Ultra-rare genetic disease, which is characterized by a prevalence rate of fewer than one in fifty thousand people, is extremely rare, poses heterogeneity in phenotype, and has limited clinical experience. An example of such challenges is Neurodevelopmental Disorder with Regression, Abnormal Movements, Loss of Speech and Seizures (NEDAMSS), which is caused by neurodegenerative pathogenic variants of the IRF2BPL gene, demonstrating a long-lasting diagnostic odyssey. The adoption of machine learning (ML) and deep learning (DL) methods presents exceptional opportunities to overcome diagnostic delays, misdiagnoses, and treatment gaps in ultra-rare disorders by utilizing high-quality pattern recognition, multimodal data integration, and predictive modeling features. A systematic review of multiple publications concludes that convolutional neural networks (CNNs) are the most widely used architecture of DL (majority of studies), then transformer models (significant portion), and graph neural networks (considerable portion). Transfer learning and few-shot learning appear as important tools to overcome the problem of data scarcity, as the reported diagnostic accuracy varies across a wide range across various types of ultra-rare disorders. The integration of ML/DL in the diagnosis of ultra-rare genetic diseases allows promising results, particularly in the case of multi-omics data integration alongside federated learning systems. Nevertheless, issues such as data standardization, model interpretability, and clinical translation remain significant obstacles to popularization.