledge graph learning, this project will aim to help scientific discoveries about yeast by using this to help evaluate hypotheses (and potentially form new hypotheses) and evaluating finding from automated Robot Scientists. This can, e.g., be done by predicting missing information in the graph or measuring similarities by new findings and the currently represented knowledge.