Dear Editor, This submission is for the "Special Issue on NeSy 2025" and is an extension to the paper "Ontology-based box embeddings and knowledge graphs for predicting phenotypic traits in Saccharomyces cerevisiae". For this extension we have further incorporated symbolic knowledge to the training of the GNN by applying semantic losses to the generated node embeddings to adhere to class hierarchies from the ontologies. We define the new loss functions, including additional regularisation methods. We evaluated this method on the same knowledge graph learning and prediction task as in the original paper. A statistically significant performance improvement was achieved when including training for semantically correct representations. Furthermore, to demonstrate how these semantic losses can be used to train GNN-based box embeddings, we have trained embeddings using these semantic losses for a smaller family tree based knowledge graph, which provides insight into the ways the different semantic losses affect training. Finally, we propose a method for how this can be used for evaluating link revisions added to the graph. We believe this work is of interest to the Neurosymbolic AI journal as it proposes a method to generate semantically correct knowledge graph embeddings using GNNs, combining symbolic knowledge with neural network models. We also show it can be useful for real world prediction tasks. This is demonstrated with the knowledge graph we have created and the ability to predict biological measurements. Along with this we include our demonstration from the original paper on how this can be used for scientific discovery, where we use it to generate an hypothesis which is successfully tested in a biological experiment. This revised version of submission 881-1890 incorporates the feedback we got from the reviewers, which is addressed in more detail in the appended response letter. Yours sincerely, on behalf of all the authors, Filip Kronström