Epilepsy, the most common neurological disorder worldwide caused by abnormal electrical activity in the brain, represents a great challenge for modern healthcare. The disorder manifests itself by recurrent periods of hyper-synchronized brain activity, or seizures, greatly hampering the quality of an affected person. Proper diagnosis and monitoring of the disease are crucial for the therapy outcomes. Electroencephalography, the most common measurement technique in epilepsy, allows for recording the patient's brain activity via scalp potentials. However, accurate analyses are time-consuming and require a trained expert's knowledge. In this paper, we exploit the non-euclidean relationships between electrodes using graph neural networks. We build our solution using a end-to-end approach, starting with determining optimal pre-processing and class balancing techniques, performing neural architecture search, and using existing algorithms to create explanations for the model's predictions. Our model achieves high performance of 0.9786 AUROC and 92.11 \% balanced accuracy in the final, 10-fold cross-validation evaluation, comparable with the existing state-of-the-art solutions, while also showing promising results in predicting the seizure onset. Obtained explanations provide insights about feature importance and changes in functional connectivity, being consistent with the current clinical knowledge. We also show that the network reliably classifies preictal activity up to 1 hour before seizure occurrence, opening a way to be used as a seizure predictor.